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
, rtx cmp
, rtx new_rtx
)
901 emit_insn (gen_sync_compare_and_swapsi (old
, mem
, cmp
, new_rtx
));
902 return s390_emit_compare (code
, gen_rtx_REG (CCZ1mode
, CC_REGNUM
), const0_rtx
);
905 /* Emit a jump instruction to TARGET. If COND is NULL_RTX, emit an
906 unconditional jump, else a conditional jump under condition COND. */
909 s390_emit_jump (rtx target
, rtx cond
)
913 target
= gen_rtx_LABEL_REF (VOIDmode
, target
);
915 target
= gen_rtx_IF_THEN_ELSE (VOIDmode
, cond
, target
, pc_rtx
);
917 insn
= gen_rtx_SET (VOIDmode
, pc_rtx
, target
);
918 emit_jump_insn (insn
);
921 /* Return branch condition mask to implement a branch
922 specified by CODE. Return -1 for invalid comparisons. */
925 s390_branch_condition_mask (rtx code
)
927 const int CC0
= 1 << 3;
928 const int CC1
= 1 << 2;
929 const int CC2
= 1 << 1;
930 const int CC3
= 1 << 0;
932 gcc_assert (GET_CODE (XEXP (code
, 0)) == REG
);
933 gcc_assert (REGNO (XEXP (code
, 0)) == CC_REGNUM
);
934 gcc_assert (XEXP (code
, 1) == const0_rtx
);
936 switch (GET_MODE (XEXP (code
, 0)))
940 switch (GET_CODE (code
))
943 case NE
: return CC1
| CC2
| CC3
;
949 switch (GET_CODE (code
))
952 case NE
: return CC0
| CC2
| CC3
;
958 switch (GET_CODE (code
))
961 case NE
: return CC0
| CC1
| CC3
;
967 switch (GET_CODE (code
))
970 case NE
: return CC0
| CC1
| CC2
;
976 switch (GET_CODE (code
))
978 case EQ
: return CC0
| CC2
;
979 case NE
: return CC1
| CC3
;
985 switch (GET_CODE (code
))
987 case LTU
: return CC2
| CC3
; /* carry */
988 case GEU
: return CC0
| CC1
; /* no carry */
994 switch (GET_CODE (code
))
996 case GTU
: return CC0
| CC1
; /* borrow */
997 case LEU
: return CC2
| CC3
; /* no borrow */
1003 switch (GET_CODE (code
))
1005 case EQ
: return CC0
| CC2
;
1006 case NE
: return CC1
| CC3
;
1007 case LTU
: return CC1
;
1008 case GTU
: return CC3
;
1009 case LEU
: return CC1
| CC2
;
1010 case GEU
: return CC2
| CC3
;
1015 switch (GET_CODE (code
))
1017 case EQ
: return CC0
;
1018 case NE
: return CC1
| CC2
| CC3
;
1019 case LTU
: return CC1
;
1020 case GTU
: return CC2
;
1021 case LEU
: return CC0
| CC1
;
1022 case GEU
: return CC0
| CC2
;
1028 switch (GET_CODE (code
))
1030 case EQ
: return CC0
;
1031 case NE
: return CC2
| CC1
| CC3
;
1032 case LTU
: return CC2
;
1033 case GTU
: return CC1
;
1034 case LEU
: return CC0
| CC2
;
1035 case GEU
: return CC0
| CC1
;
1041 switch (GET_CODE (code
))
1043 case EQ
: return CC0
;
1044 case NE
: return CC1
| CC2
| CC3
;
1045 case LT
: return CC1
| CC3
;
1046 case GT
: return CC2
;
1047 case LE
: return CC0
| CC1
| CC3
;
1048 case GE
: return CC0
| CC2
;
1054 switch (GET_CODE (code
))
1056 case EQ
: return CC0
;
1057 case NE
: return CC1
| CC2
| CC3
;
1058 case LT
: return CC1
;
1059 case GT
: return CC2
| CC3
;
1060 case LE
: return CC0
| CC1
;
1061 case GE
: return CC0
| CC2
| CC3
;
1067 switch (GET_CODE (code
))
1069 case EQ
: return CC0
;
1070 case NE
: return CC1
| CC2
| CC3
;
1071 case LT
: return CC1
;
1072 case GT
: return CC2
;
1073 case LE
: return CC0
| CC1
;
1074 case GE
: return CC0
| CC2
;
1075 case UNORDERED
: return CC3
;
1076 case ORDERED
: return CC0
| CC1
| CC2
;
1077 case UNEQ
: return CC0
| CC3
;
1078 case UNLT
: return CC1
| CC3
;
1079 case UNGT
: return CC2
| CC3
;
1080 case UNLE
: return CC0
| CC1
| CC3
;
1081 case UNGE
: return CC0
| CC2
| CC3
;
1082 case LTGT
: return CC1
| CC2
;
1088 switch (GET_CODE (code
))
1090 case EQ
: return CC0
;
1091 case NE
: return CC2
| CC1
| CC3
;
1092 case LT
: return CC2
;
1093 case GT
: return CC1
;
1094 case LE
: return CC0
| CC2
;
1095 case GE
: return CC0
| CC1
;
1096 case UNORDERED
: return CC3
;
1097 case ORDERED
: return CC0
| CC2
| CC1
;
1098 case UNEQ
: return CC0
| CC3
;
1099 case UNLT
: return CC2
| CC3
;
1100 case UNGT
: return CC1
| CC3
;
1101 case UNLE
: return CC0
| CC2
| CC3
;
1102 case UNGE
: return CC0
| CC1
| CC3
;
1103 case LTGT
: return CC2
| CC1
;
1114 /* Return branch condition mask to implement a compare and branch
1115 specified by CODE. Return -1 for invalid comparisons. */
1118 s390_compare_and_branch_condition_mask (rtx code
)
1120 const int CC0
= 1 << 3;
1121 const int CC1
= 1 << 2;
1122 const int CC2
= 1 << 1;
1124 switch (GET_CODE (code
))
1148 /* If INV is false, return assembler mnemonic string to implement
1149 a branch specified by CODE. If INV is true, return mnemonic
1150 for the corresponding inverted branch. */
1153 s390_branch_condition_mnemonic (rtx code
, int inv
)
1157 static const char *const mnemonic
[16] =
1159 NULL
, "o", "h", "nle",
1160 "l", "nhe", "lh", "ne",
1161 "e", "nlh", "he", "nl",
1162 "le", "nh", "no", NULL
1165 if (GET_CODE (XEXP (code
, 0)) == REG
1166 && REGNO (XEXP (code
, 0)) == CC_REGNUM
1167 && XEXP (code
, 1) == const0_rtx
)
1168 mask
= s390_branch_condition_mask (code
);
1170 mask
= s390_compare_and_branch_condition_mask (code
);
1172 gcc_assert (mask
>= 0);
1177 gcc_assert (mask
>= 1 && mask
<= 14);
1179 return mnemonic
[mask
];
1182 /* Return the part of op which has a value different from def.
1183 The size of the part is determined by mode.
1184 Use this function only if you already know that op really
1185 contains such a part. */
1187 unsigned HOST_WIDE_INT
1188 s390_extract_part (rtx op
, enum machine_mode mode
, int def
)
1190 unsigned HOST_WIDE_INT value
= 0;
1191 int max_parts
= HOST_BITS_PER_WIDE_INT
/ GET_MODE_BITSIZE (mode
);
1192 int part_bits
= GET_MODE_BITSIZE (mode
);
1193 unsigned HOST_WIDE_INT part_mask
1194 = ((unsigned HOST_WIDE_INT
)1 << part_bits
) - 1;
1197 for (i
= 0; i
< max_parts
; i
++)
1200 value
= (unsigned HOST_WIDE_INT
) INTVAL (op
);
1202 value
>>= part_bits
;
1204 if ((value
& part_mask
) != (def
& part_mask
))
1205 return value
& part_mask
;
1211 /* If OP is an integer constant of mode MODE with exactly one
1212 part of mode PART_MODE unequal to DEF, return the number of that
1213 part. Otherwise, return -1. */
1216 s390_single_part (rtx op
,
1217 enum machine_mode mode
,
1218 enum machine_mode part_mode
,
1221 unsigned HOST_WIDE_INT value
= 0;
1222 int n_parts
= GET_MODE_SIZE (mode
) / GET_MODE_SIZE (part_mode
);
1223 unsigned HOST_WIDE_INT part_mask
1224 = ((unsigned HOST_WIDE_INT
)1 << GET_MODE_BITSIZE (part_mode
)) - 1;
1227 if (GET_CODE (op
) != CONST_INT
)
1230 for (i
= 0; i
< n_parts
; i
++)
1233 value
= (unsigned HOST_WIDE_INT
) INTVAL (op
);
1235 value
>>= GET_MODE_BITSIZE (part_mode
);
1237 if ((value
& part_mask
) != (def
& part_mask
))
1245 return part
== -1 ? -1 : n_parts
- 1 - part
;
1248 /* Return true if IN contains a contiguous bitfield in the lower SIZE
1249 bits and no other bits are set in IN. POS and LENGTH can be used
1250 to obtain the start position and the length of the bitfield.
1252 POS gives the position of the first bit of the bitfield counting
1253 from the lowest order bit starting with zero. In order to use this
1254 value for S/390 instructions this has to be converted to "bits big
1258 s390_contiguous_bitmask_p (unsigned HOST_WIDE_INT in
, int size
,
1259 int *pos
, int *length
)
1264 unsigned HOST_WIDE_INT mask
= 1ULL;
1265 bool contiguous
= false;
1267 for (i
= 0; i
< size
; mask
<<= 1, i
++)
1291 /* Calculate a mask for all bits beyond the contiguous bits. */
1292 mask
= (-1LL & ~(((1ULL << (tmp_length
+ tmp_pos
- 1)) << 1) - 1));
1297 if (tmp_length
+ tmp_pos
- 1 > size
)
1301 *length
= tmp_length
;
1309 /* Check whether we can (and want to) split a double-word
1310 move in mode MODE from SRC to DST into two single-word
1311 moves, moving the subword FIRST_SUBWORD first. */
1314 s390_split_ok_p (rtx dst
, rtx src
, enum machine_mode mode
, int first_subword
)
1316 /* Floating point registers cannot be split. */
1317 if (FP_REG_P (src
) || FP_REG_P (dst
))
1320 /* We don't need to split if operands are directly accessible. */
1321 if (s_operand (src
, mode
) || s_operand (dst
, mode
))
1324 /* Non-offsettable memory references cannot be split. */
1325 if ((GET_CODE (src
) == MEM
&& !offsettable_memref_p (src
))
1326 || (GET_CODE (dst
) == MEM
&& !offsettable_memref_p (dst
)))
1329 /* Moving the first subword must not clobber a register
1330 needed to move the second subword. */
1331 if (register_operand (dst
, mode
))
1333 rtx subreg
= operand_subword (dst
, first_subword
, 0, mode
);
1334 if (reg_overlap_mentioned_p (subreg
, src
))
1341 /* Return true if it can be proven that [MEM1, MEM1 + SIZE]
1342 and [MEM2, MEM2 + SIZE] do overlap and false
1346 s390_overlap_p (rtx mem1
, rtx mem2
, HOST_WIDE_INT size
)
1348 rtx addr1
, addr2
, addr_delta
;
1349 HOST_WIDE_INT delta
;
1351 if (GET_CODE (mem1
) != MEM
|| GET_CODE (mem2
) != MEM
)
1357 addr1
= XEXP (mem1
, 0);
1358 addr2
= XEXP (mem2
, 0);
1360 addr_delta
= simplify_binary_operation (MINUS
, Pmode
, addr2
, addr1
);
1362 /* This overlapping check is used by peepholes merging memory block operations.
1363 Overlapping operations would otherwise be recognized by the S/390 hardware
1364 and would fall back to a slower implementation. Allowing overlapping
1365 operations would lead to slow code but not to wrong code. Therefore we are
1366 somewhat optimistic if we cannot prove that the memory blocks are
1368 That's why we return false here although this may accept operations on
1369 overlapping memory areas. */
1370 if (!addr_delta
|| GET_CODE (addr_delta
) != CONST_INT
)
1373 delta
= INTVAL (addr_delta
);
1376 || (delta
> 0 && delta
< size
)
1377 || (delta
< 0 && -delta
< size
))
1383 /* Check whether the address of memory reference MEM2 equals exactly
1384 the address of memory reference MEM1 plus DELTA. Return true if
1385 we can prove this to be the case, false otherwise. */
1388 s390_offset_p (rtx mem1
, rtx mem2
, rtx delta
)
1390 rtx addr1
, addr2
, addr_delta
;
1392 if (GET_CODE (mem1
) != MEM
|| GET_CODE (mem2
) != MEM
)
1395 addr1
= XEXP (mem1
, 0);
1396 addr2
= XEXP (mem2
, 0);
1398 addr_delta
= simplify_binary_operation (MINUS
, Pmode
, addr2
, addr1
);
1399 if (!addr_delta
|| !rtx_equal_p (addr_delta
, delta
))
1405 /* Expand logical operator CODE in mode MODE with operands OPERANDS. */
1408 s390_expand_logical_operator (enum rtx_code code
, enum machine_mode mode
,
1411 enum machine_mode wmode
= mode
;
1412 rtx dst
= operands
[0];
1413 rtx src1
= operands
[1];
1414 rtx src2
= operands
[2];
1417 /* If we cannot handle the operation directly, use a temp register. */
1418 if (!s390_logical_operator_ok_p (operands
))
1419 dst
= gen_reg_rtx (mode
);
1421 /* QImode and HImode patterns make sense only if we have a destination
1422 in memory. Otherwise perform the operation in SImode. */
1423 if ((mode
== QImode
|| mode
== HImode
) && GET_CODE (dst
) != MEM
)
1426 /* Widen operands if required. */
1429 if (GET_CODE (dst
) == SUBREG
1430 && (tem
= simplify_subreg (wmode
, dst
, mode
, 0)) != 0)
1432 else if (REG_P (dst
))
1433 dst
= gen_rtx_SUBREG (wmode
, dst
, 0);
1435 dst
= gen_reg_rtx (wmode
);
1437 if (GET_CODE (src1
) == SUBREG
1438 && (tem
= simplify_subreg (wmode
, src1
, mode
, 0)) != 0)
1440 else if (GET_MODE (src1
) != VOIDmode
)
1441 src1
= gen_rtx_SUBREG (wmode
, force_reg (mode
, src1
), 0);
1443 if (GET_CODE (src2
) == SUBREG
1444 && (tem
= simplify_subreg (wmode
, src2
, mode
, 0)) != 0)
1446 else if (GET_MODE (src2
) != VOIDmode
)
1447 src2
= gen_rtx_SUBREG (wmode
, force_reg (mode
, src2
), 0);
1450 /* Emit the instruction. */
1451 op
= gen_rtx_SET (VOIDmode
, dst
, gen_rtx_fmt_ee (code
, wmode
, src1
, src2
));
1452 clob
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
1453 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, op
, clob
)));
1455 /* Fix up the destination if needed. */
1456 if (dst
!= operands
[0])
1457 emit_move_insn (operands
[0], gen_lowpart (mode
, dst
));
1460 /* Check whether OPERANDS are OK for a logical operation (AND, IOR, XOR). */
1463 s390_logical_operator_ok_p (rtx
*operands
)
1465 /* If the destination operand is in memory, it needs to coincide
1466 with one of the source operands. After reload, it has to be
1467 the first source operand. */
1468 if (GET_CODE (operands
[0]) == MEM
)
1469 return rtx_equal_p (operands
[0], operands
[1])
1470 || (!reload_completed
&& rtx_equal_p (operands
[0], operands
[2]));
1475 /* Narrow logical operation CODE of memory operand MEMOP with immediate
1476 operand IMMOP to switch from SS to SI type instructions. */
1479 s390_narrow_logical_operator (enum rtx_code code
, rtx
*memop
, rtx
*immop
)
1481 int def
= code
== AND
? -1 : 0;
1485 gcc_assert (GET_CODE (*memop
) == MEM
);
1486 gcc_assert (!MEM_VOLATILE_P (*memop
));
1488 mask
= s390_extract_part (*immop
, QImode
, def
);
1489 part
= s390_single_part (*immop
, GET_MODE (*memop
), QImode
, def
);
1490 gcc_assert (part
>= 0);
1492 *memop
= adjust_address (*memop
, QImode
, part
);
1493 *immop
= gen_int_mode (mask
, QImode
);
1497 /* How to allocate a 'struct machine_function'. */
1499 static struct machine_function
*
1500 s390_init_machine_status (void)
1502 return ggc_alloc_cleared_machine_function ();
1506 s390_option_override (void)
1508 /* Set up function hooks. */
1509 init_machine_status
= s390_init_machine_status
;
1511 /* Architecture mode defaults according to ABI. */
1512 if (!(target_flags_explicit
& MASK_ZARCH
))
1515 target_flags
|= MASK_ZARCH
;
1517 target_flags
&= ~MASK_ZARCH
;
1520 /* Set the march default in case it hasn't been specified on
1522 if (s390_arch
== PROCESSOR_max
)
1524 s390_arch_string
= TARGET_ZARCH
? "z900" : "g5";
1525 s390_arch
= TARGET_ZARCH
? PROCESSOR_2064_Z900
: PROCESSOR_9672_G5
;
1526 s390_arch_flags
= processor_flags_table
[(int)s390_arch
];
1529 /* Determine processor to tune for. */
1530 if (s390_tune
== PROCESSOR_max
)
1532 s390_tune
= s390_arch
;
1533 s390_tune_flags
= s390_arch_flags
;
1536 /* Sanity checks. */
1537 if (TARGET_ZARCH
&& !TARGET_CPU_ZARCH
)
1538 error ("z/Architecture mode not supported on %s", s390_arch_string
);
1539 if (TARGET_64BIT
&& !TARGET_ZARCH
)
1540 error ("64-bit ABI not supported in ESA/390 mode");
1542 /* Use hardware DFP if available and not explicitly disabled by
1543 user. E.g. with -m31 -march=z10 -mzarch */
1544 if (!(target_flags_explicit
& MASK_HARD_DFP
) && TARGET_DFP
)
1545 target_flags
|= MASK_HARD_DFP
;
1547 if (TARGET_HARD_DFP
&& !TARGET_DFP
)
1549 if (target_flags_explicit
& MASK_HARD_DFP
)
1551 if (!TARGET_CPU_DFP
)
1552 error ("hardware decimal floating point instructions"
1553 " not available on %s", s390_arch_string
);
1555 error ("hardware decimal floating point instructions"
1556 " not available in ESA/390 mode");
1559 target_flags
&= ~MASK_HARD_DFP
;
1562 if ((target_flags_explicit
& MASK_SOFT_FLOAT
) && TARGET_SOFT_FLOAT
)
1564 if ((target_flags_explicit
& MASK_HARD_DFP
) && TARGET_HARD_DFP
)
1565 error ("-mhard-dfp can%'t be used in conjunction with -msoft-float");
1567 target_flags
&= ~MASK_HARD_DFP
;
1570 /* Set processor cost function. */
1573 case PROCESSOR_2084_Z990
:
1574 s390_cost
= &z990_cost
;
1576 case PROCESSOR_2094_Z9_109
:
1577 s390_cost
= &z9_109_cost
;
1579 case PROCESSOR_2097_Z10
:
1580 s390_cost
= &z10_cost
;
1581 case PROCESSOR_2817_Z196
:
1582 s390_cost
= &z196_cost
;
1585 s390_cost
= &z900_cost
;
1588 if (TARGET_BACKCHAIN
&& TARGET_PACKED_STACK
&& TARGET_HARD_FLOAT
)
1589 error ("-mbackchain -mpacked-stack -mhard-float are not supported "
1592 if (s390_stack_size
)
1594 if (s390_stack_guard
>= s390_stack_size
)
1595 error ("stack size must be greater than the stack guard value");
1596 else if (s390_stack_size
> 1 << 16)
1597 error ("stack size must not be greater than 64k");
1599 else if (s390_stack_guard
)
1600 error ("-mstack-guard implies use of -mstack-size");
1602 #ifdef TARGET_DEFAULT_LONG_DOUBLE_128
1603 if (!(target_flags_explicit
& MASK_LONG_DOUBLE_128
))
1604 target_flags
|= MASK_LONG_DOUBLE_128
;
1607 if (s390_tune
== PROCESSOR_2097_Z10
1608 || s390_tune
== PROCESSOR_2817_Z196
)
1610 maybe_set_param_value (PARAM_MAX_UNROLLED_INSNS
, 100,
1611 global_options
.x_param_values
,
1612 global_options_set
.x_param_values
);
1613 maybe_set_param_value (PARAM_MAX_UNROLL_TIMES
, 32,
1614 global_options
.x_param_values
,
1615 global_options_set
.x_param_values
);
1616 maybe_set_param_value (PARAM_MAX_COMPLETELY_PEELED_INSNS
, 2000,
1617 global_options
.x_param_values
,
1618 global_options_set
.x_param_values
);
1619 maybe_set_param_value (PARAM_MAX_COMPLETELY_PEEL_TIMES
, 64,
1620 global_options
.x_param_values
,
1621 global_options_set
.x_param_values
);
1624 maybe_set_param_value (PARAM_MAX_PENDING_LIST_LENGTH
, 256,
1625 global_options
.x_param_values
,
1626 global_options_set
.x_param_values
);
1627 /* values for loop prefetching */
1628 maybe_set_param_value (PARAM_L1_CACHE_LINE_SIZE
, 256,
1629 global_options
.x_param_values
,
1630 global_options_set
.x_param_values
);
1631 maybe_set_param_value (PARAM_L1_CACHE_SIZE
, 128,
1632 global_options
.x_param_values
,
1633 global_options_set
.x_param_values
);
1634 /* s390 has more than 2 levels and the size is much larger. Since
1635 we are always running virtualized assume that we only get a small
1636 part of the caches above l1. */
1637 maybe_set_param_value (PARAM_L2_CACHE_SIZE
, 1500,
1638 global_options
.x_param_values
,
1639 global_options_set
.x_param_values
);
1640 maybe_set_param_value (PARAM_PREFETCH_MIN_INSN_TO_MEM_RATIO
, 2,
1641 global_options
.x_param_values
,
1642 global_options_set
.x_param_values
);
1643 maybe_set_param_value (PARAM_SIMULTANEOUS_PREFETCHES
, 6,
1644 global_options
.x_param_values
,
1645 global_options_set
.x_param_values
);
1647 /* This cannot reside in s390_option_optimization_table since HAVE_prefetch
1648 requires the arch flags to be evaluated already. Since prefetching
1649 is beneficial on s390, we enable it if available. */
1650 if (flag_prefetch_loop_arrays
< 0 && HAVE_prefetch
&& optimize
>= 3)
1651 flag_prefetch_loop_arrays
= 1;
1653 /* Use the alternative scheduling-pressure algorithm by default. */
1654 maybe_set_param_value (PARAM_SCHED_PRESSURE_ALGORITHM
, 2,
1655 global_options
.x_param_values
,
1656 global_options_set
.x_param_values
);
1660 /* Don't emit DWARF3/4 unless specifically selected. The TPF
1661 debuggers do not yet support DWARF 3/4. */
1662 if (!global_options_set
.x_dwarf_strict
)
1664 if (!global_options_set
.x_dwarf_version
)
1669 /* Map for smallest class containing reg regno. */
1671 const enum reg_class regclass_map
[FIRST_PSEUDO_REGISTER
] =
1672 { GENERAL_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1673 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1674 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1675 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1676 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1677 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1678 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1679 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1680 ADDR_REGS
, CC_REGS
, ADDR_REGS
, ADDR_REGS
,
1681 ACCESS_REGS
, ACCESS_REGS
1684 /* Return attribute type of insn. */
1686 static enum attr_type
1687 s390_safe_attr_type (rtx insn
)
1689 if (recog_memoized (insn
) >= 0)
1690 return get_attr_type (insn
);
1695 /* Return true if DISP is a valid short displacement. */
1698 s390_short_displacement (rtx disp
)
1700 /* No displacement is OK. */
1704 /* Without the long displacement facility we don't need to
1705 distingiush between long and short displacement. */
1706 if (!TARGET_LONG_DISPLACEMENT
)
1709 /* Integer displacement in range. */
1710 if (GET_CODE (disp
) == CONST_INT
)
1711 return INTVAL (disp
) >= 0 && INTVAL (disp
) < 4096;
1713 /* GOT offset is not OK, the GOT can be large. */
1714 if (GET_CODE (disp
) == CONST
1715 && GET_CODE (XEXP (disp
, 0)) == UNSPEC
1716 && (XINT (XEXP (disp
, 0), 1) == UNSPEC_GOT
1717 || XINT (XEXP (disp
, 0), 1) == UNSPEC_GOTNTPOFF
))
1720 /* All other symbolic constants are literal pool references,
1721 which are OK as the literal pool must be small. */
1722 if (GET_CODE (disp
) == CONST
)
1728 /* Decompose a RTL expression ADDR for a memory address into
1729 its components, returned in OUT.
1731 Returns false if ADDR is not a valid memory address, true
1732 otherwise. If OUT is NULL, don't return the components,
1733 but check for validity only.
1735 Note: Only addresses in canonical form are recognized.
1736 LEGITIMIZE_ADDRESS should convert non-canonical forms to the
1737 canonical form so that they will be recognized. */
1740 s390_decompose_address (rtx addr
, struct s390_address
*out
)
1742 HOST_WIDE_INT offset
= 0;
1743 rtx base
= NULL_RTX
;
1744 rtx indx
= NULL_RTX
;
1745 rtx disp
= NULL_RTX
;
1747 bool pointer
= false;
1748 bool base_ptr
= false;
1749 bool indx_ptr
= false;
1750 bool literal_pool
= false;
1752 /* We may need to substitute the literal pool base register into the address
1753 below. However, at this point we do not know which register is going to
1754 be used as base, so we substitute the arg pointer register. This is going
1755 to be treated as holding a pointer below -- it shouldn't be used for any
1757 rtx fake_pool_base
= gen_rtx_REG (Pmode
, ARG_POINTER_REGNUM
);
1759 /* Decompose address into base + index + displacement. */
1761 if (GET_CODE (addr
) == REG
|| GET_CODE (addr
) == UNSPEC
)
1764 else if (GET_CODE (addr
) == PLUS
)
1766 rtx op0
= XEXP (addr
, 0);
1767 rtx op1
= XEXP (addr
, 1);
1768 enum rtx_code code0
= GET_CODE (op0
);
1769 enum rtx_code code1
= GET_CODE (op1
);
1771 if (code0
== REG
|| code0
== UNSPEC
)
1773 if (code1
== REG
|| code1
== UNSPEC
)
1775 indx
= op0
; /* index + base */
1781 base
= op0
; /* base + displacement */
1786 else if (code0
== PLUS
)
1788 indx
= XEXP (op0
, 0); /* index + base + disp */
1789 base
= XEXP (op0
, 1);
1800 disp
= addr
; /* displacement */
1802 /* Extract integer part of displacement. */
1806 if (GET_CODE (disp
) == CONST_INT
)
1808 offset
= INTVAL (disp
);
1811 else if (GET_CODE (disp
) == CONST
1812 && GET_CODE (XEXP (disp
, 0)) == PLUS
1813 && GET_CODE (XEXP (XEXP (disp
, 0), 1)) == CONST_INT
)
1815 offset
= INTVAL (XEXP (XEXP (disp
, 0), 1));
1816 disp
= XEXP (XEXP (disp
, 0), 0);
1820 /* Strip off CONST here to avoid special case tests later. */
1821 if (disp
&& GET_CODE (disp
) == CONST
)
1822 disp
= XEXP (disp
, 0);
1824 /* We can convert literal pool addresses to
1825 displacements by basing them off the base register. */
1826 if (disp
&& GET_CODE (disp
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (disp
))
1828 /* Either base or index must be free to hold the base register. */
1830 base
= fake_pool_base
, literal_pool
= true;
1832 indx
= fake_pool_base
, literal_pool
= true;
1836 /* Mark up the displacement. */
1837 disp
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, disp
),
1838 UNSPEC_LTREL_OFFSET
);
1841 /* Validate base register. */
1844 if (GET_CODE (base
) == UNSPEC
)
1845 switch (XINT (base
, 1))
1849 disp
= gen_rtx_UNSPEC (Pmode
,
1850 gen_rtvec (1, XVECEXP (base
, 0, 0)),
1851 UNSPEC_LTREL_OFFSET
);
1855 base
= XVECEXP (base
, 0, 1);
1858 case UNSPEC_LTREL_BASE
:
1859 if (XVECLEN (base
, 0) == 1)
1860 base
= fake_pool_base
, literal_pool
= true;
1862 base
= XVECEXP (base
, 0, 1);
1870 || (GET_MODE (base
) != SImode
1871 && GET_MODE (base
) != Pmode
))
1874 if (REGNO (base
) == STACK_POINTER_REGNUM
1875 || REGNO (base
) == FRAME_POINTER_REGNUM
1876 || ((reload_completed
|| reload_in_progress
)
1877 && frame_pointer_needed
1878 && REGNO (base
) == HARD_FRAME_POINTER_REGNUM
)
1879 || REGNO (base
) == ARG_POINTER_REGNUM
1881 && REGNO (base
) == PIC_OFFSET_TABLE_REGNUM
))
1882 pointer
= base_ptr
= true;
1884 if ((reload_completed
|| reload_in_progress
)
1885 && base
== cfun
->machine
->base_reg
)
1886 pointer
= base_ptr
= literal_pool
= true;
1889 /* Validate index register. */
1892 if (GET_CODE (indx
) == UNSPEC
)
1893 switch (XINT (indx
, 1))
1897 disp
= gen_rtx_UNSPEC (Pmode
,
1898 gen_rtvec (1, XVECEXP (indx
, 0, 0)),
1899 UNSPEC_LTREL_OFFSET
);
1903 indx
= XVECEXP (indx
, 0, 1);
1906 case UNSPEC_LTREL_BASE
:
1907 if (XVECLEN (indx
, 0) == 1)
1908 indx
= fake_pool_base
, literal_pool
= true;
1910 indx
= XVECEXP (indx
, 0, 1);
1918 || (GET_MODE (indx
) != SImode
1919 && GET_MODE (indx
) != Pmode
))
1922 if (REGNO (indx
) == STACK_POINTER_REGNUM
1923 || REGNO (indx
) == FRAME_POINTER_REGNUM
1924 || ((reload_completed
|| reload_in_progress
)
1925 && frame_pointer_needed
1926 && REGNO (indx
) == HARD_FRAME_POINTER_REGNUM
)
1927 || REGNO (indx
) == ARG_POINTER_REGNUM
1929 && REGNO (indx
) == PIC_OFFSET_TABLE_REGNUM
))
1930 pointer
= indx_ptr
= true;
1932 if ((reload_completed
|| reload_in_progress
)
1933 && indx
== cfun
->machine
->base_reg
)
1934 pointer
= indx_ptr
= literal_pool
= true;
1937 /* Prefer to use pointer as base, not index. */
1938 if (base
&& indx
&& !base_ptr
1939 && (indx_ptr
|| (!REG_POINTER (base
) && REG_POINTER (indx
))))
1946 /* Validate displacement. */
1949 /* If virtual registers are involved, the displacement will change later
1950 anyway as the virtual registers get eliminated. This could make a
1951 valid displacement invalid, but it is more likely to make an invalid
1952 displacement valid, because we sometimes access the register save area
1953 via negative offsets to one of those registers.
1954 Thus we don't check the displacement for validity here. If after
1955 elimination the displacement turns out to be invalid after all,
1956 this is fixed up by reload in any case. */
1957 if (base
!= arg_pointer_rtx
1958 && indx
!= arg_pointer_rtx
1959 && base
!= return_address_pointer_rtx
1960 && indx
!= return_address_pointer_rtx
1961 && base
!= frame_pointer_rtx
1962 && indx
!= frame_pointer_rtx
1963 && base
!= virtual_stack_vars_rtx
1964 && indx
!= virtual_stack_vars_rtx
)
1965 if (!DISP_IN_RANGE (offset
))
1970 /* All the special cases are pointers. */
1973 /* In the small-PIC case, the linker converts @GOT
1974 and @GOTNTPOFF offsets to possible displacements. */
1975 if (GET_CODE (disp
) == UNSPEC
1976 && (XINT (disp
, 1) == UNSPEC_GOT
1977 || XINT (disp
, 1) == UNSPEC_GOTNTPOFF
)
1983 /* Accept pool label offsets. */
1984 else if (GET_CODE (disp
) == UNSPEC
1985 && XINT (disp
, 1) == UNSPEC_POOL_OFFSET
)
1988 /* Accept literal pool references. */
1989 else if (GET_CODE (disp
) == UNSPEC
1990 && XINT (disp
, 1) == UNSPEC_LTREL_OFFSET
)
1992 /* In case CSE pulled a non literal pool reference out of
1993 the pool we have to reject the address. This is
1994 especially important when loading the GOT pointer on non
1995 zarch CPUs. In this case the literal pool contains an lt
1996 relative offset to the _GLOBAL_OFFSET_TABLE_ label which
1997 will most likely exceed the displacement. */
1998 if (GET_CODE (XVECEXP (disp
, 0, 0)) != SYMBOL_REF
1999 || !CONSTANT_POOL_ADDRESS_P (XVECEXP (disp
, 0, 0)))
2002 orig_disp
= gen_rtx_CONST (Pmode
, disp
);
2005 /* If we have an offset, make sure it does not
2006 exceed the size of the constant pool entry. */
2007 rtx sym
= XVECEXP (disp
, 0, 0);
2008 if (offset
>= GET_MODE_SIZE (get_pool_mode (sym
)))
2011 orig_disp
= plus_constant (Pmode
, orig_disp
, offset
);
2026 out
->disp
= orig_disp
;
2027 out
->pointer
= pointer
;
2028 out
->literal_pool
= literal_pool
;
2034 /* Decompose a RTL expression OP for a shift count into its components,
2035 and return the base register in BASE and the offset in OFFSET.
2037 Return true if OP is a valid shift count, false if not. */
2040 s390_decompose_shift_count (rtx op
, rtx
*base
, HOST_WIDE_INT
*offset
)
2042 HOST_WIDE_INT off
= 0;
2044 /* We can have an integer constant, an address register,
2045 or a sum of the two. */
2046 if (GET_CODE (op
) == CONST_INT
)
2051 if (op
&& GET_CODE (op
) == PLUS
&& GET_CODE (XEXP (op
, 1)) == CONST_INT
)
2053 off
= INTVAL (XEXP (op
, 1));
2056 while (op
&& GET_CODE (op
) == SUBREG
)
2057 op
= SUBREG_REG (op
);
2059 if (op
&& GET_CODE (op
) != REG
)
2071 /* Return true if CODE is a valid address without index. */
2074 s390_legitimate_address_without_index_p (rtx op
)
2076 struct s390_address addr
;
2078 if (!s390_decompose_address (XEXP (op
, 0), &addr
))
2087 /* Return true if ADDR is of kind symbol_ref or symbol_ref + const_int
2088 and return these parts in SYMREF and ADDEND. You can pass NULL in
2089 SYMREF and/or ADDEND if you are not interested in these values.
2090 Literal pool references are *not* considered symbol references. */
2093 s390_symref_operand_p (rtx addr
, rtx
*symref
, HOST_WIDE_INT
*addend
)
2095 HOST_WIDE_INT tmpaddend
= 0;
2097 if (GET_CODE (addr
) == CONST
)
2098 addr
= XEXP (addr
, 0);
2100 if (GET_CODE (addr
) == PLUS
)
2102 if (GET_CODE (XEXP (addr
, 0)) == SYMBOL_REF
2103 && !CONSTANT_POOL_ADDRESS_P (XEXP (addr
, 0))
2104 && CONST_INT_P (XEXP (addr
, 1)))
2106 tmpaddend
= INTVAL (XEXP (addr
, 1));
2107 addr
= XEXP (addr
, 0);
2113 if (GET_CODE (addr
) != SYMBOL_REF
|| CONSTANT_POOL_ADDRESS_P (addr
))
2119 *addend
= tmpaddend
;
2125 /* Return true if the address in OP is valid for constraint letter C
2126 if wrapped in a MEM rtx. Set LIT_POOL_OK to true if it literal
2127 pool MEMs should be accepted. Only the Q, R, S, T constraint
2128 letters are allowed for C. */
2131 s390_check_qrst_address (char c
, rtx op
, bool lit_pool_ok
)
2133 struct s390_address addr
;
2134 bool decomposed
= false;
2136 /* This check makes sure that no symbolic address (except literal
2137 pool references) are accepted by the R or T constraints. */
2138 if (s390_symref_operand_p (op
, NULL
, NULL
))
2141 /* Ensure literal pool references are only accepted if LIT_POOL_OK. */
2144 if (!s390_decompose_address (op
, &addr
))
2146 if (addr
.literal_pool
)
2153 case 'Q': /* no index short displacement */
2154 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2158 if (!s390_short_displacement (addr
.disp
))
2162 case 'R': /* with index short displacement */
2163 if (TARGET_LONG_DISPLACEMENT
)
2165 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2167 if (!s390_short_displacement (addr
.disp
))
2170 /* Any invalid address here will be fixed up by reload,
2171 so accept it for the most generic constraint. */
2174 case 'S': /* no index long displacement */
2175 if (!TARGET_LONG_DISPLACEMENT
)
2177 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2181 if (s390_short_displacement (addr
.disp
))
2185 case 'T': /* with index long displacement */
2186 if (!TARGET_LONG_DISPLACEMENT
)
2188 /* Any invalid address here will be fixed up by reload,
2189 so accept it for the most generic constraint. */
2190 if ((decomposed
|| s390_decompose_address (op
, &addr
))
2191 && s390_short_displacement (addr
.disp
))
2201 /* Evaluates constraint strings described by the regular expression
2202 ([A|B|Z](Q|R|S|T))|U|W|Y and returns 1 if OP is a valid operand for
2203 the constraint given in STR, or 0 else. */
2206 s390_mem_constraint (const char *str
, rtx op
)
2213 /* Check for offsettable variants of memory constraints. */
2214 if (!MEM_P (op
) || MEM_VOLATILE_P (op
))
2216 if ((reload_completed
|| reload_in_progress
)
2217 ? !offsettable_memref_p (op
) : !offsettable_nonstrict_memref_p (op
))
2219 return s390_check_qrst_address (str
[1], XEXP (op
, 0), true);
2221 /* Check for non-literal-pool variants of memory constraints. */
2224 return s390_check_qrst_address (str
[1], XEXP (op
, 0), false);
2229 if (GET_CODE (op
) != MEM
)
2231 return s390_check_qrst_address (c
, XEXP (op
, 0), true);
2233 return (s390_check_qrst_address ('Q', op
, true)
2234 || s390_check_qrst_address ('R', op
, true));
2236 return (s390_check_qrst_address ('S', op
, true)
2237 || s390_check_qrst_address ('T', op
, true));
2239 /* Simply check for the basic form of a shift count. Reload will
2240 take care of making sure we have a proper base register. */
2241 if (!s390_decompose_shift_count (op
, NULL
, NULL
))
2245 return s390_check_qrst_address (str
[1], op
, true);
2253 /* Evaluates constraint strings starting with letter O. Input
2254 parameter C is the second letter following the "O" in the constraint
2255 string. Returns 1 if VALUE meets the respective constraint and 0
2259 s390_O_constraint_str (const char c
, HOST_WIDE_INT value
)
2267 return trunc_int_for_mode (value
, SImode
) == value
;
2271 || s390_single_part (GEN_INT (value
), DImode
, SImode
, 0) == 1;
2274 return s390_single_part (GEN_INT (value
- 1), DImode
, SImode
, -1) == 1;
2282 /* Evaluates constraint strings starting with letter N. Parameter STR
2283 contains the letters following letter "N" in the constraint string.
2284 Returns true if VALUE matches the constraint. */
2287 s390_N_constraint_str (const char *str
, HOST_WIDE_INT value
)
2289 enum machine_mode mode
, part_mode
;
2291 int part
, part_goal
;
2297 part_goal
= str
[0] - '0';
2341 if (GET_MODE_SIZE (mode
) <= GET_MODE_SIZE (part_mode
))
2344 part
= s390_single_part (GEN_INT (value
), mode
, part_mode
, def
);
2347 if (part_goal
!= -1 && part_goal
!= part
)
2354 /* Returns true if the input parameter VALUE is a float zero. */
2357 s390_float_const_zero_p (rtx value
)
2359 return (GET_MODE_CLASS (GET_MODE (value
)) == MODE_FLOAT
2360 && value
== CONST0_RTX (GET_MODE (value
)));
2363 /* Implement TARGET_REGISTER_MOVE_COST. */
2366 s390_register_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED
,
2367 reg_class_t from
, reg_class_t to
)
2369 /* On s390, copy between fprs and gprs is expensive. */
2370 if ((reg_classes_intersect_p (from
, GENERAL_REGS
)
2371 && reg_classes_intersect_p (to
, FP_REGS
))
2372 || (reg_classes_intersect_p (from
, FP_REGS
)
2373 && reg_classes_intersect_p (to
, GENERAL_REGS
)))
2379 /* Implement TARGET_MEMORY_MOVE_COST. */
2382 s390_memory_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED
,
2383 reg_class_t rclass ATTRIBUTE_UNUSED
,
2384 bool in ATTRIBUTE_UNUSED
)
2389 /* Compute a (partial) cost for rtx X. Return true if the complete
2390 cost has been computed, and false if subexpressions should be
2391 scanned. In either case, *TOTAL contains the cost result.
2392 CODE contains GET_CODE (x), OUTER_CODE contains the code
2393 of the superexpression of x. */
2396 s390_rtx_costs (rtx x
, int code
, int outer_code
, int opno ATTRIBUTE_UNUSED
,
2397 int *total
, bool speed ATTRIBUTE_UNUSED
)
2420 *total
= COSTS_N_INSNS (1);
2425 *total
= COSTS_N_INSNS (1);
2429 switch (GET_MODE (x
))
2433 rtx left
= XEXP (x
, 0);
2434 rtx right
= XEXP (x
, 1);
2435 if (GET_CODE (right
) == CONST_INT
2436 && CONST_OK_FOR_K (INTVAL (right
)))
2437 *total
= s390_cost
->mhi
;
2438 else if (GET_CODE (left
) == SIGN_EXTEND
)
2439 *total
= s390_cost
->mh
;
2441 *total
= s390_cost
->ms
; /* msr, ms, msy */
2446 rtx left
= XEXP (x
, 0);
2447 rtx right
= XEXP (x
, 1);
2450 if (GET_CODE (right
) == CONST_INT
2451 && CONST_OK_FOR_K (INTVAL (right
)))
2452 *total
= s390_cost
->mghi
;
2453 else if (GET_CODE (left
) == SIGN_EXTEND
)
2454 *total
= s390_cost
->msgf
;
2456 *total
= s390_cost
->msg
; /* msgr, msg */
2458 else /* TARGET_31BIT */
2460 if (GET_CODE (left
) == SIGN_EXTEND
2461 && GET_CODE (right
) == SIGN_EXTEND
)
2462 /* mulsidi case: mr, m */
2463 *total
= s390_cost
->m
;
2464 else if (GET_CODE (left
) == ZERO_EXTEND
2465 && GET_CODE (right
) == ZERO_EXTEND
2466 && TARGET_CPU_ZARCH
)
2467 /* umulsidi case: ml, mlr */
2468 *total
= s390_cost
->ml
;
2470 /* Complex calculation is required. */
2471 *total
= COSTS_N_INSNS (40);
2477 *total
= s390_cost
->mult_df
;
2480 *total
= s390_cost
->mxbr
;
2488 switch (GET_MODE (x
))
2491 *total
= s390_cost
->madbr
;
2494 *total
= s390_cost
->maebr
;
2499 /* Negate in the third argument is free: FMSUB. */
2500 if (GET_CODE (XEXP (x
, 2)) == NEG
)
2502 *total
+= (rtx_cost (XEXP (x
, 0), FMA
, 0, speed
)
2503 + rtx_cost (XEXP (x
, 1), FMA
, 1, speed
)
2504 + rtx_cost (XEXP (XEXP (x
, 2), 0), FMA
, 2, speed
));
2511 if (GET_MODE (x
) == TImode
) /* 128 bit division */
2512 *total
= s390_cost
->dlgr
;
2513 else if (GET_MODE (x
) == DImode
)
2515 rtx right
= XEXP (x
, 1);
2516 if (GET_CODE (right
) == ZERO_EXTEND
) /* 64 by 32 bit division */
2517 *total
= s390_cost
->dlr
;
2518 else /* 64 by 64 bit division */
2519 *total
= s390_cost
->dlgr
;
2521 else if (GET_MODE (x
) == SImode
) /* 32 bit division */
2522 *total
= s390_cost
->dlr
;
2527 if (GET_MODE (x
) == DImode
)
2529 rtx right
= XEXP (x
, 1);
2530 if (GET_CODE (right
) == ZERO_EXTEND
) /* 64 by 32 bit division */
2532 *total
= s390_cost
->dsgfr
;
2534 *total
= s390_cost
->dr
;
2535 else /* 64 by 64 bit division */
2536 *total
= s390_cost
->dsgr
;
2538 else if (GET_MODE (x
) == SImode
) /* 32 bit division */
2539 *total
= s390_cost
->dlr
;
2540 else if (GET_MODE (x
) == SFmode
)
2542 *total
= s390_cost
->debr
;
2544 else if (GET_MODE (x
) == DFmode
)
2546 *total
= s390_cost
->ddbr
;
2548 else if (GET_MODE (x
) == TFmode
)
2550 *total
= s390_cost
->dxbr
;
2555 if (GET_MODE (x
) == SFmode
)
2556 *total
= s390_cost
->sqebr
;
2557 else if (GET_MODE (x
) == DFmode
)
2558 *total
= s390_cost
->sqdbr
;
2560 *total
= s390_cost
->sqxbr
;
2565 if (outer_code
== MULT
|| outer_code
== DIV
|| outer_code
== MOD
2566 || outer_code
== PLUS
|| outer_code
== MINUS
2567 || outer_code
== COMPARE
)
2572 *total
= COSTS_N_INSNS (1);
2573 if (GET_CODE (XEXP (x
, 0)) == AND
2574 && GET_CODE (XEXP (x
, 1)) == CONST_INT
2575 && GET_CODE (XEXP (XEXP (x
, 0), 1)) == CONST_INT
)
2577 rtx op0
= XEXP (XEXP (x
, 0), 0);
2578 rtx op1
= XEXP (XEXP (x
, 0), 1);
2579 rtx op2
= XEXP (x
, 1);
2581 if (memory_operand (op0
, GET_MODE (op0
))
2582 && s390_tm_ccmode (op1
, op2
, 0) != VOIDmode
)
2584 if (register_operand (op0
, GET_MODE (op0
))
2585 && s390_tm_ccmode (op1
, op2
, 1) != VOIDmode
)
2595 /* Return the cost of an address rtx ADDR. */
2598 s390_address_cost (rtx addr
, bool speed ATTRIBUTE_UNUSED
)
2600 struct s390_address ad
;
2601 if (!s390_decompose_address (addr
, &ad
))
2604 return ad
.indx
? COSTS_N_INSNS (1) + 1 : COSTS_N_INSNS (1);
2607 /* If OP is a SYMBOL_REF of a thread-local symbol, return its TLS mode,
2608 otherwise return 0. */
2611 tls_symbolic_operand (rtx op
)
2613 if (GET_CODE (op
) != SYMBOL_REF
)
2615 return SYMBOL_REF_TLS_MODEL (op
);
2618 /* Split DImode access register reference REG (on 64-bit) into its constituent
2619 low and high parts, and store them into LO and HI. Note that gen_lowpart/
2620 gen_highpart cannot be used as they assume all registers are word-sized,
2621 while our access registers have only half that size. */
2624 s390_split_access_reg (rtx reg
, rtx
*lo
, rtx
*hi
)
2626 gcc_assert (TARGET_64BIT
);
2627 gcc_assert (ACCESS_REG_P (reg
));
2628 gcc_assert (GET_MODE (reg
) == DImode
);
2629 gcc_assert (!(REGNO (reg
) & 1));
2631 *lo
= gen_rtx_REG (SImode
, REGNO (reg
) + 1);
2632 *hi
= gen_rtx_REG (SImode
, REGNO (reg
));
2635 /* Return true if OP contains a symbol reference */
2638 symbolic_reference_mentioned_p (rtx op
)
2643 if (GET_CODE (op
) == SYMBOL_REF
|| GET_CODE (op
) == LABEL_REF
)
2646 fmt
= GET_RTX_FORMAT (GET_CODE (op
));
2647 for (i
= GET_RTX_LENGTH (GET_CODE (op
)) - 1; i
>= 0; i
--)
2653 for (j
= XVECLEN (op
, i
) - 1; j
>= 0; j
--)
2654 if (symbolic_reference_mentioned_p (XVECEXP (op
, i
, j
)))
2658 else if (fmt
[i
] == 'e' && symbolic_reference_mentioned_p (XEXP (op
, i
)))
2665 /* Return true if OP contains a reference to a thread-local symbol. */
2668 tls_symbolic_reference_mentioned_p (rtx op
)
2673 if (GET_CODE (op
) == SYMBOL_REF
)
2674 return tls_symbolic_operand (op
);
2676 fmt
= GET_RTX_FORMAT (GET_CODE (op
));
2677 for (i
= GET_RTX_LENGTH (GET_CODE (op
)) - 1; i
>= 0; i
--)
2683 for (j
= XVECLEN (op
, i
) - 1; j
>= 0; j
--)
2684 if (tls_symbolic_reference_mentioned_p (XVECEXP (op
, i
, j
)))
2688 else if (fmt
[i
] == 'e' && tls_symbolic_reference_mentioned_p (XEXP (op
, i
)))
2696 /* Return true if OP is a legitimate general operand when
2697 generating PIC code. It is given that flag_pic is on
2698 and that OP satisfies CONSTANT_P or is a CONST_DOUBLE. */
2701 legitimate_pic_operand_p (rtx op
)
2703 /* Accept all non-symbolic constants. */
2704 if (!SYMBOLIC_CONST (op
))
2707 /* Reject everything else; must be handled
2708 via emit_symbolic_move. */
2712 /* Returns true if the constant value OP is a legitimate general operand.
2713 It is given that OP satisfies CONSTANT_P or is a CONST_DOUBLE. */
2716 s390_legitimate_constant_p (enum machine_mode mode
, rtx op
)
2718 /* Accept all non-symbolic constants. */
2719 if (!SYMBOLIC_CONST (op
))
2722 /* Accept immediate LARL operands. */
2723 if (TARGET_CPU_ZARCH
&& larl_operand (op
, mode
))
2726 /* Thread-local symbols are never legal constants. This is
2727 so that emit_call knows that computing such addresses
2728 might require a function call. */
2729 if (TLS_SYMBOLIC_CONST (op
))
2732 /* In the PIC case, symbolic constants must *not* be
2733 forced into the literal pool. We accept them here,
2734 so that they will be handled by emit_symbolic_move. */
2738 /* All remaining non-PIC symbolic constants are
2739 forced into the literal pool. */
2743 /* Determine if it's legal to put X into the constant pool. This
2744 is not possible if X contains the address of a symbol that is
2745 not constant (TLS) or not known at final link time (PIC). */
2748 s390_cannot_force_const_mem (enum machine_mode mode
, rtx x
)
2750 switch (GET_CODE (x
))
2754 /* Accept all non-symbolic constants. */
2758 /* Labels are OK iff we are non-PIC. */
2759 return flag_pic
!= 0;
2762 /* 'Naked' TLS symbol references are never OK,
2763 non-TLS symbols are OK iff we are non-PIC. */
2764 if (tls_symbolic_operand (x
))
2767 return flag_pic
!= 0;
2770 return s390_cannot_force_const_mem (mode
, XEXP (x
, 0));
2773 return s390_cannot_force_const_mem (mode
, XEXP (x
, 0))
2774 || s390_cannot_force_const_mem (mode
, XEXP (x
, 1));
2777 switch (XINT (x
, 1))
2779 /* Only lt-relative or GOT-relative UNSPECs are OK. */
2780 case UNSPEC_LTREL_OFFSET
:
2788 case UNSPEC_GOTNTPOFF
:
2789 case UNSPEC_INDNTPOFF
:
2792 /* If the literal pool shares the code section, be put
2793 execute template placeholders into the pool as well. */
2795 return TARGET_CPU_ZARCH
;
2807 /* Returns true if the constant value OP is a legitimate general
2808 operand during and after reload. The difference to
2809 legitimate_constant_p is that this function will not accept
2810 a constant that would need to be forced to the literal pool
2811 before it can be used as operand.
2812 This function accepts all constants which can be loaded directly
2816 legitimate_reload_constant_p (rtx op
)
2818 /* Accept la(y) operands. */
2819 if (GET_CODE (op
) == CONST_INT
2820 && DISP_IN_RANGE (INTVAL (op
)))
2823 /* Accept l(g)hi/l(g)fi operands. */
2824 if (GET_CODE (op
) == CONST_INT
2825 && (CONST_OK_FOR_K (INTVAL (op
)) || CONST_OK_FOR_Os (INTVAL (op
))))
2828 /* Accept lliXX operands. */
2830 && GET_CODE (op
) == CONST_INT
2831 && trunc_int_for_mode (INTVAL (op
), word_mode
) == INTVAL (op
)
2832 && s390_single_part (op
, word_mode
, HImode
, 0) >= 0)
2836 && GET_CODE (op
) == CONST_INT
2837 && trunc_int_for_mode (INTVAL (op
), word_mode
) == INTVAL (op
)
2838 && s390_single_part (op
, word_mode
, SImode
, 0) >= 0)
2841 /* Accept larl operands. */
2842 if (TARGET_CPU_ZARCH
2843 && larl_operand (op
, VOIDmode
))
2846 /* Accept floating-point zero operands that fit into a single GPR. */
2847 if (GET_CODE (op
) == CONST_DOUBLE
2848 && s390_float_const_zero_p (op
)
2849 && GET_MODE_SIZE (GET_MODE (op
)) <= UNITS_PER_WORD
)
2852 /* Accept double-word operands that can be split. */
2853 if (GET_CODE (op
) == CONST_INT
2854 && trunc_int_for_mode (INTVAL (op
), word_mode
) != INTVAL (op
))
2856 enum machine_mode dword_mode
= word_mode
== SImode
? DImode
: TImode
;
2857 rtx hi
= operand_subword (op
, 0, 0, dword_mode
);
2858 rtx lo
= operand_subword (op
, 1, 0, dword_mode
);
2859 return legitimate_reload_constant_p (hi
)
2860 && legitimate_reload_constant_p (lo
);
2863 /* Everything else cannot be handled without reload. */
2867 /* Returns true if the constant value OP is a legitimate fp operand
2868 during and after reload.
2869 This function accepts all constants which can be loaded directly
2873 legitimate_reload_fp_constant_p (rtx op
)
2875 /* Accept floating-point zero operands if the load zero instruction
2878 && GET_CODE (op
) == CONST_DOUBLE
2879 && s390_float_const_zero_p (op
))
2885 /* Given an rtx OP being reloaded into a reg required to be in class RCLASS,
2886 return the class of reg to actually use. */
2889 s390_preferred_reload_class (rtx op
, reg_class_t rclass
)
2891 switch (GET_CODE (op
))
2893 /* Constants we cannot reload into general registers
2894 must be forced into the literal pool. */
2897 if (reg_class_subset_p (GENERAL_REGS
, rclass
)
2898 && legitimate_reload_constant_p (op
))
2899 return GENERAL_REGS
;
2900 else if (reg_class_subset_p (ADDR_REGS
, rclass
)
2901 && legitimate_reload_constant_p (op
))
2903 else if (reg_class_subset_p (FP_REGS
, rclass
)
2904 && legitimate_reload_fp_constant_p (op
))
2908 /* If a symbolic constant or a PLUS is reloaded,
2909 it is most likely being used as an address, so
2910 prefer ADDR_REGS. If 'class' is not a superset
2911 of ADDR_REGS, e.g. FP_REGS, reject this reload. */
2915 if (!legitimate_reload_constant_p (op
))
2919 /* load address will be used. */
2920 if (reg_class_subset_p (ADDR_REGS
, rclass
))
2932 /* Return true if ADDR is SYMBOL_REF + addend with addend being a
2933 multiple of ALIGNMENT and the SYMBOL_REF being naturally
2937 s390_check_symref_alignment (rtx addr
, HOST_WIDE_INT alignment
)
2939 HOST_WIDE_INT addend
;
2942 if (!s390_symref_operand_p (addr
, &symref
, &addend
))
2945 return (!SYMBOL_REF_NOT_NATURALLY_ALIGNED_P (symref
)
2946 && !(addend
& (alignment
- 1)));
2949 /* ADDR is moved into REG using larl. If ADDR isn't a valid larl
2950 operand SCRATCH is used to reload the even part of the address and
2954 s390_reload_larl_operand (rtx reg
, rtx addr
, rtx scratch
)
2956 HOST_WIDE_INT addend
;
2959 if (!s390_symref_operand_p (addr
, &symref
, &addend
))
2963 /* Easy case. The addend is even so larl will do fine. */
2964 emit_move_insn (reg
, addr
);
2967 /* We can leave the scratch register untouched if the target
2968 register is a valid base register. */
2969 if (REGNO (reg
) < FIRST_PSEUDO_REGISTER
2970 && REGNO_REG_CLASS (REGNO (reg
)) == ADDR_REGS
)
2973 gcc_assert (REGNO (scratch
) < FIRST_PSEUDO_REGISTER
);
2974 gcc_assert (REGNO_REG_CLASS (REGNO (scratch
)) == ADDR_REGS
);
2977 emit_move_insn (scratch
,
2978 gen_rtx_CONST (Pmode
,
2979 gen_rtx_PLUS (Pmode
, symref
,
2980 GEN_INT (addend
- 1))));
2982 emit_move_insn (scratch
, symref
);
2984 /* Increment the address using la in order to avoid clobbering cc. */
2985 emit_move_insn (reg
, gen_rtx_PLUS (Pmode
, scratch
, const1_rtx
));
2989 /* Generate what is necessary to move between REG and MEM using
2990 SCRATCH. The direction is given by TOMEM. */
2993 s390_reload_symref_address (rtx reg
, rtx mem
, rtx scratch
, bool tomem
)
2995 /* Reload might have pulled a constant out of the literal pool.
2996 Force it back in. */
2997 if (CONST_INT_P (mem
) || GET_CODE (mem
) == CONST_DOUBLE
2998 || GET_CODE (mem
) == CONST
)
2999 mem
= force_const_mem (GET_MODE (reg
), mem
);
3001 gcc_assert (MEM_P (mem
));
3003 /* For a load from memory we can leave the scratch register
3004 untouched if the target register is a valid base register. */
3006 && REGNO (reg
) < FIRST_PSEUDO_REGISTER
3007 && REGNO_REG_CLASS (REGNO (reg
)) == ADDR_REGS
3008 && GET_MODE (reg
) == GET_MODE (scratch
))
3011 /* Load address into scratch register. Since we can't have a
3012 secondary reload for a secondary reload we have to cover the case
3013 where larl would need a secondary reload here as well. */
3014 s390_reload_larl_operand (scratch
, XEXP (mem
, 0), scratch
);
3016 /* Now we can use a standard load/store to do the move. */
3018 emit_move_insn (replace_equiv_address (mem
, scratch
), reg
);
3020 emit_move_insn (reg
, replace_equiv_address (mem
, scratch
));
3023 /* Inform reload about cases where moving X with a mode MODE to a register in
3024 RCLASS requires an extra scratch or immediate register. Return the class
3025 needed for the immediate register. */
3028 s390_secondary_reload (bool in_p
, rtx x
, reg_class_t rclass_i
,
3029 enum machine_mode mode
, secondary_reload_info
*sri
)
3031 enum reg_class rclass
= (enum reg_class
) rclass_i
;
3033 /* Intermediate register needed. */
3034 if (reg_classes_intersect_p (CC_REGS
, rclass
))
3035 return GENERAL_REGS
;
3039 HOST_WIDE_INT offset
;
3042 /* On z10 several optimizer steps may generate larl operands with
3045 && s390_symref_operand_p (x
, &symref
, &offset
)
3047 && !SYMBOL_REF_ALIGN1_P (symref
)
3048 && (offset
& 1) == 1)
3049 sri
->icode
= ((mode
== DImode
) ? CODE_FOR_reloaddi_larl_odd_addend_z10
3050 : CODE_FOR_reloadsi_larl_odd_addend_z10
);
3052 /* On z10 we need a scratch register when moving QI, TI or floating
3053 point mode values from or to a memory location with a SYMBOL_REF
3054 or if the symref addend of a SI or DI move is not aligned to the
3055 width of the access. */
3057 && s390_symref_operand_p (XEXP (x
, 0), NULL
, NULL
)
3058 && (mode
== QImode
|| mode
== TImode
|| FLOAT_MODE_P (mode
)
3059 || (!TARGET_ZARCH
&& mode
== DImode
)
3060 || ((mode
== HImode
|| mode
== SImode
|| mode
== DImode
)
3061 && (!s390_check_symref_alignment (XEXP (x
, 0),
3062 GET_MODE_SIZE (mode
))))))
3064 #define __SECONDARY_RELOAD_CASE(M,m) \
3067 sri->icode = in_p ? CODE_FOR_reload##m##di_toreg_z10 : \
3068 CODE_FOR_reload##m##di_tomem_z10; \
3070 sri->icode = in_p ? CODE_FOR_reload##m##si_toreg_z10 : \
3071 CODE_FOR_reload##m##si_tomem_z10; \
3074 switch (GET_MODE (x
))
3076 __SECONDARY_RELOAD_CASE (QI
, qi
);
3077 __SECONDARY_RELOAD_CASE (HI
, hi
);
3078 __SECONDARY_RELOAD_CASE (SI
, si
);
3079 __SECONDARY_RELOAD_CASE (DI
, di
);
3080 __SECONDARY_RELOAD_CASE (TI
, ti
);
3081 __SECONDARY_RELOAD_CASE (SF
, sf
);
3082 __SECONDARY_RELOAD_CASE (DF
, df
);
3083 __SECONDARY_RELOAD_CASE (TF
, tf
);
3084 __SECONDARY_RELOAD_CASE (SD
, sd
);
3085 __SECONDARY_RELOAD_CASE (DD
, dd
);
3086 __SECONDARY_RELOAD_CASE (TD
, td
);
3091 #undef __SECONDARY_RELOAD_CASE
3095 /* We need a scratch register when loading a PLUS expression which
3096 is not a legitimate operand of the LOAD ADDRESS instruction. */
3097 if (in_p
&& s390_plus_operand (x
, mode
))
3098 sri
->icode
= (TARGET_64BIT
?
3099 CODE_FOR_reloaddi_plus
: CODE_FOR_reloadsi_plus
);
3101 /* Performing a multiword move from or to memory we have to make sure the
3102 second chunk in memory is addressable without causing a displacement
3103 overflow. If that would be the case we calculate the address in
3104 a scratch register. */
3106 && GET_CODE (XEXP (x
, 0)) == PLUS
3107 && GET_CODE (XEXP (XEXP (x
, 0), 1)) == CONST_INT
3108 && !DISP_IN_RANGE (INTVAL (XEXP (XEXP (x
, 0), 1))
3109 + GET_MODE_SIZE (mode
) - 1))
3111 /* For GENERAL_REGS a displacement overflow is no problem if occurring
3112 in a s_operand address since we may fallback to lm/stm. So we only
3113 have to care about overflows in the b+i+d case. */
3114 if ((reg_classes_intersect_p (GENERAL_REGS
, rclass
)
3115 && s390_class_max_nregs (GENERAL_REGS
, mode
) > 1
3116 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == PLUS
)
3117 /* For FP_REGS no lm/stm is available so this check is triggered
3118 for displacement overflows in b+i+d and b+d like addresses. */
3119 || (reg_classes_intersect_p (FP_REGS
, rclass
)
3120 && s390_class_max_nregs (FP_REGS
, mode
) > 1))
3123 sri
->icode
= (TARGET_64BIT
?
3124 CODE_FOR_reloaddi_nonoffmem_in
:
3125 CODE_FOR_reloadsi_nonoffmem_in
);
3127 sri
->icode
= (TARGET_64BIT
?
3128 CODE_FOR_reloaddi_nonoffmem_out
:
3129 CODE_FOR_reloadsi_nonoffmem_out
);
3133 /* A scratch address register is needed when a symbolic constant is
3134 copied to r0 compiling with -fPIC. In other cases the target
3135 register might be used as temporary (see legitimize_pic_address). */
3136 if (in_p
&& SYMBOLIC_CONST (x
) && flag_pic
== 2 && rclass
!= ADDR_REGS
)
3137 sri
->icode
= (TARGET_64BIT
?
3138 CODE_FOR_reloaddi_PIC_addr
:
3139 CODE_FOR_reloadsi_PIC_addr
);
3141 /* Either scratch or no register needed. */
3145 /* Generate code to load SRC, which is PLUS that is not a
3146 legitimate operand for the LA instruction, into TARGET.
3147 SCRATCH may be used as scratch register. */
3150 s390_expand_plus_operand (rtx target
, rtx src
,
3154 struct s390_address ad
;
3156 /* src must be a PLUS; get its two operands. */
3157 gcc_assert (GET_CODE (src
) == PLUS
);
3158 gcc_assert (GET_MODE (src
) == Pmode
);
3160 /* Check if any of the two operands is already scheduled
3161 for replacement by reload. This can happen e.g. when
3162 float registers occur in an address. */
3163 sum1
= find_replacement (&XEXP (src
, 0));
3164 sum2
= find_replacement (&XEXP (src
, 1));
3165 src
= gen_rtx_PLUS (Pmode
, sum1
, sum2
);
3167 /* If the address is already strictly valid, there's nothing to do. */
3168 if (!s390_decompose_address (src
, &ad
)
3169 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
3170 || (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
))))
3172 /* Otherwise, one of the operands cannot be an address register;
3173 we reload its value into the scratch register. */
3174 if (true_regnum (sum1
) < 1 || true_regnum (sum1
) > 15)
3176 emit_move_insn (scratch
, sum1
);
3179 if (true_regnum (sum2
) < 1 || true_regnum (sum2
) > 15)
3181 emit_move_insn (scratch
, sum2
);
3185 /* According to the way these invalid addresses are generated
3186 in reload.c, it should never happen (at least on s390) that
3187 *neither* of the PLUS components, after find_replacements
3188 was applied, is an address register. */
3189 if (sum1
== scratch
&& sum2
== scratch
)
3195 src
= gen_rtx_PLUS (Pmode
, sum1
, sum2
);
3198 /* Emit the LOAD ADDRESS pattern. Note that reload of PLUS
3199 is only ever performed on addresses, so we can mark the
3200 sum as legitimate for LA in any case. */
3201 s390_load_address (target
, src
);
3205 /* Return true if ADDR is a valid memory address.
3206 STRICT specifies whether strict register checking applies. */
3209 s390_legitimate_address_p (enum machine_mode mode
, rtx addr
, bool strict
)
3211 struct s390_address ad
;
3214 && larl_operand (addr
, VOIDmode
)
3215 && (mode
== VOIDmode
3216 || s390_check_symref_alignment (addr
, GET_MODE_SIZE (mode
))))
3219 if (!s390_decompose_address (addr
, &ad
))
3224 if (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
3227 if (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
)))
3233 && !(REGNO (ad
.base
) >= FIRST_PSEUDO_REGISTER
3234 || REGNO_REG_CLASS (REGNO (ad
.base
)) == ADDR_REGS
))
3238 && !(REGNO (ad
.indx
) >= FIRST_PSEUDO_REGISTER
3239 || REGNO_REG_CLASS (REGNO (ad
.indx
)) == ADDR_REGS
))
3245 /* Return true if OP is a valid operand for the LA instruction.
3246 In 31-bit, we need to prove that the result is used as an
3247 address, as LA performs only a 31-bit addition. */
3250 legitimate_la_operand_p (rtx op
)
3252 struct s390_address addr
;
3253 if (!s390_decompose_address (op
, &addr
))
3256 return (TARGET_64BIT
|| addr
.pointer
);
3259 /* Return true if it is valid *and* preferable to use LA to
3260 compute the sum of OP1 and OP2. */
3263 preferred_la_operand_p (rtx op1
, rtx op2
)
3265 struct s390_address addr
;
3267 if (op2
!= const0_rtx
)
3268 op1
= gen_rtx_PLUS (Pmode
, op1
, op2
);
3270 if (!s390_decompose_address (op1
, &addr
))
3272 if (addr
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (addr
.base
)))
3274 if (addr
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (addr
.indx
)))
3277 /* Avoid LA instructions with index register on z196; it is
3278 preferable to use regular add instructions when possible. */
3279 if (addr
.indx
&& s390_tune
== PROCESSOR_2817_Z196
)
3282 if (!TARGET_64BIT
&& !addr
.pointer
)
3288 if ((addr
.base
&& REG_P (addr
.base
) && REG_POINTER (addr
.base
))
3289 || (addr
.indx
&& REG_P (addr
.indx
) && REG_POINTER (addr
.indx
)))
3295 /* Emit a forced load-address operation to load SRC into DST.
3296 This will use the LOAD ADDRESS instruction even in situations
3297 where legitimate_la_operand_p (SRC) returns false. */
3300 s390_load_address (rtx dst
, rtx src
)
3303 emit_move_insn (dst
, src
);
3305 emit_insn (gen_force_la_31 (dst
, src
));
3308 /* Return a legitimate reference for ORIG (an address) using the
3309 register REG. If REG is 0, a new pseudo is generated.
3311 There are two types of references that must be handled:
3313 1. Global data references must load the address from the GOT, via
3314 the PIC reg. An insn is emitted to do this load, and the reg is
3317 2. Static data references, constant pool addresses, and code labels
3318 compute the address as an offset from the GOT, whose base is in
3319 the PIC reg. Static data objects have SYMBOL_FLAG_LOCAL set to
3320 differentiate them from global data objects. The returned
3321 address is the PIC reg + an unspec constant.
3323 TARGET_LEGITIMIZE_ADDRESS_P rejects symbolic references unless the PIC
3324 reg also appears in the address. */
3327 legitimize_pic_address (rtx orig
, rtx reg
)
3333 gcc_assert (!TLS_SYMBOLIC_CONST (addr
));
3335 if (GET_CODE (addr
) == LABEL_REF
3336 || (GET_CODE (addr
) == SYMBOL_REF
&& SYMBOL_REF_LOCAL_P (addr
)))
3338 /* This is a local symbol. */
3339 if (TARGET_CPU_ZARCH
&& larl_operand (addr
, VOIDmode
))
3341 /* Access local symbols PC-relative via LARL.
3342 This is the same as in the non-PIC case, so it is
3343 handled automatically ... */
3347 /* Access local symbols relative to the GOT. */
3349 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3351 if (reload_in_progress
|| reload_completed
)
3352 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3354 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTOFF
);
3355 addr
= gen_rtx_CONST (Pmode
, addr
);
3356 addr
= force_const_mem (Pmode
, addr
);
3357 emit_move_insn (temp
, addr
);
3359 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3362 s390_load_address (reg
, new_rtx
);
3367 else if (GET_CODE (addr
) == SYMBOL_REF
)
3370 reg
= gen_reg_rtx (Pmode
);
3374 /* Assume GOT offset < 4k. This is handled the same way
3375 in both 31- and 64-bit code (@GOT). */
3377 if (reload_in_progress
|| reload_completed
)
3378 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3380 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOT
);
3381 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3382 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, new_rtx
);
3383 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3384 emit_move_insn (reg
, new_rtx
);
3387 else if (TARGET_CPU_ZARCH
)
3389 /* If the GOT offset might be >= 4k, we determine the position
3390 of the GOT entry via a PC-relative LARL (@GOTENT). */
3392 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3394 gcc_assert (REGNO (temp
) >= FIRST_PSEUDO_REGISTER
3395 || REGNO_REG_CLASS (REGNO (temp
)) == ADDR_REGS
);
3397 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTENT
);
3398 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3399 emit_move_insn (temp
, new_rtx
);
3401 new_rtx
= gen_const_mem (Pmode
, temp
);
3402 emit_move_insn (reg
, new_rtx
);
3407 /* If the GOT offset might be >= 4k, we have to load it
3408 from the literal pool (@GOT). */
3410 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3412 gcc_assert (REGNO (temp
) >= FIRST_PSEUDO_REGISTER
3413 || REGNO_REG_CLASS (REGNO (temp
)) == ADDR_REGS
);
3415 if (reload_in_progress
|| reload_completed
)
3416 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3418 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOT
);
3419 addr
= gen_rtx_CONST (Pmode
, addr
);
3420 addr
= force_const_mem (Pmode
, addr
);
3421 emit_move_insn (temp
, addr
);
3423 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3424 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3425 emit_move_insn (reg
, new_rtx
);
3431 if (GET_CODE (addr
) == CONST
)
3433 addr
= XEXP (addr
, 0);
3434 if (GET_CODE (addr
) == UNSPEC
)
3436 gcc_assert (XVECLEN (addr
, 0) == 1);
3437 switch (XINT (addr
, 1))
3439 /* If someone moved a GOT-relative UNSPEC
3440 out of the literal pool, force them back in. */
3443 new_rtx
= force_const_mem (Pmode
, orig
);
3446 /* @GOT is OK as is if small. */
3449 new_rtx
= force_const_mem (Pmode
, orig
);
3452 /* @GOTENT is OK as is. */
3456 /* @PLT is OK as is on 64-bit, must be converted to
3457 GOT-relative @PLTOFF on 31-bit. */
3459 if (!TARGET_CPU_ZARCH
)
3461 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3463 if (reload_in_progress
|| reload_completed
)
3464 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3466 addr
= XVECEXP (addr
, 0, 0);
3467 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
),
3469 addr
= gen_rtx_CONST (Pmode
, addr
);
3470 addr
= force_const_mem (Pmode
, addr
);
3471 emit_move_insn (temp
, addr
);
3473 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3476 s390_load_address (reg
, new_rtx
);
3482 /* Everything else cannot happen. */
3488 gcc_assert (GET_CODE (addr
) == PLUS
);
3490 if (GET_CODE (addr
) == PLUS
)
3492 rtx op0
= XEXP (addr
, 0), op1
= XEXP (addr
, 1);
3494 gcc_assert (!TLS_SYMBOLIC_CONST (op0
));
3495 gcc_assert (!TLS_SYMBOLIC_CONST (op1
));
3497 /* Check first to see if this is a constant offset
3498 from a local symbol reference. */
3499 if ((GET_CODE (op0
) == LABEL_REF
3500 || (GET_CODE (op0
) == SYMBOL_REF
&& SYMBOL_REF_LOCAL_P (op0
)))
3501 && GET_CODE (op1
) == CONST_INT
)
3503 if (TARGET_CPU_ZARCH
3504 && larl_operand (op0
, VOIDmode
)
3505 && INTVAL (op1
) < (HOST_WIDE_INT
)1 << 31
3506 && INTVAL (op1
) >= -((HOST_WIDE_INT
)1 << 31))
3508 if (INTVAL (op1
) & 1)
3510 /* LARL can't handle odd offsets, so emit a
3511 pair of LARL and LA. */
3512 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3514 if (!DISP_IN_RANGE (INTVAL (op1
)))
3516 HOST_WIDE_INT even
= INTVAL (op1
) - 1;
3517 op0
= gen_rtx_PLUS (Pmode
, op0
, GEN_INT (even
));
3518 op0
= gen_rtx_CONST (Pmode
, op0
);
3522 emit_move_insn (temp
, op0
);
3523 new_rtx
= gen_rtx_PLUS (Pmode
, temp
, op1
);
3527 s390_load_address (reg
, new_rtx
);
3533 /* If the offset is even, we can just use LARL.
3534 This will happen automatically. */
3539 /* Access local symbols relative to the GOT. */
3541 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3543 if (reload_in_progress
|| reload_completed
)
3544 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3546 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op0
),
3548 addr
= gen_rtx_PLUS (Pmode
, addr
, op1
);
3549 addr
= gen_rtx_CONST (Pmode
, addr
);
3550 addr
= force_const_mem (Pmode
, addr
);
3551 emit_move_insn (temp
, addr
);
3553 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3556 s390_load_address (reg
, new_rtx
);
3562 /* Now, check whether it is a GOT relative symbol plus offset
3563 that was pulled out of the literal pool. Force it back in. */
3565 else if (GET_CODE (op0
) == UNSPEC
3566 && GET_CODE (op1
) == CONST_INT
3567 && XINT (op0
, 1) == UNSPEC_GOTOFF
)
3569 gcc_assert (XVECLEN (op0
, 0) == 1);
3571 new_rtx
= force_const_mem (Pmode
, orig
);
3574 /* Otherwise, compute the sum. */
3577 base
= legitimize_pic_address (XEXP (addr
, 0), reg
);
3578 new_rtx
= legitimize_pic_address (XEXP (addr
, 1),
3579 base
== reg
? NULL_RTX
: reg
);
3580 if (GET_CODE (new_rtx
) == CONST_INT
)
3581 new_rtx
= plus_constant (Pmode
, base
, INTVAL (new_rtx
));
3584 if (GET_CODE (new_rtx
) == PLUS
&& CONSTANT_P (XEXP (new_rtx
, 1)))
3586 base
= gen_rtx_PLUS (Pmode
, base
, XEXP (new_rtx
, 0));
3587 new_rtx
= XEXP (new_rtx
, 1);
3589 new_rtx
= gen_rtx_PLUS (Pmode
, base
, new_rtx
);
3592 if (GET_CODE (new_rtx
) == CONST
)
3593 new_rtx
= XEXP (new_rtx
, 0);
3594 new_rtx
= force_operand (new_rtx
, 0);
3601 /* Load the thread pointer into a register. */
3604 s390_get_thread_pointer (void)
3606 rtx tp
= gen_reg_rtx (Pmode
);
3608 emit_move_insn (tp
, gen_rtx_REG (Pmode
, TP_REGNUM
));
3609 mark_reg_pointer (tp
, BITS_PER_WORD
);
3614 /* Emit a tls call insn. The call target is the SYMBOL_REF stored
3615 in s390_tls_symbol which always refers to __tls_get_offset.
3616 The returned offset is written to RESULT_REG and an USE rtx is
3617 generated for TLS_CALL. */
3619 static GTY(()) rtx s390_tls_symbol
;
3622 s390_emit_tls_call_insn (rtx result_reg
, rtx tls_call
)
3627 emit_insn (s390_load_got ());
3629 if (!s390_tls_symbol
)
3630 s390_tls_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "__tls_get_offset");
3632 insn
= s390_emit_call (s390_tls_symbol
, tls_call
, result_reg
,
3633 gen_rtx_REG (Pmode
, RETURN_REGNUM
));
3635 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), result_reg
);
3636 RTL_CONST_CALL_P (insn
) = 1;
3639 /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute
3640 this (thread-local) address. REG may be used as temporary. */
3643 legitimize_tls_address (rtx addr
, rtx reg
)
3645 rtx new_rtx
, tls_call
, temp
, base
, r2
, insn
;
3647 if (GET_CODE (addr
) == SYMBOL_REF
)
3648 switch (tls_symbolic_operand (addr
))
3650 case TLS_MODEL_GLOBAL_DYNAMIC
:
3652 r2
= gen_rtx_REG (Pmode
, 2);
3653 tls_call
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_TLSGD
);
3654 new_rtx
= gen_rtx_CONST (Pmode
, tls_call
);
3655 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3656 emit_move_insn (r2
, new_rtx
);
3657 s390_emit_tls_call_insn (r2
, tls_call
);
3658 insn
= get_insns ();
3661 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_NTPOFF
);
3662 temp
= gen_reg_rtx (Pmode
);
3663 emit_libcall_block (insn
, temp
, r2
, new_rtx
);
3665 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3668 s390_load_address (reg
, new_rtx
);
3673 case TLS_MODEL_LOCAL_DYNAMIC
:
3675 r2
= gen_rtx_REG (Pmode
, 2);
3676 tls_call
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, const0_rtx
), UNSPEC_TLSLDM
);
3677 new_rtx
= gen_rtx_CONST (Pmode
, tls_call
);
3678 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3679 emit_move_insn (r2
, new_rtx
);
3680 s390_emit_tls_call_insn (r2
, tls_call
);
3681 insn
= get_insns ();
3684 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, const0_rtx
), UNSPEC_TLSLDM_NTPOFF
);
3685 temp
= gen_reg_rtx (Pmode
);
3686 emit_libcall_block (insn
, temp
, r2
, new_rtx
);
3688 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3689 base
= gen_reg_rtx (Pmode
);
3690 s390_load_address (base
, new_rtx
);
3692 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_DTPOFF
);
3693 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3694 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3695 temp
= gen_reg_rtx (Pmode
);
3696 emit_move_insn (temp
, new_rtx
);
3698 new_rtx
= gen_rtx_PLUS (Pmode
, base
, temp
);
3701 s390_load_address (reg
, new_rtx
);
3706 case TLS_MODEL_INITIAL_EXEC
:
3709 /* Assume GOT offset < 4k. This is handled the same way
3710 in both 31- and 64-bit code. */
3712 if (reload_in_progress
|| reload_completed
)
3713 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3715 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTNTPOFF
);
3716 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3717 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, new_rtx
);
3718 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3719 temp
= gen_reg_rtx (Pmode
);
3720 emit_move_insn (temp
, new_rtx
);
3722 else if (TARGET_CPU_ZARCH
)
3724 /* If the GOT offset might be >= 4k, we determine the position
3725 of the GOT entry via a PC-relative LARL. */
3727 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_INDNTPOFF
);
3728 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3729 temp
= gen_reg_rtx (Pmode
);
3730 emit_move_insn (temp
, new_rtx
);
3732 new_rtx
= gen_const_mem (Pmode
, temp
);
3733 temp
= gen_reg_rtx (Pmode
);
3734 emit_move_insn (temp
, new_rtx
);
3738 /* If the GOT offset might be >= 4k, we have to load it
3739 from the literal pool. */
3741 if (reload_in_progress
|| reload_completed
)
3742 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3744 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTNTPOFF
);
3745 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3746 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3747 temp
= gen_reg_rtx (Pmode
);
3748 emit_move_insn (temp
, new_rtx
);
3750 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3751 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3753 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, new_rtx
, addr
), UNSPEC_TLS_LOAD
);
3754 temp
= gen_reg_rtx (Pmode
);
3755 emit_insn (gen_rtx_SET (Pmode
, temp
, new_rtx
));
3759 /* In position-dependent code, load the absolute address of
3760 the GOT entry from the literal pool. */
3762 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_INDNTPOFF
);
3763 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3764 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3765 temp
= gen_reg_rtx (Pmode
);
3766 emit_move_insn (temp
, new_rtx
);
3769 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3770 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, new_rtx
, addr
), UNSPEC_TLS_LOAD
);
3771 temp
= gen_reg_rtx (Pmode
);
3772 emit_insn (gen_rtx_SET (Pmode
, temp
, new_rtx
));
3775 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3778 s390_load_address (reg
, new_rtx
);
3783 case TLS_MODEL_LOCAL_EXEC
:
3784 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_NTPOFF
);
3785 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3786 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3787 temp
= gen_reg_rtx (Pmode
);
3788 emit_move_insn (temp
, new_rtx
);
3790 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3793 s390_load_address (reg
, new_rtx
);
3802 else if (GET_CODE (addr
) == CONST
&& GET_CODE (XEXP (addr
, 0)) == UNSPEC
)
3804 switch (XINT (XEXP (addr
, 0), 1))
3806 case UNSPEC_INDNTPOFF
:
3807 gcc_assert (TARGET_CPU_ZARCH
);
3816 else if (GET_CODE (addr
) == CONST
&& GET_CODE (XEXP (addr
, 0)) == PLUS
3817 && GET_CODE (XEXP (XEXP (addr
, 0), 1)) == CONST_INT
)
3819 new_rtx
= XEXP (XEXP (addr
, 0), 0);
3820 if (GET_CODE (new_rtx
) != SYMBOL_REF
)
3821 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3823 new_rtx
= legitimize_tls_address (new_rtx
, reg
);
3824 new_rtx
= plus_constant (Pmode
, new_rtx
,
3825 INTVAL (XEXP (XEXP (addr
, 0), 1)));
3826 new_rtx
= force_operand (new_rtx
, 0);
3830 gcc_unreachable (); /* for now ... */
3835 /* Emit insns making the address in operands[1] valid for a standard
3836 move to operands[0]. operands[1] is replaced by an address which
3837 should be used instead of the former RTX to emit the move
3841 emit_symbolic_move (rtx
*operands
)
3843 rtx temp
= !can_create_pseudo_p () ? operands
[0] : gen_reg_rtx (Pmode
);
3845 if (GET_CODE (operands
[0]) == MEM
)
3846 operands
[1] = force_reg (Pmode
, operands
[1]);
3847 else if (TLS_SYMBOLIC_CONST (operands
[1]))
3848 operands
[1] = legitimize_tls_address (operands
[1], temp
);
3850 operands
[1] = legitimize_pic_address (operands
[1], temp
);
3853 /* Try machine-dependent ways of modifying an illegitimate address X
3854 to be legitimate. If we find one, return the new, valid address.
3856 OLDX is the address as it was before break_out_memory_refs was called.
3857 In some cases it is useful to look at this to decide what needs to be done.
3859 MODE is the mode of the operand pointed to by X.
3861 When -fpic is used, special handling is needed for symbolic references.
3862 See comments by legitimize_pic_address for details. */
3865 s390_legitimize_address (rtx x
, rtx oldx ATTRIBUTE_UNUSED
,
3866 enum machine_mode mode ATTRIBUTE_UNUSED
)
3868 rtx constant_term
= const0_rtx
;
3870 if (TLS_SYMBOLIC_CONST (x
))
3872 x
= legitimize_tls_address (x
, 0);
3874 if (s390_legitimate_address_p (mode
, x
, FALSE
))
3877 else if (GET_CODE (x
) == PLUS
3878 && (TLS_SYMBOLIC_CONST (XEXP (x
, 0))
3879 || TLS_SYMBOLIC_CONST (XEXP (x
, 1))))
3885 if (SYMBOLIC_CONST (x
)
3886 || (GET_CODE (x
) == PLUS
3887 && (SYMBOLIC_CONST (XEXP (x
, 0))
3888 || SYMBOLIC_CONST (XEXP (x
, 1)))))
3889 x
= legitimize_pic_address (x
, 0);
3891 if (s390_legitimate_address_p (mode
, x
, FALSE
))
3895 x
= eliminate_constant_term (x
, &constant_term
);
3897 /* Optimize loading of large displacements by splitting them
3898 into the multiple of 4K and the rest; this allows the
3899 former to be CSE'd if possible.
3901 Don't do this if the displacement is added to a register
3902 pointing into the stack frame, as the offsets will
3903 change later anyway. */
3905 if (GET_CODE (constant_term
) == CONST_INT
3906 && !TARGET_LONG_DISPLACEMENT
3907 && !DISP_IN_RANGE (INTVAL (constant_term
))
3908 && !(REG_P (x
) && REGNO_PTR_FRAME_P (REGNO (x
))))
3910 HOST_WIDE_INT lower
= INTVAL (constant_term
) & 0xfff;
3911 HOST_WIDE_INT upper
= INTVAL (constant_term
) ^ lower
;
3913 rtx temp
= gen_reg_rtx (Pmode
);
3914 rtx val
= force_operand (GEN_INT (upper
), temp
);
3916 emit_move_insn (temp
, val
);
3918 x
= gen_rtx_PLUS (Pmode
, x
, temp
);
3919 constant_term
= GEN_INT (lower
);
3922 if (GET_CODE (x
) == PLUS
)
3924 if (GET_CODE (XEXP (x
, 0)) == REG
)
3926 rtx temp
= gen_reg_rtx (Pmode
);
3927 rtx val
= force_operand (XEXP (x
, 1), temp
);
3929 emit_move_insn (temp
, val
);
3931 x
= gen_rtx_PLUS (Pmode
, XEXP (x
, 0), temp
);
3934 else if (GET_CODE (XEXP (x
, 1)) == REG
)
3936 rtx temp
= gen_reg_rtx (Pmode
);
3937 rtx val
= force_operand (XEXP (x
, 0), temp
);
3939 emit_move_insn (temp
, val
);
3941 x
= gen_rtx_PLUS (Pmode
, temp
, XEXP (x
, 1));
3945 if (constant_term
!= const0_rtx
)
3946 x
= gen_rtx_PLUS (Pmode
, x
, constant_term
);
3951 /* Try a machine-dependent way of reloading an illegitimate address AD
3952 operand. If we find one, push the reload and return the new address.
3954 MODE is the mode of the enclosing MEM. OPNUM is the operand number
3955 and TYPE is the reload type of the current reload. */
3958 legitimize_reload_address (rtx ad
, enum machine_mode mode ATTRIBUTE_UNUSED
,
3959 int opnum
, int type
)
3961 if (!optimize
|| TARGET_LONG_DISPLACEMENT
)
3964 if (GET_CODE (ad
) == PLUS
)
3966 rtx tem
= simplify_binary_operation (PLUS
, Pmode
,
3967 XEXP (ad
, 0), XEXP (ad
, 1));
3972 if (GET_CODE (ad
) == PLUS
3973 && GET_CODE (XEXP (ad
, 0)) == REG
3974 && GET_CODE (XEXP (ad
, 1)) == CONST_INT
3975 && !DISP_IN_RANGE (INTVAL (XEXP (ad
, 1))))
3977 HOST_WIDE_INT lower
= INTVAL (XEXP (ad
, 1)) & 0xfff;
3978 HOST_WIDE_INT upper
= INTVAL (XEXP (ad
, 1)) ^ lower
;
3979 rtx cst
, tem
, new_rtx
;
3981 cst
= GEN_INT (upper
);
3982 if (!legitimate_reload_constant_p (cst
))
3983 cst
= force_const_mem (Pmode
, cst
);
3985 tem
= gen_rtx_PLUS (Pmode
, XEXP (ad
, 0), cst
);
3986 new_rtx
= gen_rtx_PLUS (Pmode
, tem
, GEN_INT (lower
));
3988 push_reload (XEXP (tem
, 1), 0, &XEXP (tem
, 1), 0,
3989 BASE_REG_CLASS
, Pmode
, VOIDmode
, 0, 0,
3990 opnum
, (enum reload_type
) type
);
3997 /* Emit code to move LEN bytes from DST to SRC. */
4000 s390_expand_movmem (rtx dst
, rtx src
, rtx len
)
4002 /* When tuning for z10 or higher we rely on the Glibc functions to
4003 do the right thing. Only for constant lengths below 64k we will
4004 generate inline code. */
4005 if (s390_tune
>= PROCESSOR_2097_Z10
4006 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > (1<<16)))
4009 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) >= 0 && INTVAL (len
) <= 256)
4011 if (INTVAL (len
) > 0)
4012 emit_insn (gen_movmem_short (dst
, src
, GEN_INT (INTVAL (len
) - 1)));
4015 else if (TARGET_MVCLE
)
4017 emit_insn (gen_movmem_long (dst
, src
, convert_to_mode (Pmode
, len
, 1)));
4022 rtx dst_addr
, src_addr
, count
, blocks
, temp
;
4023 rtx loop_start_label
= gen_label_rtx ();
4024 rtx loop_end_label
= gen_label_rtx ();
4025 rtx end_label
= gen_label_rtx ();
4026 enum machine_mode mode
;
4028 mode
= GET_MODE (len
);
4029 if (mode
== VOIDmode
)
4032 dst_addr
= gen_reg_rtx (Pmode
);
4033 src_addr
= gen_reg_rtx (Pmode
);
4034 count
= gen_reg_rtx (mode
);
4035 blocks
= gen_reg_rtx (mode
);
4037 convert_move (count
, len
, 1);
4038 emit_cmp_and_jump_insns (count
, const0_rtx
,
4039 EQ
, NULL_RTX
, mode
, 1, end_label
);
4041 emit_move_insn (dst_addr
, force_operand (XEXP (dst
, 0), NULL_RTX
));
4042 emit_move_insn (src_addr
, force_operand (XEXP (src
, 0), NULL_RTX
));
4043 dst
= change_address (dst
, VOIDmode
, dst_addr
);
4044 src
= change_address (src
, VOIDmode
, src_addr
);
4046 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4049 emit_move_insn (count
, temp
);
4051 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4054 emit_move_insn (blocks
, temp
);
4056 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4057 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4059 emit_label (loop_start_label
);
4062 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 768))
4066 /* Issue a read prefetch for the +3 cache line. */
4067 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, src_addr
, GEN_INT (768)),
4068 const0_rtx
, const0_rtx
);
4069 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4070 emit_insn (prefetch
);
4072 /* Issue a write prefetch for the +3 cache line. */
4073 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (768)),
4074 const1_rtx
, const0_rtx
);
4075 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4076 emit_insn (prefetch
);
4079 emit_insn (gen_movmem_short (dst
, src
, GEN_INT (255)));
4080 s390_load_address (dst_addr
,
4081 gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (256)));
4082 s390_load_address (src_addr
,
4083 gen_rtx_PLUS (Pmode
, src_addr
, GEN_INT (256)));
4085 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4088 emit_move_insn (blocks
, temp
);
4090 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4091 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4093 emit_jump (loop_start_label
);
4094 emit_label (loop_end_label
);
4096 emit_insn (gen_movmem_short (dst
, src
,
4097 convert_to_mode (Pmode
, count
, 1)));
4098 emit_label (end_label
);
4103 /* Emit code to set LEN bytes at DST to VAL.
4104 Make use of clrmem if VAL is zero. */
4107 s390_expand_setmem (rtx dst
, rtx len
, rtx val
)
4109 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) == 0)
4112 gcc_assert (GET_CODE (val
) == CONST_INT
|| GET_MODE (val
) == QImode
);
4114 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) > 0 && INTVAL (len
) <= 257)
4116 if (val
== const0_rtx
&& INTVAL (len
) <= 256)
4117 emit_insn (gen_clrmem_short (dst
, GEN_INT (INTVAL (len
) - 1)));
4120 /* Initialize memory by storing the first byte. */
4121 emit_move_insn (adjust_address (dst
, QImode
, 0), val
);
4123 if (INTVAL (len
) > 1)
4125 /* Initiate 1 byte overlap move.
4126 The first byte of DST is propagated through DSTP1.
4127 Prepare a movmem for: DST+1 = DST (length = LEN - 1).
4128 DST is set to size 1 so the rest of the memory location
4129 does not count as source operand. */
4130 rtx dstp1
= adjust_address (dst
, VOIDmode
, 1);
4131 set_mem_size (dst
, 1);
4133 emit_insn (gen_movmem_short (dstp1
, dst
,
4134 GEN_INT (INTVAL (len
) - 2)));
4139 else if (TARGET_MVCLE
)
4141 val
= force_not_mem (convert_modes (Pmode
, QImode
, val
, 1));
4142 emit_insn (gen_setmem_long (dst
, convert_to_mode (Pmode
, len
, 1), val
));
4147 rtx dst_addr
, count
, blocks
, temp
, dstp1
= NULL_RTX
;
4148 rtx loop_start_label
= gen_label_rtx ();
4149 rtx loop_end_label
= gen_label_rtx ();
4150 rtx end_label
= gen_label_rtx ();
4151 enum machine_mode mode
;
4153 mode
= GET_MODE (len
);
4154 if (mode
== VOIDmode
)
4157 dst_addr
= gen_reg_rtx (Pmode
);
4158 count
= gen_reg_rtx (mode
);
4159 blocks
= gen_reg_rtx (mode
);
4161 convert_move (count
, len
, 1);
4162 emit_cmp_and_jump_insns (count
, const0_rtx
,
4163 EQ
, NULL_RTX
, mode
, 1, end_label
);
4165 emit_move_insn (dst_addr
, force_operand (XEXP (dst
, 0), NULL_RTX
));
4166 dst
= change_address (dst
, VOIDmode
, dst_addr
);
4168 if (val
== const0_rtx
)
4169 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4173 dstp1
= adjust_address (dst
, VOIDmode
, 1);
4174 set_mem_size (dst
, 1);
4176 /* Initialize memory by storing the first byte. */
4177 emit_move_insn (adjust_address (dst
, QImode
, 0), val
);
4179 /* If count is 1 we are done. */
4180 emit_cmp_and_jump_insns (count
, const1_rtx
,
4181 EQ
, NULL_RTX
, mode
, 1, end_label
);
4183 temp
= expand_binop (mode
, add_optab
, count
, GEN_INT (-2), count
, 1,
4187 emit_move_insn (count
, temp
);
4189 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4192 emit_move_insn (blocks
, temp
);
4194 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4195 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4197 emit_label (loop_start_label
);
4200 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 1024))
4202 /* Issue a write prefetch for the +4 cache line. */
4203 rtx prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, dst_addr
,
4205 const1_rtx
, const0_rtx
);
4206 emit_insn (prefetch
);
4207 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4210 if (val
== const0_rtx
)
4211 emit_insn (gen_clrmem_short (dst
, GEN_INT (255)));
4213 emit_insn (gen_movmem_short (dstp1
, dst
, GEN_INT (255)));
4214 s390_load_address (dst_addr
,
4215 gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (256)));
4217 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4220 emit_move_insn (blocks
, temp
);
4222 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4223 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4225 emit_jump (loop_start_label
);
4226 emit_label (loop_end_label
);
4228 if (val
== const0_rtx
)
4229 emit_insn (gen_clrmem_short (dst
, convert_to_mode (Pmode
, count
, 1)));
4231 emit_insn (gen_movmem_short (dstp1
, dst
, convert_to_mode (Pmode
, count
, 1)));
4232 emit_label (end_label
);
4236 /* Emit code to compare LEN bytes at OP0 with those at OP1,
4237 and return the result in TARGET. */
4240 s390_expand_cmpmem (rtx target
, rtx op0
, rtx op1
, rtx len
)
4242 rtx ccreg
= gen_rtx_REG (CCUmode
, CC_REGNUM
);
4245 /* When tuning for z10 or higher we rely on the Glibc functions to
4246 do the right thing. Only for constant lengths below 64k we will
4247 generate inline code. */
4248 if (s390_tune
>= PROCESSOR_2097_Z10
4249 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > (1<<16)))
4252 /* As the result of CMPINT is inverted compared to what we need,
4253 we have to swap the operands. */
4254 tmp
= op0
; op0
= op1
; op1
= tmp
;
4256 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) >= 0 && INTVAL (len
) <= 256)
4258 if (INTVAL (len
) > 0)
4260 emit_insn (gen_cmpmem_short (op0
, op1
, GEN_INT (INTVAL (len
) - 1)));
4261 emit_insn (gen_cmpint (target
, ccreg
));
4264 emit_move_insn (target
, const0_rtx
);
4266 else if (TARGET_MVCLE
)
4268 emit_insn (gen_cmpmem_long (op0
, op1
, convert_to_mode (Pmode
, len
, 1)));
4269 emit_insn (gen_cmpint (target
, ccreg
));
4273 rtx addr0
, addr1
, count
, blocks
, temp
;
4274 rtx loop_start_label
= gen_label_rtx ();
4275 rtx loop_end_label
= gen_label_rtx ();
4276 rtx end_label
= gen_label_rtx ();
4277 enum machine_mode mode
;
4279 mode
= GET_MODE (len
);
4280 if (mode
== VOIDmode
)
4283 addr0
= gen_reg_rtx (Pmode
);
4284 addr1
= gen_reg_rtx (Pmode
);
4285 count
= gen_reg_rtx (mode
);
4286 blocks
= gen_reg_rtx (mode
);
4288 convert_move (count
, len
, 1);
4289 emit_cmp_and_jump_insns (count
, const0_rtx
,
4290 EQ
, NULL_RTX
, mode
, 1, end_label
);
4292 emit_move_insn (addr0
, force_operand (XEXP (op0
, 0), NULL_RTX
));
4293 emit_move_insn (addr1
, force_operand (XEXP (op1
, 0), NULL_RTX
));
4294 op0
= change_address (op0
, VOIDmode
, addr0
);
4295 op1
= change_address (op1
, VOIDmode
, addr1
);
4297 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4300 emit_move_insn (count
, temp
);
4302 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4305 emit_move_insn (blocks
, temp
);
4307 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4308 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4310 emit_label (loop_start_label
);
4313 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 512))
4317 /* Issue a read prefetch for the +2 cache line of operand 1. */
4318 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, addr0
, GEN_INT (512)),
4319 const0_rtx
, const0_rtx
);
4320 emit_insn (prefetch
);
4321 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4323 /* Issue a read prefetch for the +2 cache line of operand 2. */
4324 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, addr1
, GEN_INT (512)),
4325 const0_rtx
, const0_rtx
);
4326 emit_insn (prefetch
);
4327 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4330 emit_insn (gen_cmpmem_short (op0
, op1
, GEN_INT (255)));
4331 temp
= gen_rtx_NE (VOIDmode
, ccreg
, const0_rtx
);
4332 temp
= gen_rtx_IF_THEN_ELSE (VOIDmode
, temp
,
4333 gen_rtx_LABEL_REF (VOIDmode
, end_label
), pc_rtx
);
4334 temp
= gen_rtx_SET (VOIDmode
, pc_rtx
, temp
);
4335 emit_jump_insn (temp
);
4337 s390_load_address (addr0
,
4338 gen_rtx_PLUS (Pmode
, addr0
, GEN_INT (256)));
4339 s390_load_address (addr1
,
4340 gen_rtx_PLUS (Pmode
, addr1
, GEN_INT (256)));
4342 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4345 emit_move_insn (blocks
, temp
);
4347 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4348 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4350 emit_jump (loop_start_label
);
4351 emit_label (loop_end_label
);
4353 emit_insn (gen_cmpmem_short (op0
, op1
,
4354 convert_to_mode (Pmode
, count
, 1)));
4355 emit_label (end_label
);
4357 emit_insn (gen_cmpint (target
, ccreg
));
4363 /* Expand conditional increment or decrement using alc/slb instructions.
4364 Should generate code setting DST to either SRC or SRC + INCREMENT,
4365 depending on the result of the comparison CMP_OP0 CMP_CODE CMP_OP1.
4366 Returns true if successful, false otherwise.
4368 That makes it possible to implement some if-constructs without jumps e.g.:
4369 (borrow = CC0 | CC1 and carry = CC2 | CC3)
4370 unsigned int a, b, c;
4371 if (a < b) c++; -> CCU b > a -> CC2; c += carry;
4372 if (a < b) c--; -> CCL3 a - b -> borrow; c -= borrow;
4373 if (a <= b) c++; -> CCL3 b - a -> borrow; c += carry;
4374 if (a <= b) c--; -> CCU a <= b -> borrow; c -= borrow;
4376 Checks for EQ and NE with a nonzero value need an additional xor e.g.:
4377 if (a == b) c++; -> CCL3 a ^= b; 0 - a -> borrow; c += carry;
4378 if (a == b) c--; -> CCU a ^= b; a <= 0 -> CC0 | CC1; c -= borrow;
4379 if (a != b) c++; -> CCU a ^= b; a > 0 -> CC2; c += carry;
4380 if (a != b) c--; -> CCL3 a ^= b; 0 - a -> borrow; c -= borrow; */
4383 s390_expand_addcc (enum rtx_code cmp_code
, rtx cmp_op0
, rtx cmp_op1
,
4384 rtx dst
, rtx src
, rtx increment
)
4386 enum machine_mode cmp_mode
;
4387 enum machine_mode cc_mode
;
4393 if ((GET_MODE (cmp_op0
) == SImode
|| GET_MODE (cmp_op0
) == VOIDmode
)
4394 && (GET_MODE (cmp_op1
) == SImode
|| GET_MODE (cmp_op1
) == VOIDmode
))
4396 else if ((GET_MODE (cmp_op0
) == DImode
|| GET_MODE (cmp_op0
) == VOIDmode
)
4397 && (GET_MODE (cmp_op1
) == DImode
|| GET_MODE (cmp_op1
) == VOIDmode
))
4402 /* Try ADD LOGICAL WITH CARRY. */
4403 if (increment
== const1_rtx
)
4405 /* Determine CC mode to use. */
4406 if (cmp_code
== EQ
|| cmp_code
== NE
)
4408 if (cmp_op1
!= const0_rtx
)
4410 cmp_op0
= expand_simple_binop (cmp_mode
, XOR
, cmp_op0
, cmp_op1
,
4411 NULL_RTX
, 0, OPTAB_WIDEN
);
4412 cmp_op1
= const0_rtx
;
4415 cmp_code
= cmp_code
== EQ
? LEU
: GTU
;
4418 if (cmp_code
== LTU
|| cmp_code
== LEU
)
4423 cmp_code
= swap_condition (cmp_code
);
4440 /* Emit comparison instruction pattern. */
4441 if (!register_operand (cmp_op0
, cmp_mode
))
4442 cmp_op0
= force_reg (cmp_mode
, cmp_op0
);
4444 insn
= gen_rtx_SET (VOIDmode
, gen_rtx_REG (cc_mode
, CC_REGNUM
),
4445 gen_rtx_COMPARE (cc_mode
, cmp_op0
, cmp_op1
));
4446 /* We use insn_invalid_p here to add clobbers if required. */
4447 ret
= insn_invalid_p (emit_insn (insn
), false);
4450 /* Emit ALC instruction pattern. */
4451 op_res
= gen_rtx_fmt_ee (cmp_code
, GET_MODE (dst
),
4452 gen_rtx_REG (cc_mode
, CC_REGNUM
),
4455 if (src
!= const0_rtx
)
4457 if (!register_operand (src
, GET_MODE (dst
)))
4458 src
= force_reg (GET_MODE (dst
), src
);
4460 op_res
= gen_rtx_PLUS (GET_MODE (dst
), op_res
, src
);
4461 op_res
= gen_rtx_PLUS (GET_MODE (dst
), op_res
, const0_rtx
);
4464 p
= rtvec_alloc (2);
4466 gen_rtx_SET (VOIDmode
, dst
, op_res
);
4468 gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4469 emit_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
4474 /* Try SUBTRACT LOGICAL WITH BORROW. */
4475 if (increment
== constm1_rtx
)
4477 /* Determine CC mode to use. */
4478 if (cmp_code
== EQ
|| cmp_code
== NE
)
4480 if (cmp_op1
!= const0_rtx
)
4482 cmp_op0
= expand_simple_binop (cmp_mode
, XOR
, cmp_op0
, cmp_op1
,
4483 NULL_RTX
, 0, OPTAB_WIDEN
);
4484 cmp_op1
= const0_rtx
;
4487 cmp_code
= cmp_code
== EQ
? LEU
: GTU
;
4490 if (cmp_code
== GTU
|| cmp_code
== GEU
)
4495 cmp_code
= swap_condition (cmp_code
);
4512 /* Emit comparison instruction pattern. */
4513 if (!register_operand (cmp_op0
, cmp_mode
))
4514 cmp_op0
= force_reg (cmp_mode
, cmp_op0
);
4516 insn
= gen_rtx_SET (VOIDmode
, gen_rtx_REG (cc_mode
, CC_REGNUM
),
4517 gen_rtx_COMPARE (cc_mode
, cmp_op0
, cmp_op1
));
4518 /* We use insn_invalid_p here to add clobbers if required. */
4519 ret
= insn_invalid_p (emit_insn (insn
), false);
4522 /* Emit SLB instruction pattern. */
4523 if (!register_operand (src
, GET_MODE (dst
)))
4524 src
= force_reg (GET_MODE (dst
), src
);
4526 op_res
= gen_rtx_MINUS (GET_MODE (dst
),
4527 gen_rtx_MINUS (GET_MODE (dst
), src
, const0_rtx
),
4528 gen_rtx_fmt_ee (cmp_code
, GET_MODE (dst
),
4529 gen_rtx_REG (cc_mode
, CC_REGNUM
),
4531 p
= rtvec_alloc (2);
4533 gen_rtx_SET (VOIDmode
, dst
, op_res
);
4535 gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4536 emit_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
4544 /* Expand code for the insv template. Return true if successful. */
4547 s390_expand_insv (rtx dest
, rtx op1
, rtx op2
, rtx src
)
4549 int bitsize
= INTVAL (op1
);
4550 int bitpos
= INTVAL (op2
);
4552 /* On z10 we can use the risbg instruction to implement insv. */
4554 && ((GET_MODE (dest
) == DImode
&& GET_MODE (src
) == DImode
)
4555 || (GET_MODE (dest
) == SImode
&& GET_MODE (src
) == SImode
)))
4560 op
= gen_rtx_SET (GET_MODE(src
),
4561 gen_rtx_ZERO_EXTRACT (GET_MODE (dest
), dest
, op1
, op2
),
4563 clobber
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4564 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, op
, clobber
)));
4569 /* We need byte alignment. */
4570 if (bitsize
% BITS_PER_UNIT
)
4574 && memory_operand (dest
, VOIDmode
)
4575 && (register_operand (src
, word_mode
)
4576 || const_int_operand (src
, VOIDmode
)))
4578 /* Emit standard pattern if possible. */
4579 enum machine_mode mode
= smallest_mode_for_size (bitsize
, MODE_INT
);
4580 if (GET_MODE_BITSIZE (mode
) == bitsize
)
4581 emit_move_insn (adjust_address (dest
, mode
, 0), gen_lowpart (mode
, src
));
4583 /* (set (ze (mem)) (const_int)). */
4584 else if (const_int_operand (src
, VOIDmode
))
4586 int size
= bitsize
/ BITS_PER_UNIT
;
4587 rtx src_mem
= adjust_address (force_const_mem (word_mode
, src
), BLKmode
,
4588 GET_MODE_SIZE (word_mode
) - size
);
4590 dest
= adjust_address (dest
, BLKmode
, 0);
4591 set_mem_size (dest
, size
);
4592 s390_expand_movmem (dest
, src_mem
, GEN_INT (size
));
4595 /* (set (ze (mem)) (reg)). */
4596 else if (register_operand (src
, word_mode
))
4598 if (bitsize
<= GET_MODE_BITSIZE (SImode
))
4599 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
, op1
,
4603 /* Emit st,stcmh sequence. */
4604 int stcmh_width
= bitsize
- GET_MODE_BITSIZE (SImode
);
4605 int size
= stcmh_width
/ BITS_PER_UNIT
;
4607 emit_move_insn (adjust_address (dest
, SImode
, size
),
4608 gen_lowpart (SImode
, src
));
4609 set_mem_size (dest
, size
);
4610 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
, GEN_INT
4611 (stcmh_width
), const0_rtx
),
4612 gen_rtx_LSHIFTRT (word_mode
, src
, GEN_INT
4613 (GET_MODE_BITSIZE (SImode
))));
4622 /* (set (ze (reg)) (const_int)). */
4624 && register_operand (dest
, word_mode
)
4625 && (bitpos
% 16) == 0
4626 && (bitsize
% 16) == 0
4627 && const_int_operand (src
, VOIDmode
))
4629 HOST_WIDE_INT val
= INTVAL (src
);
4630 int regpos
= bitpos
+ bitsize
;
4632 while (regpos
> bitpos
)
4634 enum machine_mode putmode
;
4637 if (TARGET_EXTIMM
&& (regpos
% 32 == 0) && (regpos
>= bitpos
+ 32))
4642 putsize
= GET_MODE_BITSIZE (putmode
);
4644 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
,
4647 gen_int_mode (val
, putmode
));
4650 gcc_assert (regpos
== bitpos
);
4657 /* A subroutine of s390_expand_cs_hqi and s390_expand_atomic which returns a
4658 register that holds VAL of mode MODE shifted by COUNT bits. */
4661 s390_expand_mask_and_shift (rtx val
, enum machine_mode mode
, rtx count
)
4663 val
= expand_simple_binop (SImode
, AND
, val
, GEN_INT (GET_MODE_MASK (mode
)),
4664 NULL_RTX
, 1, OPTAB_DIRECT
);
4665 return expand_simple_binop (SImode
, ASHIFT
, val
, count
,
4666 NULL_RTX
, 1, OPTAB_DIRECT
);
4669 /* Structure to hold the initial parameters for a compare_and_swap operation
4670 in HImode and QImode. */
4672 struct alignment_context
4674 rtx memsi
; /* SI aligned memory location. */
4675 rtx shift
; /* Bit offset with regard to lsb. */
4676 rtx modemask
; /* Mask of the HQImode shifted by SHIFT bits. */
4677 rtx modemaski
; /* ~modemask */
4678 bool aligned
; /* True if memory is aligned, false else. */
4681 /* A subroutine of s390_expand_cs_hqi and s390_expand_atomic to initialize
4682 structure AC for transparent simplifying, if the memory alignment is known
4683 to be at least 32bit. MEM is the memory location for the actual operation
4684 and MODE its mode. */
4687 init_alignment_context (struct alignment_context
*ac
, rtx mem
,
4688 enum machine_mode mode
)
4690 ac
->shift
= GEN_INT (GET_MODE_SIZE (SImode
) - GET_MODE_SIZE (mode
));
4691 ac
->aligned
= (MEM_ALIGN (mem
) >= GET_MODE_BITSIZE (SImode
));
4694 ac
->memsi
= adjust_address (mem
, SImode
, 0); /* Memory is aligned. */
4697 /* Alignment is unknown. */
4698 rtx byteoffset
, addr
, align
;
4700 /* Force the address into a register. */
4701 addr
= force_reg (Pmode
, XEXP (mem
, 0));
4703 /* Align it to SImode. */
4704 align
= expand_simple_binop (Pmode
, AND
, addr
,
4705 GEN_INT (-GET_MODE_SIZE (SImode
)),
4706 NULL_RTX
, 1, OPTAB_DIRECT
);
4708 ac
->memsi
= gen_rtx_MEM (SImode
, align
);
4709 MEM_VOLATILE_P (ac
->memsi
) = MEM_VOLATILE_P (mem
);
4710 set_mem_alias_set (ac
->memsi
, ALIAS_SET_MEMORY_BARRIER
);
4711 set_mem_align (ac
->memsi
, GET_MODE_BITSIZE (SImode
));
4713 /* Calculate shiftcount. */
4714 byteoffset
= expand_simple_binop (Pmode
, AND
, addr
,
4715 GEN_INT (GET_MODE_SIZE (SImode
) - 1),
4716 NULL_RTX
, 1, OPTAB_DIRECT
);
4717 /* As we already have some offset, evaluate the remaining distance. */
4718 ac
->shift
= expand_simple_binop (SImode
, MINUS
, ac
->shift
, byteoffset
,
4719 NULL_RTX
, 1, OPTAB_DIRECT
);
4722 /* Shift is the byte count, but we need the bitcount. */
4723 ac
->shift
= expand_simple_binop (SImode
, MULT
, ac
->shift
, GEN_INT (BITS_PER_UNIT
),
4724 NULL_RTX
, 1, OPTAB_DIRECT
);
4725 /* Calculate masks. */
4726 ac
->modemask
= expand_simple_binop (SImode
, ASHIFT
,
4727 GEN_INT (GET_MODE_MASK (mode
)), ac
->shift
,
4728 NULL_RTX
, 1, OPTAB_DIRECT
);
4729 ac
->modemaski
= expand_simple_unop (SImode
, NOT
, ac
->modemask
, NULL_RTX
, 1);
4732 /* Expand an atomic compare and swap operation for HImode and QImode. MEM is
4733 the memory location, CMP the old value to compare MEM with and NEW_RTX the value
4734 to set if CMP == MEM.
4735 CMP is never in memory for compare_and_swap_cc because
4736 expand_bool_compare_and_swap puts it into a register for later compare. */
4739 s390_expand_cs_hqi (enum machine_mode mode
, rtx target
, rtx mem
, rtx cmp
, rtx new_rtx
)
4741 struct alignment_context ac
;
4742 rtx cmpv
, newv
, val
, resv
, cc
;
4743 rtx res
= gen_reg_rtx (SImode
);
4744 rtx csloop
= gen_label_rtx ();
4745 rtx csend
= gen_label_rtx ();
4747 gcc_assert (register_operand (target
, VOIDmode
));
4748 gcc_assert (MEM_P (mem
));
4750 init_alignment_context (&ac
, mem
, mode
);
4752 /* Shift the values to the correct bit positions. */
4753 if (!(ac
.aligned
&& MEM_P (cmp
)))
4754 cmp
= s390_expand_mask_and_shift (cmp
, mode
, ac
.shift
);
4755 if (!(ac
.aligned
&& MEM_P (new_rtx
)))
4756 new_rtx
= s390_expand_mask_and_shift (new_rtx
, mode
, ac
.shift
);
4758 /* Load full word. Subsequent loads are performed by CS. */
4759 val
= expand_simple_binop (SImode
, AND
, ac
.memsi
, ac
.modemaski
,
4760 NULL_RTX
, 1, OPTAB_DIRECT
);
4762 /* Start CS loop. */
4763 emit_label (csloop
);
4764 /* val = "<mem>00..0<mem>"
4765 * cmp = "00..0<cmp>00..0"
4766 * new = "00..0<new>00..0"
4769 /* Patch cmp and new with val at correct position. */
4770 if (ac
.aligned
&& MEM_P (cmp
))
4772 cmpv
= force_reg (SImode
, val
);
4773 store_bit_field (cmpv
, GET_MODE_BITSIZE (mode
), 0,
4777 cmpv
= force_reg (SImode
, expand_simple_binop (SImode
, IOR
, cmp
, val
,
4778 NULL_RTX
, 1, OPTAB_DIRECT
));
4779 if (ac
.aligned
&& MEM_P (new_rtx
))
4781 newv
= force_reg (SImode
, val
);
4782 store_bit_field (newv
, GET_MODE_BITSIZE (mode
), 0,
4783 0, 0, SImode
, new_rtx
);
4786 newv
= force_reg (SImode
, expand_simple_binop (SImode
, IOR
, new_rtx
, val
,
4787 NULL_RTX
, 1, OPTAB_DIRECT
));
4789 /* Jump to end if we're done (likely?). */
4790 s390_emit_jump (csend
, s390_emit_compare_and_swap (EQ
, res
, ac
.memsi
,
4793 /* Check for changes outside mode. */
4794 resv
= expand_simple_binop (SImode
, AND
, res
, ac
.modemaski
,
4795 NULL_RTX
, 1, OPTAB_DIRECT
);
4796 cc
= s390_emit_compare (NE
, resv
, val
);
4797 emit_move_insn (val
, resv
);
4798 /* Loop internal if so. */
4799 s390_emit_jump (csloop
, cc
);
4803 /* Return the correct part of the bitfield. */
4804 convert_move (target
, expand_simple_binop (SImode
, LSHIFTRT
, res
, ac
.shift
,
4805 NULL_RTX
, 1, OPTAB_DIRECT
), 1);
4808 /* Expand an atomic operation CODE of mode MODE. MEM is the memory location
4809 and VAL the value to play with. If AFTER is true then store the value
4810 MEM holds after the operation, if AFTER is false then store the value MEM
4811 holds before the operation. If TARGET is zero then discard that value, else
4812 store it to TARGET. */
4815 s390_expand_atomic (enum machine_mode mode
, enum rtx_code code
,
4816 rtx target
, rtx mem
, rtx val
, bool after
)
4818 struct alignment_context ac
;
4820 rtx new_rtx
= gen_reg_rtx (SImode
);
4821 rtx orig
= gen_reg_rtx (SImode
);
4822 rtx csloop
= gen_label_rtx ();
4824 gcc_assert (!target
|| register_operand (target
, VOIDmode
));
4825 gcc_assert (MEM_P (mem
));
4827 init_alignment_context (&ac
, mem
, mode
);
4829 /* Shift val to the correct bit positions.
4830 Preserve "icm", but prevent "ex icm". */
4831 if (!(ac
.aligned
&& code
== SET
&& MEM_P (val
)))
4832 val
= s390_expand_mask_and_shift (val
, mode
, ac
.shift
);
4834 /* Further preparation insns. */
4835 if (code
== PLUS
|| code
== MINUS
)
4836 emit_move_insn (orig
, val
);
4837 else if (code
== MULT
|| code
== AND
) /* val = "11..1<val>11..1" */
4838 val
= expand_simple_binop (SImode
, XOR
, val
, ac
.modemaski
,
4839 NULL_RTX
, 1, OPTAB_DIRECT
);
4841 /* Load full word. Subsequent loads are performed by CS. */
4842 cmp
= force_reg (SImode
, ac
.memsi
);
4844 /* Start CS loop. */
4845 emit_label (csloop
);
4846 emit_move_insn (new_rtx
, cmp
);
4848 /* Patch new with val at correct position. */
4853 val
= expand_simple_binop (SImode
, code
, new_rtx
, orig
,
4854 NULL_RTX
, 1, OPTAB_DIRECT
);
4855 val
= expand_simple_binop (SImode
, AND
, val
, ac
.modemask
,
4856 NULL_RTX
, 1, OPTAB_DIRECT
);
4859 if (ac
.aligned
&& MEM_P (val
))
4860 store_bit_field (new_rtx
, GET_MODE_BITSIZE (mode
), 0,
4864 new_rtx
= expand_simple_binop (SImode
, AND
, new_rtx
, ac
.modemaski
,
4865 NULL_RTX
, 1, OPTAB_DIRECT
);
4866 new_rtx
= expand_simple_binop (SImode
, IOR
, new_rtx
, val
,
4867 NULL_RTX
, 1, OPTAB_DIRECT
);
4873 new_rtx
= expand_simple_binop (SImode
, code
, new_rtx
, val
,
4874 NULL_RTX
, 1, OPTAB_DIRECT
);
4876 case MULT
: /* NAND */
4877 new_rtx
= expand_simple_binop (SImode
, AND
, new_rtx
, val
,
4878 NULL_RTX
, 1, OPTAB_DIRECT
);
4879 new_rtx
= expand_simple_binop (SImode
, XOR
, new_rtx
, ac
.modemask
,
4880 NULL_RTX
, 1, OPTAB_DIRECT
);
4886 s390_emit_jump (csloop
, s390_emit_compare_and_swap (NE
, cmp
,
4887 ac
.memsi
, cmp
, new_rtx
));
4889 /* Return the correct part of the bitfield. */
4891 convert_move (target
, expand_simple_binop (SImode
, LSHIFTRT
,
4892 after
? new_rtx
: cmp
, ac
.shift
,
4893 NULL_RTX
, 1, OPTAB_DIRECT
), 1);
4896 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
4897 We need to emit DTP-relative relocations. */
4899 static void s390_output_dwarf_dtprel (FILE *, int, rtx
) ATTRIBUTE_UNUSED
;
4902 s390_output_dwarf_dtprel (FILE *file
, int size
, rtx x
)
4907 fputs ("\t.long\t", file
);
4910 fputs ("\t.quad\t", file
);
4915 output_addr_const (file
, x
);
4916 fputs ("@DTPOFF", file
);
4919 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
4920 /* Implement TARGET_MANGLE_TYPE. */
4923 s390_mangle_type (const_tree type
)
4925 if (TYPE_MAIN_VARIANT (type
) == long_double_type_node
4926 && TARGET_LONG_DOUBLE_128
)
4929 /* For all other types, use normal C++ mangling. */
4934 /* In the name of slightly smaller debug output, and to cater to
4935 general assembler lossage, recognize various UNSPEC sequences
4936 and turn them back into a direct symbol reference. */
4939 s390_delegitimize_address (rtx orig_x
)
4943 orig_x
= delegitimize_mem_from_attrs (orig_x
);
4946 /* Extract the symbol ref from:
4947 (plus:SI (reg:SI 12 %r12)
4948 (const:SI (unspec:SI [(symbol_ref/f:SI ("*.LC0"))]
4949 UNSPEC_GOTOFF/PLTOFF)))
4951 (plus:SI (reg:SI 12 %r12)
4952 (const:SI (plus:SI (unspec:SI [(symbol_ref:SI ("L"))]
4953 UNSPEC_GOTOFF/PLTOFF)
4954 (const_int 4 [0x4])))) */
4955 if (GET_CODE (x
) == PLUS
4956 && REG_P (XEXP (x
, 0))
4957 && REGNO (XEXP (x
, 0)) == PIC_OFFSET_TABLE_REGNUM
4958 && GET_CODE (XEXP (x
, 1)) == CONST
)
4960 HOST_WIDE_INT offset
= 0;
4962 /* The const operand. */
4963 y
= XEXP (XEXP (x
, 1), 0);
4965 if (GET_CODE (y
) == PLUS
4966 && GET_CODE (XEXP (y
, 1)) == CONST_INT
)
4968 offset
= INTVAL (XEXP (y
, 1));
4972 if (GET_CODE (y
) == UNSPEC
4973 && (XINT (y
, 1) == UNSPEC_GOTOFF
4974 || XINT (y
, 1) == UNSPEC_PLTOFF
))
4975 return plus_constant (Pmode
, XVECEXP (y
, 0, 0), offset
);
4978 if (GET_CODE (x
) != MEM
)
4982 if (GET_CODE (x
) == PLUS
4983 && GET_CODE (XEXP (x
, 1)) == CONST
4984 && GET_CODE (XEXP (x
, 0)) == REG
4985 && REGNO (XEXP (x
, 0)) == PIC_OFFSET_TABLE_REGNUM
)
4987 y
= XEXP (XEXP (x
, 1), 0);
4988 if (GET_CODE (y
) == UNSPEC
4989 && XINT (y
, 1) == UNSPEC_GOT
)
4990 y
= XVECEXP (y
, 0, 0);
4994 else if (GET_CODE (x
) == CONST
)
4996 /* Extract the symbol ref from:
4997 (mem:QI (const:DI (unspec:DI [(symbol_ref:DI ("foo"))]
4998 UNSPEC_PLT/GOTENT))) */
5001 if (GET_CODE (y
) == UNSPEC
5002 && (XINT (y
, 1) == UNSPEC_GOTENT
5003 || XINT (y
, 1) == UNSPEC_PLT
))
5004 y
= XVECEXP (y
, 0, 0);
5011 if (GET_MODE (orig_x
) != Pmode
)
5013 if (GET_MODE (orig_x
) == BLKmode
)
5015 y
= lowpart_subreg (GET_MODE (orig_x
), y
, Pmode
);
5022 /* Output operand OP to stdio stream FILE.
5023 OP is an address (register + offset) which is not used to address data;
5024 instead the rightmost bits are interpreted as the value. */
5027 print_shift_count_operand (FILE *file
, rtx op
)
5029 HOST_WIDE_INT offset
;
5032 /* Extract base register and offset. */
5033 if (!s390_decompose_shift_count (op
, &base
, &offset
))
5039 gcc_assert (GET_CODE (base
) == REG
);
5040 gcc_assert (REGNO (base
) < FIRST_PSEUDO_REGISTER
);
5041 gcc_assert (REGNO_REG_CLASS (REGNO (base
)) == ADDR_REGS
);
5044 /* Offsets are constricted to twelve bits. */
5045 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, offset
& ((1 << 12) - 1));
5047 fprintf (file
, "(%s)", reg_names
[REGNO (base
)]);
5050 /* See 'get_some_local_dynamic_name'. */
5053 get_some_local_dynamic_name_1 (rtx
*px
, void *data ATTRIBUTE_UNUSED
)
5057 if (GET_CODE (x
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (x
))
5059 x
= get_pool_constant (x
);
5060 return for_each_rtx (&x
, get_some_local_dynamic_name_1
, 0);
5063 if (GET_CODE (x
) == SYMBOL_REF
5064 && tls_symbolic_operand (x
) == TLS_MODEL_LOCAL_DYNAMIC
)
5066 cfun
->machine
->some_ld_name
= XSTR (x
, 0);
5073 /* Locate some local-dynamic symbol still in use by this function
5074 so that we can print its name in local-dynamic base patterns. */
5077 get_some_local_dynamic_name (void)
5081 if (cfun
->machine
->some_ld_name
)
5082 return cfun
->machine
->some_ld_name
;
5084 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
5086 && for_each_rtx (&PATTERN (insn
), get_some_local_dynamic_name_1
, 0))
5087 return cfun
->machine
->some_ld_name
;
5092 /* Output machine-dependent UNSPECs occurring in address constant X
5093 in assembler syntax to stdio stream FILE. Returns true if the
5094 constant X could be recognized, false otherwise. */
5097 s390_output_addr_const_extra (FILE *file
, rtx x
)
5099 if (GET_CODE (x
) == UNSPEC
&& XVECLEN (x
, 0) == 1)
5100 switch (XINT (x
, 1))
5103 output_addr_const (file
, XVECEXP (x
, 0, 0));
5104 fprintf (file
, "@GOTENT");
5107 output_addr_const (file
, XVECEXP (x
, 0, 0));
5108 fprintf (file
, "@GOT");
5111 output_addr_const (file
, XVECEXP (x
, 0, 0));
5112 fprintf (file
, "@GOTOFF");
5115 output_addr_const (file
, XVECEXP (x
, 0, 0));
5116 fprintf (file
, "@PLT");
5119 output_addr_const (file
, XVECEXP (x
, 0, 0));
5120 fprintf (file
, "@PLTOFF");
5123 output_addr_const (file
, XVECEXP (x
, 0, 0));
5124 fprintf (file
, "@TLSGD");
5127 assemble_name (file
, get_some_local_dynamic_name ());
5128 fprintf (file
, "@TLSLDM");
5131 output_addr_const (file
, XVECEXP (x
, 0, 0));
5132 fprintf (file
, "@DTPOFF");
5135 output_addr_const (file
, XVECEXP (x
, 0, 0));
5136 fprintf (file
, "@NTPOFF");
5138 case UNSPEC_GOTNTPOFF
:
5139 output_addr_const (file
, XVECEXP (x
, 0, 0));
5140 fprintf (file
, "@GOTNTPOFF");
5142 case UNSPEC_INDNTPOFF
:
5143 output_addr_const (file
, XVECEXP (x
, 0, 0));
5144 fprintf (file
, "@INDNTPOFF");
5148 if (GET_CODE (x
) == UNSPEC
&& XVECLEN (x
, 0) == 2)
5149 switch (XINT (x
, 1))
5151 case UNSPEC_POOL_OFFSET
:
5152 x
= gen_rtx_MINUS (GET_MODE (x
), XVECEXP (x
, 0, 0), XVECEXP (x
, 0, 1));
5153 output_addr_const (file
, x
);
5159 /* Output address operand ADDR in assembler syntax to
5160 stdio stream FILE. */
5163 print_operand_address (FILE *file
, rtx addr
)
5165 struct s390_address ad
;
5167 if (s390_symref_operand_p (addr
, NULL
, NULL
))
5171 output_operand_lossage ("symbolic memory references are "
5172 "only supported on z10 or later");
5175 output_addr_const (file
, addr
);
5179 if (!s390_decompose_address (addr
, &ad
)
5180 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5181 || (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
))))
5182 output_operand_lossage ("cannot decompose address");
5185 output_addr_const (file
, ad
.disp
);
5187 fprintf (file
, "0");
5189 if (ad
.base
&& ad
.indx
)
5190 fprintf (file
, "(%s,%s)", reg_names
[REGNO (ad
.indx
)],
5191 reg_names
[REGNO (ad
.base
)]);
5193 fprintf (file
, "(%s)", reg_names
[REGNO (ad
.base
)]);
5196 /* Output operand X in assembler syntax to stdio stream FILE.
5197 CODE specified the format flag. The following format flags
5200 'C': print opcode suffix for branch condition.
5201 'D': print opcode suffix for inverse branch condition.
5202 'E': print opcode suffix for branch on index instruction.
5203 'J': print tls_load/tls_gdcall/tls_ldcall suffix
5204 'G': print the size of the operand in bytes.
5205 'O': print only the displacement of a memory reference.
5206 'R': print only the base register of a memory reference.
5207 'S': print S-type memory reference (base+displacement).
5208 'N': print the second word of a DImode operand.
5209 'M': print the second word of a TImode operand.
5210 'Y': print shift count operand.
5212 'b': print integer X as if it's an unsigned byte.
5213 'c': print integer X as if it's an signed byte.
5214 'x': print integer X as if it's an unsigned halfword.
5215 'h': print integer X as if it's a signed halfword.
5216 'i': print the first nonzero HImode part of X.
5217 'j': print the first HImode part unequal to -1 of X.
5218 'k': print the first nonzero SImode part of X.
5219 'm': print the first SImode part unequal to -1 of X.
5220 'o': print integer X as if it's an unsigned 32bit word. */
5223 print_operand (FILE *file
, rtx x
, int code
)
5228 fprintf (file
, s390_branch_condition_mnemonic (x
, FALSE
));
5232 fprintf (file
, s390_branch_condition_mnemonic (x
, TRUE
));
5236 if (GET_CODE (x
) == LE
)
5237 fprintf (file
, "l");
5238 else if (GET_CODE (x
) == GT
)
5239 fprintf (file
, "h");
5241 output_operand_lossage ("invalid comparison operator "
5242 "for 'E' output modifier");
5246 if (GET_CODE (x
) == SYMBOL_REF
)
5248 fprintf (file
, "%s", ":tls_load:");
5249 output_addr_const (file
, x
);
5251 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLSGD
)
5253 fprintf (file
, "%s", ":tls_gdcall:");
5254 output_addr_const (file
, XVECEXP (x
, 0, 0));
5256 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLSLDM
)
5258 fprintf (file
, "%s", ":tls_ldcall:");
5259 assemble_name (file
, get_some_local_dynamic_name ());
5262 output_operand_lossage ("invalid reference for 'J' output modifier");
5266 fprintf (file
, "%u", GET_MODE_SIZE (GET_MODE (x
)));
5271 struct s390_address ad
;
5276 output_operand_lossage ("memory reference expected for "
5277 "'O' output modifier");
5281 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5284 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5287 output_operand_lossage ("invalid address for 'O' output modifier");
5292 output_addr_const (file
, ad
.disp
);
5294 fprintf (file
, "0");
5300 struct s390_address ad
;
5305 output_operand_lossage ("memory reference expected for "
5306 "'R' output modifier");
5310 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5313 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5316 output_operand_lossage ("invalid address for 'R' output modifier");
5321 fprintf (file
, "%s", reg_names
[REGNO (ad
.base
)]);
5323 fprintf (file
, "0");
5329 struct s390_address ad
;
5334 output_operand_lossage ("memory reference expected for "
5335 "'S' output modifier");
5338 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5341 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5344 output_operand_lossage ("invalid address for 'S' output modifier");
5349 output_addr_const (file
, ad
.disp
);
5351 fprintf (file
, "0");
5354 fprintf (file
, "(%s)", reg_names
[REGNO (ad
.base
)]);
5359 if (GET_CODE (x
) == REG
)
5360 x
= gen_rtx_REG (GET_MODE (x
), REGNO (x
) + 1);
5361 else if (GET_CODE (x
) == MEM
)
5362 x
= change_address (x
, VOIDmode
,
5363 plus_constant (Pmode
, XEXP (x
, 0), 4));
5365 output_operand_lossage ("register or memory expression expected "
5366 "for 'N' output modifier");
5370 if (GET_CODE (x
) == REG
)
5371 x
= gen_rtx_REG (GET_MODE (x
), REGNO (x
) + 1);
5372 else if (GET_CODE (x
) == MEM
)
5373 x
= change_address (x
, VOIDmode
,
5374 plus_constant (Pmode
, XEXP (x
, 0), 8));
5376 output_operand_lossage ("register or memory expression expected "
5377 "for 'M' output modifier");
5381 print_shift_count_operand (file
, x
);
5385 switch (GET_CODE (x
))
5388 fprintf (file
, "%s", reg_names
[REGNO (x
)]);
5392 output_address (XEXP (x
, 0));
5399 output_addr_const (file
, x
);
5404 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xff);
5405 else if (code
== 'c')
5406 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ((INTVAL (x
) & 0xff) ^ 0x80) - 0x80);
5407 else if (code
== 'x')
5408 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xffff);
5409 else if (code
== 'h')
5410 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ((INTVAL (x
) & 0xffff) ^ 0x8000) - 0x8000);
5411 else if (code
== 'i')
5412 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5413 s390_extract_part (x
, HImode
, 0));
5414 else if (code
== 'j')
5415 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5416 s390_extract_part (x
, HImode
, -1));
5417 else if (code
== 'k')
5418 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5419 s390_extract_part (x
, SImode
, 0));
5420 else if (code
== 'm')
5421 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5422 s390_extract_part (x
, SImode
, -1));
5423 else if (code
== 'o')
5424 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xffffffff);
5426 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
));
5430 gcc_assert (GET_MODE (x
) == VOIDmode
);
5432 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, CONST_DOUBLE_LOW (x
) & 0xff);
5433 else if (code
== 'x')
5434 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, CONST_DOUBLE_LOW (x
) & 0xffff);
5435 else if (code
== 'h')
5436 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5437 ((CONST_DOUBLE_LOW (x
) & 0xffff) ^ 0x8000) - 0x8000);
5441 output_operand_lossage ("invalid constant - try using "
5442 "an output modifier");
5444 output_operand_lossage ("invalid constant for output modifier '%c'",
5451 output_operand_lossage ("invalid expression - try using "
5452 "an output modifier");
5454 output_operand_lossage ("invalid expression for output "
5455 "modifier '%c'", code
);
5460 /* Target hook for assembling integer objects. We need to define it
5461 here to work a round a bug in some versions of GAS, which couldn't
5462 handle values smaller than INT_MIN when printed in decimal. */
5465 s390_assemble_integer (rtx x
, unsigned int size
, int aligned_p
)
5467 if (size
== 8 && aligned_p
5468 && GET_CODE (x
) == CONST_INT
&& INTVAL (x
) < INT_MIN
)
5470 fprintf (asm_out_file
, "\t.quad\t" HOST_WIDE_INT_PRINT_HEX
"\n",
5474 return default_assemble_integer (x
, size
, aligned_p
);
5477 /* Returns true if register REGNO is used for forming
5478 a memory address in expression X. */
5481 reg_used_in_mem_p (int regno
, rtx x
)
5483 enum rtx_code code
= GET_CODE (x
);
5489 if (refers_to_regno_p (regno
, regno
+1,
5493 else if (code
== SET
5494 && GET_CODE (SET_DEST (x
)) == PC
)
5496 if (refers_to_regno_p (regno
, regno
+1,
5501 fmt
= GET_RTX_FORMAT (code
);
5502 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
5505 && reg_used_in_mem_p (regno
, XEXP (x
, i
)))
5508 else if (fmt
[i
] == 'E')
5509 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5510 if (reg_used_in_mem_p (regno
, XVECEXP (x
, i
, j
)))
5516 /* Returns true if expression DEP_RTX sets an address register
5517 used by instruction INSN to address memory. */
5520 addr_generation_dependency_p (rtx dep_rtx
, rtx insn
)
5524 if (GET_CODE (dep_rtx
) == INSN
)
5525 dep_rtx
= PATTERN (dep_rtx
);
5527 if (GET_CODE (dep_rtx
) == SET
)
5529 target
= SET_DEST (dep_rtx
);
5530 if (GET_CODE (target
) == STRICT_LOW_PART
)
5531 target
= XEXP (target
, 0);
5532 while (GET_CODE (target
) == SUBREG
)
5533 target
= SUBREG_REG (target
);
5535 if (GET_CODE (target
) == REG
)
5537 int regno
= REGNO (target
);
5539 if (s390_safe_attr_type (insn
) == TYPE_LA
)
5541 pat
= PATTERN (insn
);
5542 if (GET_CODE (pat
) == PARALLEL
)
5544 gcc_assert (XVECLEN (pat
, 0) == 2);
5545 pat
= XVECEXP (pat
, 0, 0);
5547 gcc_assert (GET_CODE (pat
) == SET
);
5548 return refers_to_regno_p (regno
, regno
+1, SET_SRC (pat
), 0);
5550 else if (get_attr_atype (insn
) == ATYPE_AGEN
)
5551 return reg_used_in_mem_p (regno
, PATTERN (insn
));
5557 /* Return 1, if dep_insn sets register used in insn in the agen unit. */
5560 s390_agen_dep_p (rtx dep_insn
, rtx insn
)
5562 rtx dep_rtx
= PATTERN (dep_insn
);
5565 if (GET_CODE (dep_rtx
) == SET
5566 && addr_generation_dependency_p (dep_rtx
, insn
))
5568 else if (GET_CODE (dep_rtx
) == PARALLEL
)
5570 for (i
= 0; i
< XVECLEN (dep_rtx
, 0); i
++)
5572 if (addr_generation_dependency_p (XVECEXP (dep_rtx
, 0, i
), insn
))
5580 /* A C statement (sans semicolon) to update the integer scheduling priority
5581 INSN_PRIORITY (INSN). Increase the priority to execute the INSN earlier,
5582 reduce the priority to execute INSN later. Do not define this macro if
5583 you do not need to adjust the scheduling priorities of insns.
5585 A STD instruction should be scheduled earlier,
5586 in order to use the bypass. */
5588 s390_adjust_priority (rtx insn ATTRIBUTE_UNUSED
, int priority
)
5590 if (! INSN_P (insn
))
5593 if (s390_tune
!= PROCESSOR_2084_Z990
5594 && s390_tune
!= PROCESSOR_2094_Z9_109
5595 && s390_tune
!= PROCESSOR_2097_Z10
5596 && s390_tune
!= PROCESSOR_2817_Z196
)
5599 switch (s390_safe_attr_type (insn
))
5603 priority
= priority
<< 3;
5607 priority
= priority
<< 1;
5616 /* The number of instructions that can be issued per cycle. */
5619 s390_issue_rate (void)
5623 case PROCESSOR_2084_Z990
:
5624 case PROCESSOR_2094_Z9_109
:
5625 case PROCESSOR_2817_Z196
:
5627 case PROCESSOR_2097_Z10
:
5635 s390_first_cycle_multipass_dfa_lookahead (void)
5640 /* Annotate every literal pool reference in X by an UNSPEC_LTREF expression.
5641 Fix up MEMs as required. */
5644 annotate_constant_pool_refs (rtx
*x
)
5649 gcc_assert (GET_CODE (*x
) != SYMBOL_REF
5650 || !CONSTANT_POOL_ADDRESS_P (*x
));
5652 /* Literal pool references can only occur inside a MEM ... */
5653 if (GET_CODE (*x
) == MEM
)
5655 rtx memref
= XEXP (*x
, 0);
5657 if (GET_CODE (memref
) == SYMBOL_REF
5658 && CONSTANT_POOL_ADDRESS_P (memref
))
5660 rtx base
= cfun
->machine
->base_reg
;
5661 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, memref
, base
),
5664 *x
= replace_equiv_address (*x
, addr
);
5668 if (GET_CODE (memref
) == CONST
5669 && GET_CODE (XEXP (memref
, 0)) == PLUS
5670 && GET_CODE (XEXP (XEXP (memref
, 0), 1)) == CONST_INT
5671 && GET_CODE (XEXP (XEXP (memref
, 0), 0)) == SYMBOL_REF
5672 && CONSTANT_POOL_ADDRESS_P (XEXP (XEXP (memref
, 0), 0)))
5674 HOST_WIDE_INT off
= INTVAL (XEXP (XEXP (memref
, 0), 1));
5675 rtx sym
= XEXP (XEXP (memref
, 0), 0);
5676 rtx base
= cfun
->machine
->base_reg
;
5677 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, sym
, base
),
5680 *x
= replace_equiv_address (*x
, plus_constant (Pmode
, addr
, off
));
5685 /* ... or a load-address type pattern. */
5686 if (GET_CODE (*x
) == SET
)
5688 rtx addrref
= SET_SRC (*x
);
5690 if (GET_CODE (addrref
) == SYMBOL_REF
5691 && CONSTANT_POOL_ADDRESS_P (addrref
))
5693 rtx base
= cfun
->machine
->base_reg
;
5694 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, addrref
, base
),
5697 SET_SRC (*x
) = addr
;
5701 if (GET_CODE (addrref
) == CONST
5702 && GET_CODE (XEXP (addrref
, 0)) == PLUS
5703 && GET_CODE (XEXP (XEXP (addrref
, 0), 1)) == CONST_INT
5704 && GET_CODE (XEXP (XEXP (addrref
, 0), 0)) == SYMBOL_REF
5705 && CONSTANT_POOL_ADDRESS_P (XEXP (XEXP (addrref
, 0), 0)))
5707 HOST_WIDE_INT off
= INTVAL (XEXP (XEXP (addrref
, 0), 1));
5708 rtx sym
= XEXP (XEXP (addrref
, 0), 0);
5709 rtx base
= cfun
->machine
->base_reg
;
5710 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, sym
, base
),
5713 SET_SRC (*x
) = plus_constant (Pmode
, addr
, off
);
5718 /* Annotate LTREL_BASE as well. */
5719 if (GET_CODE (*x
) == UNSPEC
5720 && XINT (*x
, 1) == UNSPEC_LTREL_BASE
)
5722 rtx base
= cfun
->machine
->base_reg
;
5723 *x
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, XVECEXP (*x
, 0, 0), base
),
5728 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5729 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5733 annotate_constant_pool_refs (&XEXP (*x
, i
));
5735 else if (fmt
[i
] == 'E')
5737 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5738 annotate_constant_pool_refs (&XVECEXP (*x
, i
, j
));
5743 /* Split all branches that exceed the maximum distance.
5744 Returns true if this created a new literal pool entry. */
5747 s390_split_branches (void)
5749 rtx temp_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
5750 int new_literal
= 0, ret
;
5751 rtx insn
, pat
, tmp
, target
;
5754 /* We need correct insn addresses. */
5756 shorten_branches (get_insns ());
5758 /* Find all branches that exceed 64KB, and split them. */
5760 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
5762 if (GET_CODE (insn
) != JUMP_INSN
)
5765 pat
= PATTERN (insn
);
5766 if (GET_CODE (pat
) == PARALLEL
&& XVECLEN (pat
, 0) > 2)
5767 pat
= XVECEXP (pat
, 0, 0);
5768 if (GET_CODE (pat
) != SET
|| SET_DEST (pat
) != pc_rtx
)
5771 if (GET_CODE (SET_SRC (pat
)) == LABEL_REF
)
5773 label
= &SET_SRC (pat
);
5775 else if (GET_CODE (SET_SRC (pat
)) == IF_THEN_ELSE
)
5777 if (GET_CODE (XEXP (SET_SRC (pat
), 1)) == LABEL_REF
)
5778 label
= &XEXP (SET_SRC (pat
), 1);
5779 else if (GET_CODE (XEXP (SET_SRC (pat
), 2)) == LABEL_REF
)
5780 label
= &XEXP (SET_SRC (pat
), 2);
5787 if (get_attr_length (insn
) <= 4)
5790 /* We are going to use the return register as scratch register,
5791 make sure it will be saved/restored by the prologue/epilogue. */
5792 cfun_frame_layout
.save_return_addr_p
= 1;
5797 tmp
= force_const_mem (Pmode
, *label
);
5798 tmp
= emit_insn_before (gen_rtx_SET (Pmode
, temp_reg
, tmp
), insn
);
5799 INSN_ADDRESSES_NEW (tmp
, -1);
5800 annotate_constant_pool_refs (&PATTERN (tmp
));
5807 target
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, *label
),
5808 UNSPEC_LTREL_OFFSET
);
5809 target
= gen_rtx_CONST (Pmode
, target
);
5810 target
= force_const_mem (Pmode
, target
);
5811 tmp
= emit_insn_before (gen_rtx_SET (Pmode
, temp_reg
, target
), insn
);
5812 INSN_ADDRESSES_NEW (tmp
, -1);
5813 annotate_constant_pool_refs (&PATTERN (tmp
));
5815 target
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, XEXP (target
, 0),
5816 cfun
->machine
->base_reg
),
5818 target
= gen_rtx_PLUS (Pmode
, temp_reg
, target
);
5821 ret
= validate_change (insn
, label
, target
, 0);
5829 /* Find an annotated literal pool symbol referenced in RTX X,
5830 and store it at REF. Will abort if X contains references to
5831 more than one such pool symbol; multiple references to the same
5832 symbol are allowed, however.
5834 The rtx pointed to by REF must be initialized to NULL_RTX
5835 by the caller before calling this routine. */
5838 find_constant_pool_ref (rtx x
, rtx
*ref
)
5843 /* Ignore LTREL_BASE references. */
5844 if (GET_CODE (x
) == UNSPEC
5845 && XINT (x
, 1) == UNSPEC_LTREL_BASE
)
5847 /* Likewise POOL_ENTRY insns. */
5848 if (GET_CODE (x
) == UNSPEC_VOLATILE
5849 && XINT (x
, 1) == UNSPECV_POOL_ENTRY
)
5852 gcc_assert (GET_CODE (x
) != SYMBOL_REF
5853 || !CONSTANT_POOL_ADDRESS_P (x
));
5855 if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_LTREF
)
5857 rtx sym
= XVECEXP (x
, 0, 0);
5858 gcc_assert (GET_CODE (sym
) == SYMBOL_REF
5859 && CONSTANT_POOL_ADDRESS_P (sym
));
5861 if (*ref
== NULL_RTX
)
5864 gcc_assert (*ref
== sym
);
5869 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
5870 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
5874 find_constant_pool_ref (XEXP (x
, i
), ref
);
5876 else if (fmt
[i
] == 'E')
5878 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5879 find_constant_pool_ref (XVECEXP (x
, i
, j
), ref
);
5884 /* Replace every reference to the annotated literal pool
5885 symbol REF in X by its base plus OFFSET. */
5888 replace_constant_pool_ref (rtx
*x
, rtx ref
, rtx offset
)
5893 gcc_assert (*x
!= ref
);
5895 if (GET_CODE (*x
) == UNSPEC
5896 && XINT (*x
, 1) == UNSPEC_LTREF
5897 && XVECEXP (*x
, 0, 0) == ref
)
5899 *x
= gen_rtx_PLUS (Pmode
, XVECEXP (*x
, 0, 1), offset
);
5903 if (GET_CODE (*x
) == PLUS
5904 && GET_CODE (XEXP (*x
, 1)) == CONST_INT
5905 && GET_CODE (XEXP (*x
, 0)) == UNSPEC
5906 && XINT (XEXP (*x
, 0), 1) == UNSPEC_LTREF
5907 && XVECEXP (XEXP (*x
, 0), 0, 0) == ref
)
5909 rtx addr
= gen_rtx_PLUS (Pmode
, XVECEXP (XEXP (*x
, 0), 0, 1), offset
);
5910 *x
= plus_constant (Pmode
, addr
, INTVAL (XEXP (*x
, 1)));
5914 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5915 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5919 replace_constant_pool_ref (&XEXP (*x
, i
), ref
, offset
);
5921 else if (fmt
[i
] == 'E')
5923 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5924 replace_constant_pool_ref (&XVECEXP (*x
, i
, j
), ref
, offset
);
5929 /* Check whether X contains an UNSPEC_LTREL_BASE.
5930 Return its constant pool symbol if found, NULL_RTX otherwise. */
5933 find_ltrel_base (rtx x
)
5938 if (GET_CODE (x
) == UNSPEC
5939 && XINT (x
, 1) == UNSPEC_LTREL_BASE
)
5940 return XVECEXP (x
, 0, 0);
5942 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
5943 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
5947 rtx fnd
= find_ltrel_base (XEXP (x
, i
));
5951 else if (fmt
[i
] == 'E')
5953 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5955 rtx fnd
= find_ltrel_base (XVECEXP (x
, i
, j
));
5965 /* Replace any occurrence of UNSPEC_LTREL_BASE in X with its base. */
5968 replace_ltrel_base (rtx
*x
)
5973 if (GET_CODE (*x
) == UNSPEC
5974 && XINT (*x
, 1) == UNSPEC_LTREL_BASE
)
5976 *x
= XVECEXP (*x
, 0, 1);
5980 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5981 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5985 replace_ltrel_base (&XEXP (*x
, i
));
5987 else if (fmt
[i
] == 'E')
5989 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5990 replace_ltrel_base (&XVECEXP (*x
, i
, j
));
5996 /* We keep a list of constants which we have to add to internal
5997 constant tables in the middle of large functions. */
5999 #define NR_C_MODES 11
6000 enum machine_mode constant_modes
[NR_C_MODES
] =
6002 TFmode
, TImode
, TDmode
,
6003 DFmode
, DImode
, DDmode
,
6004 SFmode
, SImode
, SDmode
,
6011 struct constant
*next
;
6016 struct constant_pool
6018 struct constant_pool
*next
;
6022 rtx emit_pool_after
;
6024 struct constant
*constants
[NR_C_MODES
];
6025 struct constant
*execute
;
6030 /* Allocate new constant_pool structure. */
6032 static struct constant_pool
*
6033 s390_alloc_pool (void)
6035 struct constant_pool
*pool
;
6038 pool
= (struct constant_pool
*) xmalloc (sizeof *pool
);
6040 for (i
= 0; i
< NR_C_MODES
; i
++)
6041 pool
->constants
[i
] = NULL
;
6043 pool
->execute
= NULL
;
6044 pool
->label
= gen_label_rtx ();
6045 pool
->first_insn
= NULL_RTX
;
6046 pool
->pool_insn
= NULL_RTX
;
6047 pool
->insns
= BITMAP_ALLOC (NULL
);
6049 pool
->emit_pool_after
= NULL_RTX
;
6054 /* Create new constant pool covering instructions starting at INSN
6055 and chain it to the end of POOL_LIST. */
6057 static struct constant_pool
*
6058 s390_start_pool (struct constant_pool
**pool_list
, rtx insn
)
6060 struct constant_pool
*pool
, **prev
;
6062 pool
= s390_alloc_pool ();
6063 pool
->first_insn
= insn
;
6065 for (prev
= pool_list
; *prev
; prev
= &(*prev
)->next
)
6072 /* End range of instructions covered by POOL at INSN and emit
6073 placeholder insn representing the pool. */
6076 s390_end_pool (struct constant_pool
*pool
, rtx insn
)
6078 rtx pool_size
= GEN_INT (pool
->size
+ 8 /* alignment slop */);
6081 insn
= get_last_insn ();
6083 pool
->pool_insn
= emit_insn_after (gen_pool (pool_size
), insn
);
6084 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6087 /* Add INSN to the list of insns covered by POOL. */
6090 s390_add_pool_insn (struct constant_pool
*pool
, rtx insn
)
6092 bitmap_set_bit (pool
->insns
, INSN_UID (insn
));
6095 /* Return pool out of POOL_LIST that covers INSN. */
6097 static struct constant_pool
*
6098 s390_find_pool (struct constant_pool
*pool_list
, rtx insn
)
6100 struct constant_pool
*pool
;
6102 for (pool
= pool_list
; pool
; pool
= pool
->next
)
6103 if (bitmap_bit_p (pool
->insns
, INSN_UID (insn
)))
6109 /* Add constant VAL of mode MODE to the constant pool POOL. */
6112 s390_add_constant (struct constant_pool
*pool
, rtx val
, enum machine_mode mode
)
6117 for (i
= 0; i
< NR_C_MODES
; i
++)
6118 if (constant_modes
[i
] == mode
)
6120 gcc_assert (i
!= NR_C_MODES
);
6122 for (c
= pool
->constants
[i
]; c
!= NULL
; c
= c
->next
)
6123 if (rtx_equal_p (val
, c
->value
))
6128 c
= (struct constant
*) xmalloc (sizeof *c
);
6130 c
->label
= gen_label_rtx ();
6131 c
->next
= pool
->constants
[i
];
6132 pool
->constants
[i
] = c
;
6133 pool
->size
+= GET_MODE_SIZE (mode
);
6137 /* Return an rtx that represents the offset of X from the start of
6141 s390_pool_offset (struct constant_pool
*pool
, rtx x
)
6145 label
= gen_rtx_LABEL_REF (GET_MODE (x
), pool
->label
);
6146 x
= gen_rtx_UNSPEC (GET_MODE (x
), gen_rtvec (2, x
, label
),
6147 UNSPEC_POOL_OFFSET
);
6148 return gen_rtx_CONST (GET_MODE (x
), x
);
6151 /* Find constant VAL of mode MODE in the constant pool POOL.
6152 Return an RTX describing the distance from the start of
6153 the pool to the location of the new constant. */
6156 s390_find_constant (struct constant_pool
*pool
, rtx val
,
6157 enum machine_mode mode
)
6162 for (i
= 0; i
< NR_C_MODES
; i
++)
6163 if (constant_modes
[i
] == mode
)
6165 gcc_assert (i
!= NR_C_MODES
);
6167 for (c
= pool
->constants
[i
]; c
!= NULL
; c
= c
->next
)
6168 if (rtx_equal_p (val
, c
->value
))
6173 return s390_pool_offset (pool
, gen_rtx_LABEL_REF (Pmode
, c
->label
));
6176 /* Check whether INSN is an execute. Return the label_ref to its
6177 execute target template if so, NULL_RTX otherwise. */
6180 s390_execute_label (rtx insn
)
6182 if (GET_CODE (insn
) == INSN
6183 && GET_CODE (PATTERN (insn
)) == PARALLEL
6184 && GET_CODE (XVECEXP (PATTERN (insn
), 0, 0)) == UNSPEC
6185 && XINT (XVECEXP (PATTERN (insn
), 0, 0), 1) == UNSPEC_EXECUTE
)
6186 return XVECEXP (XVECEXP (PATTERN (insn
), 0, 0), 0, 2);
6191 /* Add execute target for INSN to the constant pool POOL. */
6194 s390_add_execute (struct constant_pool
*pool
, rtx insn
)
6198 for (c
= pool
->execute
; c
!= NULL
; c
= c
->next
)
6199 if (INSN_UID (insn
) == INSN_UID (c
->value
))
6204 c
= (struct constant
*) xmalloc (sizeof *c
);
6206 c
->label
= gen_label_rtx ();
6207 c
->next
= pool
->execute
;
6213 /* Find execute target for INSN in the constant pool POOL.
6214 Return an RTX describing the distance from the start of
6215 the pool to the location of the execute target. */
6218 s390_find_execute (struct constant_pool
*pool
, rtx insn
)
6222 for (c
= pool
->execute
; c
!= NULL
; c
= c
->next
)
6223 if (INSN_UID (insn
) == INSN_UID (c
->value
))
6228 return s390_pool_offset (pool
, gen_rtx_LABEL_REF (Pmode
, c
->label
));
6231 /* For an execute INSN, extract the execute target template. */
6234 s390_execute_target (rtx insn
)
6236 rtx pattern
= PATTERN (insn
);
6237 gcc_assert (s390_execute_label (insn
));
6239 if (XVECLEN (pattern
, 0) == 2)
6241 pattern
= copy_rtx (XVECEXP (pattern
, 0, 1));
6245 rtvec vec
= rtvec_alloc (XVECLEN (pattern
, 0) - 1);
6248 for (i
= 0; i
< XVECLEN (pattern
, 0) - 1; i
++)
6249 RTVEC_ELT (vec
, i
) = copy_rtx (XVECEXP (pattern
, 0, i
+ 1));
6251 pattern
= gen_rtx_PARALLEL (VOIDmode
, vec
);
6257 /* Indicate that INSN cannot be duplicated. This is the case for
6258 execute insns that carry a unique label. */
6261 s390_cannot_copy_insn_p (rtx insn
)
6263 rtx label
= s390_execute_label (insn
);
6264 return label
&& label
!= const0_rtx
;
6267 /* Dump out the constants in POOL. If REMOTE_LABEL is true,
6268 do not emit the pool base label. */
6271 s390_dump_pool (struct constant_pool
*pool
, bool remote_label
)
6274 rtx insn
= pool
->pool_insn
;
6277 /* Switch to rodata section. */
6278 if (TARGET_CPU_ZARCH
)
6280 insn
= emit_insn_after (gen_pool_section_start (), insn
);
6281 INSN_ADDRESSES_NEW (insn
, -1);
6284 /* Ensure minimum pool alignment. */
6285 if (TARGET_CPU_ZARCH
)
6286 insn
= emit_insn_after (gen_pool_align (GEN_INT (8)), insn
);
6288 insn
= emit_insn_after (gen_pool_align (GEN_INT (4)), insn
);
6289 INSN_ADDRESSES_NEW (insn
, -1);
6291 /* Emit pool base label. */
6294 insn
= emit_label_after (pool
->label
, insn
);
6295 INSN_ADDRESSES_NEW (insn
, -1);
6298 /* Dump constants in descending alignment requirement order,
6299 ensuring proper alignment for every constant. */
6300 for (i
= 0; i
< NR_C_MODES
; i
++)
6301 for (c
= pool
->constants
[i
]; c
; c
= c
->next
)
6303 /* Convert UNSPEC_LTREL_OFFSET unspecs to pool-relative references. */
6304 rtx value
= copy_rtx (c
->value
);
6305 if (GET_CODE (value
) == CONST
6306 && GET_CODE (XEXP (value
, 0)) == UNSPEC
6307 && XINT (XEXP (value
, 0), 1) == UNSPEC_LTREL_OFFSET
6308 && XVECLEN (XEXP (value
, 0), 0) == 1)
6309 value
= s390_pool_offset (pool
, XVECEXP (XEXP (value
, 0), 0, 0));
6311 insn
= emit_label_after (c
->label
, insn
);
6312 INSN_ADDRESSES_NEW (insn
, -1);
6314 value
= gen_rtx_UNSPEC_VOLATILE (constant_modes
[i
],
6315 gen_rtvec (1, value
),
6316 UNSPECV_POOL_ENTRY
);
6317 insn
= emit_insn_after (value
, insn
);
6318 INSN_ADDRESSES_NEW (insn
, -1);
6321 /* Ensure minimum alignment for instructions. */
6322 insn
= emit_insn_after (gen_pool_align (GEN_INT (2)), insn
);
6323 INSN_ADDRESSES_NEW (insn
, -1);
6325 /* Output in-pool execute template insns. */
6326 for (c
= pool
->execute
; c
; c
= c
->next
)
6328 insn
= emit_label_after (c
->label
, insn
);
6329 INSN_ADDRESSES_NEW (insn
, -1);
6331 insn
= emit_insn_after (s390_execute_target (c
->value
), insn
);
6332 INSN_ADDRESSES_NEW (insn
, -1);
6335 /* Switch back to previous section. */
6336 if (TARGET_CPU_ZARCH
)
6338 insn
= emit_insn_after (gen_pool_section_end (), insn
);
6339 INSN_ADDRESSES_NEW (insn
, -1);
6342 insn
= emit_barrier_after (insn
);
6343 INSN_ADDRESSES_NEW (insn
, -1);
6345 /* Remove placeholder insn. */
6346 remove_insn (pool
->pool_insn
);
6349 /* Free all memory used by POOL. */
6352 s390_free_pool (struct constant_pool
*pool
)
6354 struct constant
*c
, *next
;
6357 for (i
= 0; i
< NR_C_MODES
; i
++)
6358 for (c
= pool
->constants
[i
]; c
; c
= next
)
6364 for (c
= pool
->execute
; c
; c
= next
)
6370 BITMAP_FREE (pool
->insns
);
6375 /* Collect main literal pool. Return NULL on overflow. */
6377 static struct constant_pool
*
6378 s390_mainpool_start (void)
6380 struct constant_pool
*pool
;
6383 pool
= s390_alloc_pool ();
6385 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6387 if (GET_CODE (insn
) == INSN
6388 && GET_CODE (PATTERN (insn
)) == SET
6389 && GET_CODE (SET_SRC (PATTERN (insn
))) == UNSPEC_VOLATILE
6390 && XINT (SET_SRC (PATTERN (insn
)), 1) == UNSPECV_MAIN_POOL
)
6392 gcc_assert (!pool
->pool_insn
);
6393 pool
->pool_insn
= insn
;
6396 if (!TARGET_CPU_ZARCH
&& s390_execute_label (insn
))
6398 s390_add_execute (pool
, insn
);
6400 else if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6402 rtx pool_ref
= NULL_RTX
;
6403 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6406 rtx constant
= get_pool_constant (pool_ref
);
6407 enum machine_mode mode
= get_pool_mode (pool_ref
);
6408 s390_add_constant (pool
, constant
, mode
);
6412 /* If hot/cold partitioning is enabled we have to make sure that
6413 the literal pool is emitted in the same section where the
6414 initialization of the literal pool base pointer takes place.
6415 emit_pool_after is only used in the non-overflow case on non
6416 Z cpus where we can emit the literal pool at the end of the
6417 function body within the text section. */
6419 && NOTE_KIND (insn
) == NOTE_INSN_SWITCH_TEXT_SECTIONS
6420 && !pool
->emit_pool_after
)
6421 pool
->emit_pool_after
= PREV_INSN (insn
);
6424 gcc_assert (pool
->pool_insn
|| pool
->size
== 0);
6426 if (pool
->size
>= 4096)
6428 /* We're going to chunkify the pool, so remove the main
6429 pool placeholder insn. */
6430 remove_insn (pool
->pool_insn
);
6432 s390_free_pool (pool
);
6436 /* If the functions ends with the section where the literal pool
6437 should be emitted set the marker to its end. */
6438 if (pool
&& !pool
->emit_pool_after
)
6439 pool
->emit_pool_after
= get_last_insn ();
6444 /* POOL holds the main literal pool as collected by s390_mainpool_start.
6445 Modify the current function to output the pool constants as well as
6446 the pool register setup instruction. */
6449 s390_mainpool_finish (struct constant_pool
*pool
)
6451 rtx base_reg
= cfun
->machine
->base_reg
;
6454 /* If the pool is empty, we're done. */
6455 if (pool
->size
== 0)
6457 /* We don't actually need a base register after all. */
6458 cfun
->machine
->base_reg
= NULL_RTX
;
6460 if (pool
->pool_insn
)
6461 remove_insn (pool
->pool_insn
);
6462 s390_free_pool (pool
);
6466 /* We need correct insn addresses. */
6467 shorten_branches (get_insns ());
6469 /* On zSeries, we use a LARL to load the pool register. The pool is
6470 located in the .rodata section, so we emit it after the function. */
6471 if (TARGET_CPU_ZARCH
)
6473 insn
= gen_main_base_64 (base_reg
, pool
->label
);
6474 insn
= emit_insn_after (insn
, pool
->pool_insn
);
6475 INSN_ADDRESSES_NEW (insn
, -1);
6476 remove_insn (pool
->pool_insn
);
6478 insn
= get_last_insn ();
6479 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6480 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6482 s390_dump_pool (pool
, 0);
6485 /* On S/390, if the total size of the function's code plus literal pool
6486 does not exceed 4096 bytes, we use BASR to set up a function base
6487 pointer, and emit the literal pool at the end of the function. */
6488 else if (INSN_ADDRESSES (INSN_UID (pool
->emit_pool_after
))
6489 + pool
->size
+ 8 /* alignment slop */ < 4096)
6491 insn
= gen_main_base_31_small (base_reg
, pool
->label
);
6492 insn
= emit_insn_after (insn
, pool
->pool_insn
);
6493 INSN_ADDRESSES_NEW (insn
, -1);
6494 remove_insn (pool
->pool_insn
);
6496 insn
= emit_label_after (pool
->label
, insn
);
6497 INSN_ADDRESSES_NEW (insn
, -1);
6499 /* emit_pool_after will be set by s390_mainpool_start to the
6500 last insn of the section where the literal pool should be
6502 insn
= pool
->emit_pool_after
;
6504 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6505 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6507 s390_dump_pool (pool
, 1);
6510 /* Otherwise, we emit an inline literal pool and use BASR to branch
6511 over it, setting up the pool register at the same time. */
6514 rtx pool_end
= gen_label_rtx ();
6516 insn
= gen_main_base_31_large (base_reg
, pool
->label
, pool_end
);
6517 insn
= emit_jump_insn_after (insn
, pool
->pool_insn
);
6518 JUMP_LABEL (insn
) = pool_end
;
6519 INSN_ADDRESSES_NEW (insn
, -1);
6520 remove_insn (pool
->pool_insn
);
6522 insn
= emit_label_after (pool
->label
, insn
);
6523 INSN_ADDRESSES_NEW (insn
, -1);
6525 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6526 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6528 insn
= emit_label_after (pool_end
, pool
->pool_insn
);
6529 INSN_ADDRESSES_NEW (insn
, -1);
6531 s390_dump_pool (pool
, 1);
6535 /* Replace all literal pool references. */
6537 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6540 replace_ltrel_base (&PATTERN (insn
));
6542 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6544 rtx addr
, pool_ref
= NULL_RTX
;
6545 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6548 if (s390_execute_label (insn
))
6549 addr
= s390_find_execute (pool
, insn
);
6551 addr
= s390_find_constant (pool
, get_pool_constant (pool_ref
),
6552 get_pool_mode (pool_ref
));
6554 replace_constant_pool_ref (&PATTERN (insn
), pool_ref
, addr
);
6555 INSN_CODE (insn
) = -1;
6561 /* Free the pool. */
6562 s390_free_pool (pool
);
6565 /* POOL holds the main literal pool as collected by s390_mainpool_start.
6566 We have decided we cannot use this pool, so revert all changes
6567 to the current function that were done by s390_mainpool_start. */
6569 s390_mainpool_cancel (struct constant_pool
*pool
)
6571 /* We didn't actually change the instruction stream, so simply
6572 free the pool memory. */
6573 s390_free_pool (pool
);
6577 /* Chunkify the literal pool. */
6579 #define S390_POOL_CHUNK_MIN 0xc00
6580 #define S390_POOL_CHUNK_MAX 0xe00
6582 static struct constant_pool
*
6583 s390_chunkify_start (void)
6585 struct constant_pool
*curr_pool
= NULL
, *pool_list
= NULL
;
6588 rtx pending_ltrel
= NULL_RTX
;
6591 rtx (*gen_reload_base
) (rtx
, rtx
) =
6592 TARGET_CPU_ZARCH
? gen_reload_base_64
: gen_reload_base_31
;
6595 /* We need correct insn addresses. */
6597 shorten_branches (get_insns ());
6599 /* Scan all insns and move literals to pool chunks. */
6601 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6603 bool section_switch_p
= false;
6605 /* Check for pending LTREL_BASE. */
6608 rtx ltrel_base
= find_ltrel_base (PATTERN (insn
));
6611 gcc_assert (ltrel_base
== pending_ltrel
);
6612 pending_ltrel
= NULL_RTX
;
6616 if (!TARGET_CPU_ZARCH
&& s390_execute_label (insn
))
6619 curr_pool
= s390_start_pool (&pool_list
, insn
);
6621 s390_add_execute (curr_pool
, insn
);
6622 s390_add_pool_insn (curr_pool
, insn
);
6624 else if (GET_CODE (insn
) == INSN
|| CALL_P (insn
))
6626 rtx pool_ref
= NULL_RTX
;
6627 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6630 rtx constant
= get_pool_constant (pool_ref
);
6631 enum machine_mode mode
= get_pool_mode (pool_ref
);
6634 curr_pool
= s390_start_pool (&pool_list
, insn
);
6636 s390_add_constant (curr_pool
, constant
, mode
);
6637 s390_add_pool_insn (curr_pool
, insn
);
6639 /* Don't split the pool chunk between a LTREL_OFFSET load
6640 and the corresponding LTREL_BASE. */
6641 if (GET_CODE (constant
) == CONST
6642 && GET_CODE (XEXP (constant
, 0)) == UNSPEC
6643 && XINT (XEXP (constant
, 0), 1) == UNSPEC_LTREL_OFFSET
)
6645 gcc_assert (!pending_ltrel
);
6646 pending_ltrel
= pool_ref
;
6651 if (GET_CODE (insn
) == JUMP_INSN
|| GET_CODE (insn
) == CODE_LABEL
)
6654 s390_add_pool_insn (curr_pool
, insn
);
6655 /* An LTREL_BASE must follow within the same basic block. */
6656 gcc_assert (!pending_ltrel
);
6660 switch (NOTE_KIND (insn
))
6662 case NOTE_INSN_SWITCH_TEXT_SECTIONS
:
6663 section_switch_p
= true;
6665 case NOTE_INSN_VAR_LOCATION
:
6666 case NOTE_INSN_CALL_ARG_LOCATION
:
6673 || INSN_ADDRESSES_SIZE () <= (size_t) INSN_UID (insn
)
6674 || INSN_ADDRESSES (INSN_UID (insn
)) == -1)
6677 if (TARGET_CPU_ZARCH
)
6679 if (curr_pool
->size
< S390_POOL_CHUNK_MAX
)
6682 s390_end_pool (curr_pool
, NULL_RTX
);
6687 int chunk_size
= INSN_ADDRESSES (INSN_UID (insn
))
6688 - INSN_ADDRESSES (INSN_UID (curr_pool
->first_insn
))
6691 /* We will later have to insert base register reload insns.
6692 Those will have an effect on code size, which we need to
6693 consider here. This calculation makes rather pessimistic
6694 worst-case assumptions. */
6695 if (GET_CODE (insn
) == CODE_LABEL
)
6698 if (chunk_size
< S390_POOL_CHUNK_MIN
6699 && curr_pool
->size
< S390_POOL_CHUNK_MIN
6700 && !section_switch_p
)
6703 /* Pool chunks can only be inserted after BARRIERs ... */
6704 if (GET_CODE (insn
) == BARRIER
)
6706 s390_end_pool (curr_pool
, insn
);
6711 /* ... so if we don't find one in time, create one. */
6712 else if (chunk_size
> S390_POOL_CHUNK_MAX
6713 || curr_pool
->size
> S390_POOL_CHUNK_MAX
6714 || section_switch_p
)
6716 rtx label
, jump
, barrier
, next
, prev
;
6718 if (!section_switch_p
)
6720 /* We can insert the barrier only after a 'real' insn. */
6721 if (GET_CODE (insn
) != INSN
&& GET_CODE (insn
) != CALL_INSN
)
6723 if (get_attr_length (insn
) == 0)
6725 /* Don't separate LTREL_BASE from the corresponding
6726 LTREL_OFFSET load. */
6733 next
= NEXT_INSN (insn
);
6737 && (NOTE_KIND (next
) == NOTE_INSN_VAR_LOCATION
6738 || NOTE_KIND (next
) == NOTE_INSN_CALL_ARG_LOCATION
));
6742 gcc_assert (!pending_ltrel
);
6744 /* The old pool has to end before the section switch
6745 note in order to make it part of the current
6747 insn
= PREV_INSN (insn
);
6750 label
= gen_label_rtx ();
6752 if (prev
&& NOTE_P (prev
))
6753 prev
= prev_nonnote_insn (prev
);
6755 jump
= emit_jump_insn_after_setloc (gen_jump (label
), insn
,
6756 INSN_LOCATOR (prev
));
6758 jump
= emit_jump_insn_after_noloc (gen_jump (label
), insn
);
6759 barrier
= emit_barrier_after (jump
);
6760 insn
= emit_label_after (label
, barrier
);
6761 JUMP_LABEL (jump
) = label
;
6762 LABEL_NUSES (label
) = 1;
6764 INSN_ADDRESSES_NEW (jump
, -1);
6765 INSN_ADDRESSES_NEW (barrier
, -1);
6766 INSN_ADDRESSES_NEW (insn
, -1);
6768 s390_end_pool (curr_pool
, barrier
);
6776 s390_end_pool (curr_pool
, NULL_RTX
);
6777 gcc_assert (!pending_ltrel
);
6779 /* Find all labels that are branched into
6780 from an insn belonging to a different chunk. */
6782 far_labels
= BITMAP_ALLOC (NULL
);
6784 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6786 /* Labels marked with LABEL_PRESERVE_P can be target
6787 of non-local jumps, so we have to mark them.
6788 The same holds for named labels.
6790 Don't do that, however, if it is the label before
6793 if (GET_CODE (insn
) == CODE_LABEL
6794 && (LABEL_PRESERVE_P (insn
) || LABEL_NAME (insn
)))
6796 rtx vec_insn
= next_real_insn (insn
);
6797 rtx vec_pat
= vec_insn
&& GET_CODE (vec_insn
) == JUMP_INSN
?
6798 PATTERN (vec_insn
) : NULL_RTX
;
6800 || !(GET_CODE (vec_pat
) == ADDR_VEC
6801 || GET_CODE (vec_pat
) == ADDR_DIFF_VEC
))
6802 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (insn
));
6805 /* If we have a direct jump (conditional or unconditional)
6806 or a casesi jump, check all potential targets. */
6807 else if (GET_CODE (insn
) == JUMP_INSN
)
6809 rtx pat
= PATTERN (insn
);
6810 if (GET_CODE (pat
) == PARALLEL
&& XVECLEN (pat
, 0) > 2)
6811 pat
= XVECEXP (pat
, 0, 0);
6813 if (GET_CODE (pat
) == SET
)
6815 rtx label
= JUMP_LABEL (insn
);
6818 if (s390_find_pool (pool_list
, label
)
6819 != s390_find_pool (pool_list
, insn
))
6820 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (label
));
6823 else if (GET_CODE (pat
) == PARALLEL
6824 && XVECLEN (pat
, 0) == 2
6825 && GET_CODE (XVECEXP (pat
, 0, 0)) == SET
6826 && GET_CODE (XVECEXP (pat
, 0, 1)) == USE
6827 && GET_CODE (XEXP (XVECEXP (pat
, 0, 1), 0)) == LABEL_REF
)
6829 /* Find the jump table used by this casesi jump. */
6830 rtx vec_label
= XEXP (XEXP (XVECEXP (pat
, 0, 1), 0), 0);
6831 rtx vec_insn
= next_real_insn (vec_label
);
6832 rtx vec_pat
= vec_insn
&& GET_CODE (vec_insn
) == JUMP_INSN
?
6833 PATTERN (vec_insn
) : NULL_RTX
;
6835 && (GET_CODE (vec_pat
) == ADDR_VEC
6836 || GET_CODE (vec_pat
) == ADDR_DIFF_VEC
))
6838 int i
, diff_p
= GET_CODE (vec_pat
) == ADDR_DIFF_VEC
;
6840 for (i
= 0; i
< XVECLEN (vec_pat
, diff_p
); i
++)
6842 rtx label
= XEXP (XVECEXP (vec_pat
, diff_p
, i
), 0);
6844 if (s390_find_pool (pool_list
, label
)
6845 != s390_find_pool (pool_list
, insn
))
6846 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (label
));
6853 /* Insert base register reload insns before every pool. */
6855 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6857 rtx new_insn
= gen_reload_base (cfun
->machine
->base_reg
,
6859 rtx insn
= curr_pool
->first_insn
;
6860 INSN_ADDRESSES_NEW (emit_insn_before (new_insn
, insn
), -1);
6863 /* Insert base register reload insns at every far label. */
6865 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6866 if (GET_CODE (insn
) == CODE_LABEL
6867 && bitmap_bit_p (far_labels
, CODE_LABEL_NUMBER (insn
)))
6869 struct constant_pool
*pool
= s390_find_pool (pool_list
, insn
);
6872 rtx new_insn
= gen_reload_base (cfun
->machine
->base_reg
,
6874 INSN_ADDRESSES_NEW (emit_insn_after (new_insn
, insn
), -1);
6879 BITMAP_FREE (far_labels
);
6882 /* Recompute insn addresses. */
6884 init_insn_lengths ();
6885 shorten_branches (get_insns ());
6890 /* POOL_LIST is a chunk list as prepared by s390_chunkify_start.
6891 After we have decided to use this list, finish implementing
6892 all changes to the current function as required. */
6895 s390_chunkify_finish (struct constant_pool
*pool_list
)
6897 struct constant_pool
*curr_pool
= NULL
;
6901 /* Replace all literal pool references. */
6903 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6906 replace_ltrel_base (&PATTERN (insn
));
6908 curr_pool
= s390_find_pool (pool_list
, insn
);
6912 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6914 rtx addr
, pool_ref
= NULL_RTX
;
6915 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6918 if (s390_execute_label (insn
))
6919 addr
= s390_find_execute (curr_pool
, insn
);
6921 addr
= s390_find_constant (curr_pool
,
6922 get_pool_constant (pool_ref
),
6923 get_pool_mode (pool_ref
));
6925 replace_constant_pool_ref (&PATTERN (insn
), pool_ref
, addr
);
6926 INSN_CODE (insn
) = -1;
6931 /* Dump out all literal pools. */
6933 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6934 s390_dump_pool (curr_pool
, 0);
6936 /* Free pool list. */
6940 struct constant_pool
*next
= pool_list
->next
;
6941 s390_free_pool (pool_list
);
6946 /* POOL_LIST is a chunk list as prepared by s390_chunkify_start.
6947 We have decided we cannot use this list, so revert all changes
6948 to the current function that were done by s390_chunkify_start. */
6951 s390_chunkify_cancel (struct constant_pool
*pool_list
)
6953 struct constant_pool
*curr_pool
= NULL
;
6956 /* Remove all pool placeholder insns. */
6958 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6960 /* Did we insert an extra barrier? Remove it. */
6961 rtx barrier
= PREV_INSN (curr_pool
->pool_insn
);
6962 rtx jump
= barrier
? PREV_INSN (barrier
) : NULL_RTX
;
6963 rtx label
= NEXT_INSN (curr_pool
->pool_insn
);
6965 if (jump
&& GET_CODE (jump
) == JUMP_INSN
6966 && barrier
&& GET_CODE (barrier
) == BARRIER
6967 && label
&& GET_CODE (label
) == CODE_LABEL
6968 && GET_CODE (PATTERN (jump
)) == SET
6969 && SET_DEST (PATTERN (jump
)) == pc_rtx
6970 && GET_CODE (SET_SRC (PATTERN (jump
))) == LABEL_REF
6971 && XEXP (SET_SRC (PATTERN (jump
)), 0) == label
)
6974 remove_insn (barrier
);
6975 remove_insn (label
);
6978 remove_insn (curr_pool
->pool_insn
);
6981 /* Remove all base register reload insns. */
6983 for (insn
= get_insns (); insn
; )
6985 rtx next_insn
= NEXT_INSN (insn
);
6987 if (GET_CODE (insn
) == INSN
6988 && GET_CODE (PATTERN (insn
)) == SET
6989 && GET_CODE (SET_SRC (PATTERN (insn
))) == UNSPEC
6990 && XINT (SET_SRC (PATTERN (insn
)), 1) == UNSPEC_RELOAD_BASE
)
6996 /* Free pool list. */
7000 struct constant_pool
*next
= pool_list
->next
;
7001 s390_free_pool (pool_list
);
7006 /* Output the constant pool entry EXP in mode MODE with alignment ALIGN. */
7009 s390_output_pool_entry (rtx exp
, enum machine_mode mode
, unsigned int align
)
7013 switch (GET_MODE_CLASS (mode
))
7016 case MODE_DECIMAL_FLOAT
:
7017 gcc_assert (GET_CODE (exp
) == CONST_DOUBLE
);
7019 REAL_VALUE_FROM_CONST_DOUBLE (r
, exp
);
7020 assemble_real (r
, mode
, align
);
7024 assemble_integer (exp
, GET_MODE_SIZE (mode
), align
, 1);
7025 mark_symbol_refs_as_used (exp
);
7034 /* Return an RTL expression representing the value of the return address
7035 for the frame COUNT steps up from the current frame. FRAME is the
7036 frame pointer of that frame. */
7039 s390_return_addr_rtx (int count
, rtx frame ATTRIBUTE_UNUSED
)
7044 /* Without backchain, we fail for all but the current frame. */
7046 if (!TARGET_BACKCHAIN
&& count
> 0)
7049 /* For the current frame, we need to make sure the initial
7050 value of RETURN_REGNUM is actually saved. */
7054 /* On non-z architectures branch splitting could overwrite r14. */
7055 if (TARGET_CPU_ZARCH
)
7056 return get_hard_reg_initial_val (Pmode
, RETURN_REGNUM
);
7059 cfun_frame_layout
.save_return_addr_p
= true;
7060 return gen_rtx_MEM (Pmode
, return_address_pointer_rtx
);
7064 if (TARGET_PACKED_STACK
)
7065 offset
= -2 * UNITS_PER_LONG
;
7067 offset
= RETURN_REGNUM
* UNITS_PER_LONG
;
7069 addr
= plus_constant (Pmode
, frame
, offset
);
7070 addr
= memory_address (Pmode
, addr
);
7071 return gen_rtx_MEM (Pmode
, addr
);
7074 /* Return an RTL expression representing the back chain stored in
7075 the current stack frame. */
7078 s390_back_chain_rtx (void)
7082 gcc_assert (TARGET_BACKCHAIN
);
7084 if (TARGET_PACKED_STACK
)
7085 chain
= plus_constant (Pmode
, stack_pointer_rtx
,
7086 STACK_POINTER_OFFSET
- UNITS_PER_LONG
);
7088 chain
= stack_pointer_rtx
;
7090 chain
= gen_rtx_MEM (Pmode
, chain
);
7094 /* Find first call clobbered register unused in a function.
7095 This could be used as base register in a leaf function
7096 or for holding the return address before epilogue. */
7099 find_unused_clobbered_reg (void)
7102 for (i
= 0; i
< 6; i
++)
7103 if (!df_regs_ever_live_p (i
))
7109 /* Helper function for s390_regs_ever_clobbered. Sets the fields in DATA for all
7110 clobbered hard regs in SETREG. */
7113 s390_reg_clobbered_rtx (rtx setreg
, const_rtx set_insn ATTRIBUTE_UNUSED
, void *data
)
7115 int *regs_ever_clobbered
= (int *)data
;
7116 unsigned int i
, regno
;
7117 enum machine_mode mode
= GET_MODE (setreg
);
7119 if (GET_CODE (setreg
) == SUBREG
)
7121 rtx inner
= SUBREG_REG (setreg
);
7122 if (!GENERAL_REG_P (inner
))
7124 regno
= subreg_regno (setreg
);
7126 else if (GENERAL_REG_P (setreg
))
7127 regno
= REGNO (setreg
);
7132 i
< regno
+ HARD_REGNO_NREGS (regno
, mode
);
7134 regs_ever_clobbered
[i
] = 1;
7137 /* Walks through all basic blocks of the current function looking
7138 for clobbered hard regs using s390_reg_clobbered_rtx. The fields
7139 of the passed integer array REGS_EVER_CLOBBERED are set to one for
7140 each of those regs. */
7143 s390_regs_ever_clobbered (int *regs_ever_clobbered
)
7149 memset (regs_ever_clobbered
, 0, 16 * sizeof (int));
7151 /* For non-leaf functions we have to consider all call clobbered regs to be
7155 for (i
= 0; i
< 16; i
++)
7156 regs_ever_clobbered
[i
] = call_really_used_regs
[i
];
7159 /* Make the "magic" eh_return registers live if necessary. For regs_ever_live
7160 this work is done by liveness analysis (mark_regs_live_at_end).
7161 Special care is needed for functions containing landing pads. Landing pads
7162 may use the eh registers, but the code which sets these registers is not
7163 contained in that function. Hence s390_regs_ever_clobbered is not able to
7164 deal with this automatically. */
7165 if (crtl
->calls_eh_return
|| cfun
->machine
->has_landing_pad_p
)
7166 for (i
= 0; EH_RETURN_DATA_REGNO (i
) != INVALID_REGNUM
; i
++)
7167 if (crtl
->calls_eh_return
7168 || (cfun
->machine
->has_landing_pad_p
7169 && df_regs_ever_live_p (EH_RETURN_DATA_REGNO (i
))))
7170 regs_ever_clobbered
[EH_RETURN_DATA_REGNO (i
)] = 1;
7172 /* For nonlocal gotos all call-saved registers have to be saved.
7173 This flag is also set for the unwinding code in libgcc.
7174 See expand_builtin_unwind_init. For regs_ever_live this is done by
7176 if (cfun
->has_nonlocal_label
)
7177 for (i
= 0; i
< 16; i
++)
7178 if (!call_really_used_regs
[i
])
7179 regs_ever_clobbered
[i
] = 1;
7181 FOR_EACH_BB (cur_bb
)
7183 FOR_BB_INSNS (cur_bb
, cur_insn
)
7185 if (INSN_P (cur_insn
))
7186 note_stores (PATTERN (cur_insn
),
7187 s390_reg_clobbered_rtx
,
7188 regs_ever_clobbered
);
7193 /* Determine the frame area which actually has to be accessed
7194 in the function epilogue. The values are stored at the
7195 given pointers AREA_BOTTOM (address of the lowest used stack
7196 address) and AREA_TOP (address of the first item which does
7197 not belong to the stack frame). */
7200 s390_frame_area (int *area_bottom
, int *area_top
)
7208 if (cfun_frame_layout
.first_restore_gpr
!= -1)
7210 b
= (cfun_frame_layout
.gprs_offset
7211 + cfun_frame_layout
.first_restore_gpr
* UNITS_PER_LONG
);
7212 t
= b
+ (cfun_frame_layout
.last_restore_gpr
7213 - cfun_frame_layout
.first_restore_gpr
+ 1) * UNITS_PER_LONG
;
7216 if (TARGET_64BIT
&& cfun_save_high_fprs_p
)
7218 b
= MIN (b
, cfun_frame_layout
.f8_offset
);
7219 t
= MAX (t
, (cfun_frame_layout
.f8_offset
7220 + cfun_frame_layout
.high_fprs
* 8));
7224 for (i
= 2; i
< 4; i
++)
7225 if (cfun_fpr_bit_p (i
))
7227 b
= MIN (b
, cfun_frame_layout
.f4_offset
+ (i
- 2) * 8);
7228 t
= MAX (t
, cfun_frame_layout
.f4_offset
+ (i
- 1) * 8);
7235 /* Fill cfun->machine with info about register usage of current function.
7236 Return in CLOBBERED_REGS which GPRs are currently considered set. */
7239 s390_register_info (int clobbered_regs
[])
7243 /* fprs 8 - 15 are call saved for 64 Bit ABI. */
7244 cfun_frame_layout
.fpr_bitmap
= 0;
7245 cfun_frame_layout
.high_fprs
= 0;
7247 for (i
= 24; i
< 32; i
++)
7248 if (df_regs_ever_live_p (i
) && !global_regs
[i
])
7250 cfun_set_fpr_bit (i
- 16);
7251 cfun_frame_layout
.high_fprs
++;
7254 /* Find first and last gpr to be saved. We trust regs_ever_live
7255 data, except that we don't save and restore global registers.
7257 Also, all registers with special meaning to the compiler need
7258 to be handled extra. */
7260 s390_regs_ever_clobbered (clobbered_regs
);
7262 for (i
= 0; i
< 16; i
++)
7263 clobbered_regs
[i
] = clobbered_regs
[i
] && !global_regs
[i
] && !fixed_regs
[i
];
7265 if (frame_pointer_needed
)
7266 clobbered_regs
[HARD_FRAME_POINTER_REGNUM
] = 1;
7269 clobbered_regs
[PIC_OFFSET_TABLE_REGNUM
]
7270 |= df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM
);
7272 clobbered_regs
[BASE_REGNUM
]
7273 |= (cfun
->machine
->base_reg
7274 && REGNO (cfun
->machine
->base_reg
) == BASE_REGNUM
);
7276 clobbered_regs
[RETURN_REGNUM
]
7278 || TARGET_TPF_PROFILING
7279 || cfun
->machine
->split_branches_pending_p
7280 || cfun_frame_layout
.save_return_addr_p
7281 || crtl
->calls_eh_return
7284 clobbered_regs
[STACK_POINTER_REGNUM
]
7286 || TARGET_TPF_PROFILING
7287 || cfun_save_high_fprs_p
7288 || get_frame_size () > 0
7289 || cfun
->calls_alloca
7292 for (i
= 6; i
< 16; i
++)
7293 if (df_regs_ever_live_p (i
) || clobbered_regs
[i
])
7295 for (j
= 15; j
> i
; j
--)
7296 if (df_regs_ever_live_p (j
) || clobbered_regs
[j
])
7301 /* Nothing to save/restore. */
7302 cfun_frame_layout
.first_save_gpr_slot
= -1;
7303 cfun_frame_layout
.last_save_gpr_slot
= -1;
7304 cfun_frame_layout
.first_save_gpr
= -1;
7305 cfun_frame_layout
.first_restore_gpr
= -1;
7306 cfun_frame_layout
.last_save_gpr
= -1;
7307 cfun_frame_layout
.last_restore_gpr
= -1;
7311 /* Save slots for gprs from i to j. */
7312 cfun_frame_layout
.first_save_gpr_slot
= i
;
7313 cfun_frame_layout
.last_save_gpr_slot
= j
;
7315 for (i
= cfun_frame_layout
.first_save_gpr_slot
;
7316 i
< cfun_frame_layout
.last_save_gpr_slot
+ 1;
7318 if (clobbered_regs
[i
])
7321 for (j
= cfun_frame_layout
.last_save_gpr_slot
; j
> i
; j
--)
7322 if (clobbered_regs
[j
])
7325 if (i
== cfun_frame_layout
.last_save_gpr_slot
+ 1)
7327 /* Nothing to save/restore. */
7328 cfun_frame_layout
.first_save_gpr
= -1;
7329 cfun_frame_layout
.first_restore_gpr
= -1;
7330 cfun_frame_layout
.last_save_gpr
= -1;
7331 cfun_frame_layout
.last_restore_gpr
= -1;
7335 /* Save / Restore from gpr i to j. */
7336 cfun_frame_layout
.first_save_gpr
= i
;
7337 cfun_frame_layout
.first_restore_gpr
= i
;
7338 cfun_frame_layout
.last_save_gpr
= j
;
7339 cfun_frame_layout
.last_restore_gpr
= j
;
7345 /* Varargs functions need to save gprs 2 to 6. */
7346 if (cfun
->va_list_gpr_size
7347 && crtl
->args
.info
.gprs
< GP_ARG_NUM_REG
)
7349 int min_gpr
= crtl
->args
.info
.gprs
;
7350 int max_gpr
= min_gpr
+ cfun
->va_list_gpr_size
;
7351 if (max_gpr
> GP_ARG_NUM_REG
)
7352 max_gpr
= GP_ARG_NUM_REG
;
7354 if (cfun_frame_layout
.first_save_gpr
== -1
7355 || cfun_frame_layout
.first_save_gpr
> 2 + min_gpr
)
7357 cfun_frame_layout
.first_save_gpr
= 2 + min_gpr
;
7358 cfun_frame_layout
.first_save_gpr_slot
= 2 + min_gpr
;
7361 if (cfun_frame_layout
.last_save_gpr
== -1
7362 || cfun_frame_layout
.last_save_gpr
< 2 + max_gpr
- 1)
7364 cfun_frame_layout
.last_save_gpr
= 2 + max_gpr
- 1;
7365 cfun_frame_layout
.last_save_gpr_slot
= 2 + max_gpr
- 1;
7369 /* Mark f0, f2 for 31 bit and f0-f4 for 64 bit to be saved. */
7370 if (TARGET_HARD_FLOAT
&& cfun
->va_list_fpr_size
7371 && crtl
->args
.info
.fprs
< FP_ARG_NUM_REG
)
7373 int min_fpr
= crtl
->args
.info
.fprs
;
7374 int max_fpr
= min_fpr
+ cfun
->va_list_fpr_size
;
7375 if (max_fpr
> FP_ARG_NUM_REG
)
7376 max_fpr
= FP_ARG_NUM_REG
;
7378 /* ??? This is currently required to ensure proper location
7379 of the fpr save slots within the va_list save area. */
7380 if (TARGET_PACKED_STACK
)
7383 for (i
= min_fpr
; i
< max_fpr
; i
++)
7384 cfun_set_fpr_bit (i
);
7389 for (i
= 2; i
< 4; i
++)
7390 if (df_regs_ever_live_p (i
+ 16) && !global_regs
[i
+ 16])
7391 cfun_set_fpr_bit (i
);
7394 /* Fill cfun->machine with info about frame of current function. */
7397 s390_frame_info (void)
7401 cfun_frame_layout
.frame_size
= get_frame_size ();
7402 if (!TARGET_64BIT
&& cfun_frame_layout
.frame_size
> 0x7fff0000)
7403 fatal_error ("total size of local variables exceeds architecture limit");
7405 if (!TARGET_PACKED_STACK
)
7407 cfun_frame_layout
.backchain_offset
= 0;
7408 cfun_frame_layout
.f0_offset
= 16 * UNITS_PER_LONG
;
7409 cfun_frame_layout
.f4_offset
= cfun_frame_layout
.f0_offset
+ 2 * 8;
7410 cfun_frame_layout
.f8_offset
= -cfun_frame_layout
.high_fprs
* 8;
7411 cfun_frame_layout
.gprs_offset
= (cfun_frame_layout
.first_save_gpr_slot
7414 else if (TARGET_BACKCHAIN
) /* kernel stack layout */
7416 cfun_frame_layout
.backchain_offset
= (STACK_POINTER_OFFSET
7418 cfun_frame_layout
.gprs_offset
7419 = (cfun_frame_layout
.backchain_offset
7420 - (STACK_POINTER_REGNUM
- cfun_frame_layout
.first_save_gpr_slot
+ 1)
7425 cfun_frame_layout
.f4_offset
7426 = (cfun_frame_layout
.gprs_offset
7427 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7429 cfun_frame_layout
.f0_offset
7430 = (cfun_frame_layout
.f4_offset
7431 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7435 /* On 31 bit we have to care about alignment of the
7436 floating point regs to provide fastest access. */
7437 cfun_frame_layout
.f0_offset
7438 = ((cfun_frame_layout
.gprs_offset
7439 & ~(STACK_BOUNDARY
/ BITS_PER_UNIT
- 1))
7440 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7442 cfun_frame_layout
.f4_offset
7443 = (cfun_frame_layout
.f0_offset
7444 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7447 else /* no backchain */
7449 cfun_frame_layout
.f4_offset
7450 = (STACK_POINTER_OFFSET
7451 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7453 cfun_frame_layout
.f0_offset
7454 = (cfun_frame_layout
.f4_offset
7455 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7457 cfun_frame_layout
.gprs_offset
7458 = cfun_frame_layout
.f0_offset
- cfun_gprs_save_area_size
;
7462 && !TARGET_TPF_PROFILING
7463 && cfun_frame_layout
.frame_size
== 0
7464 && !cfun_save_high_fprs_p
7465 && !cfun
->calls_alloca
7469 if (!TARGET_PACKED_STACK
)
7470 cfun_frame_layout
.frame_size
+= (STACK_POINTER_OFFSET
7471 + crtl
->outgoing_args_size
7472 + cfun_frame_layout
.high_fprs
* 8);
7475 if (TARGET_BACKCHAIN
)
7476 cfun_frame_layout
.frame_size
+= UNITS_PER_LONG
;
7478 /* No alignment trouble here because f8-f15 are only saved under
7480 cfun_frame_layout
.f8_offset
= (MIN (MIN (cfun_frame_layout
.f0_offset
,
7481 cfun_frame_layout
.f4_offset
),
7482 cfun_frame_layout
.gprs_offset
)
7483 - cfun_frame_layout
.high_fprs
* 8);
7485 cfun_frame_layout
.frame_size
+= cfun_frame_layout
.high_fprs
* 8;
7487 for (i
= 0; i
< 8; i
++)
7488 if (cfun_fpr_bit_p (i
))
7489 cfun_frame_layout
.frame_size
+= 8;
7491 cfun_frame_layout
.frame_size
+= cfun_gprs_save_area_size
;
7493 /* If under 31 bit an odd number of gprs has to be saved we have to adjust
7494 the frame size to sustain 8 byte alignment of stack frames. */
7495 cfun_frame_layout
.frame_size
= ((cfun_frame_layout
.frame_size
+
7496 STACK_BOUNDARY
/ BITS_PER_UNIT
- 1)
7497 & ~(STACK_BOUNDARY
/ BITS_PER_UNIT
- 1));
7499 cfun_frame_layout
.frame_size
+= crtl
->outgoing_args_size
;
7503 /* Generate frame layout. Fills in register and frame data for the current
7504 function in cfun->machine. This routine can be called multiple times;
7505 it will re-do the complete frame layout every time. */
7508 s390_init_frame_layout (void)
7510 HOST_WIDE_INT frame_size
;
7512 int clobbered_regs
[16];
7514 /* On S/390 machines, we may need to perform branch splitting, which
7515 will require both base and return address register. We have no
7516 choice but to assume we're going to need them until right at the
7517 end of the machine dependent reorg phase. */
7518 if (!TARGET_CPU_ZARCH
)
7519 cfun
->machine
->split_branches_pending_p
= true;
7523 frame_size
= cfun_frame_layout
.frame_size
;
7525 /* Try to predict whether we'll need the base register. */
7526 base_used
= cfun
->machine
->split_branches_pending_p
7527 || crtl
->uses_const_pool
7528 || (!DISP_IN_RANGE (frame_size
)
7529 && !CONST_OK_FOR_K (frame_size
));
7531 /* Decide which register to use as literal pool base. In small
7532 leaf functions, try to use an unused call-clobbered register
7533 as base register to avoid save/restore overhead. */
7535 cfun
->machine
->base_reg
= NULL_RTX
;
7536 else if (crtl
->is_leaf
&& !df_regs_ever_live_p (5))
7537 cfun
->machine
->base_reg
= gen_rtx_REG (Pmode
, 5);
7539 cfun
->machine
->base_reg
= gen_rtx_REG (Pmode
, BASE_REGNUM
);
7541 s390_register_info (clobbered_regs
);
7544 while (frame_size
!= cfun_frame_layout
.frame_size
);
7547 /* Update frame layout. Recompute actual register save data based on
7548 current info and update regs_ever_live for the special registers.
7549 May be called multiple times, but may never cause *more* registers
7550 to be saved than s390_init_frame_layout allocated room for. */
7553 s390_update_frame_layout (void)
7555 int clobbered_regs
[16];
7557 s390_register_info (clobbered_regs
);
7559 df_set_regs_ever_live (BASE_REGNUM
,
7560 clobbered_regs
[BASE_REGNUM
] ? true : false);
7561 df_set_regs_ever_live (RETURN_REGNUM
,
7562 clobbered_regs
[RETURN_REGNUM
] ? true : false);
7563 df_set_regs_ever_live (STACK_POINTER_REGNUM
,
7564 clobbered_regs
[STACK_POINTER_REGNUM
] ? true : false);
7566 if (cfun
->machine
->base_reg
)
7567 df_set_regs_ever_live (REGNO (cfun
->machine
->base_reg
), true);
7570 /* Return true if it is legal to put a value with MODE into REGNO. */
7573 s390_hard_regno_mode_ok (unsigned int regno
, enum machine_mode mode
)
7575 switch (REGNO_REG_CLASS (regno
))
7578 if (REGNO_PAIR_OK (regno
, mode
))
7580 if (mode
== SImode
|| mode
== DImode
)
7583 if (FLOAT_MODE_P (mode
) && GET_MODE_CLASS (mode
) != MODE_VECTOR_FLOAT
)
7588 if (FRAME_REGNO_P (regno
) && mode
== Pmode
)
7593 if (REGNO_PAIR_OK (regno
, mode
))
7596 || (mode
!= TFmode
&& mode
!= TCmode
&& mode
!= TDmode
))
7601 if (GET_MODE_CLASS (mode
) == MODE_CC
)
7605 if (REGNO_PAIR_OK (regno
, mode
))
7607 if (mode
== SImode
|| mode
== Pmode
)
7618 /* Return nonzero if register OLD_REG can be renamed to register NEW_REG. */
7621 s390_hard_regno_rename_ok (unsigned int old_reg
, unsigned int new_reg
)
7623 /* Once we've decided upon a register to use as base register, it must
7624 no longer be used for any other purpose. */
7625 if (cfun
->machine
->base_reg
)
7626 if (REGNO (cfun
->machine
->base_reg
) == old_reg
7627 || REGNO (cfun
->machine
->base_reg
) == new_reg
)
7633 /* Maximum number of registers to represent a value of mode MODE
7634 in a register of class RCLASS. */
7637 s390_class_max_nregs (enum reg_class rclass
, enum machine_mode mode
)
7642 if (GET_MODE_CLASS (mode
) == MODE_COMPLEX_FLOAT
)
7643 return 2 * ((GET_MODE_SIZE (mode
) / 2 + 8 - 1) / 8);
7645 return (GET_MODE_SIZE (mode
) + 8 - 1) / 8;
7647 return (GET_MODE_SIZE (mode
) + 4 - 1) / 4;
7651 return (GET_MODE_SIZE (mode
) + UNITS_PER_WORD
- 1) / UNITS_PER_WORD
;
7654 /* Return true if register FROM can be eliminated via register TO. */
7657 s390_can_eliminate (const int from
, const int to
)
7659 /* On zSeries machines, we have not marked the base register as fixed.
7660 Instead, we have an elimination rule BASE_REGNUM -> BASE_REGNUM.
7661 If a function requires the base register, we say here that this
7662 elimination cannot be performed. This will cause reload to free
7663 up the base register (as if it were fixed). On the other hand,
7664 if the current function does *not* require the base register, we
7665 say here the elimination succeeds, which in turn allows reload
7666 to allocate the base register for any other purpose. */
7667 if (from
== BASE_REGNUM
&& to
== BASE_REGNUM
)
7669 if (TARGET_CPU_ZARCH
)
7671 s390_init_frame_layout ();
7672 return cfun
->machine
->base_reg
== NULL_RTX
;
7678 /* Everything else must point into the stack frame. */
7679 gcc_assert (to
== STACK_POINTER_REGNUM
7680 || to
== HARD_FRAME_POINTER_REGNUM
);
7682 gcc_assert (from
== FRAME_POINTER_REGNUM
7683 || from
== ARG_POINTER_REGNUM
7684 || from
== RETURN_ADDRESS_POINTER_REGNUM
);
7686 /* Make sure we actually saved the return address. */
7687 if (from
== RETURN_ADDRESS_POINTER_REGNUM
)
7688 if (!crtl
->calls_eh_return
7690 && !cfun_frame_layout
.save_return_addr_p
)
7696 /* Return offset between register FROM and TO initially after prolog. */
7699 s390_initial_elimination_offset (int from
, int to
)
7701 HOST_WIDE_INT offset
;
7704 /* ??? Why are we called for non-eliminable pairs? */
7705 if (!s390_can_eliminate (from
, to
))
7710 case FRAME_POINTER_REGNUM
:
7711 offset
= (get_frame_size()
7712 + STACK_POINTER_OFFSET
7713 + crtl
->outgoing_args_size
);
7716 case ARG_POINTER_REGNUM
:
7717 s390_init_frame_layout ();
7718 offset
= cfun_frame_layout
.frame_size
+ STACK_POINTER_OFFSET
;
7721 case RETURN_ADDRESS_POINTER_REGNUM
:
7722 s390_init_frame_layout ();
7723 index
= RETURN_REGNUM
- cfun_frame_layout
.first_save_gpr_slot
;
7724 gcc_assert (index
>= 0);
7725 offset
= cfun_frame_layout
.frame_size
+ cfun_frame_layout
.gprs_offset
;
7726 offset
+= index
* UNITS_PER_LONG
;
7740 /* Emit insn to save fpr REGNUM at offset OFFSET relative
7741 to register BASE. Return generated insn. */
7744 save_fpr (rtx base
, int offset
, int regnum
)
7747 addr
= gen_rtx_MEM (DFmode
, plus_constant (Pmode
, base
, offset
));
7749 if (regnum
>= 16 && regnum
<= (16 + FP_ARG_NUM_REG
))
7750 set_mem_alias_set (addr
, get_varargs_alias_set ());
7752 set_mem_alias_set (addr
, get_frame_alias_set ());
7754 return emit_move_insn (addr
, gen_rtx_REG (DFmode
, regnum
));
7757 /* Emit insn to restore fpr REGNUM from offset OFFSET relative
7758 to register BASE. Return generated insn. */
7761 restore_fpr (rtx base
, int offset
, int regnum
)
7764 addr
= gen_rtx_MEM (DFmode
, plus_constant (Pmode
, base
, offset
));
7765 set_mem_alias_set (addr
, get_frame_alias_set ());
7767 return emit_move_insn (gen_rtx_REG (DFmode
, regnum
), addr
);
7770 /* Return true if REGNO is a global register, but not one
7771 of the special ones that need to be saved/restored in anyway. */
7774 global_not_special_regno_p (int regno
)
7776 return (global_regs
[regno
]
7777 /* These registers are special and need to be
7778 restored in any case. */
7779 && !(regno
== STACK_POINTER_REGNUM
7780 || regno
== RETURN_REGNUM
7781 || regno
== BASE_REGNUM
7782 || (flag_pic
&& regno
== (int)PIC_OFFSET_TABLE_REGNUM
)));
7785 /* Generate insn to save registers FIRST to LAST into
7786 the register save area located at offset OFFSET
7787 relative to register BASE. */
7790 save_gprs (rtx base
, int offset
, int first
, int last
)
7792 rtx addr
, insn
, note
;
7795 addr
= plus_constant (Pmode
, base
, offset
);
7796 addr
= gen_rtx_MEM (Pmode
, addr
);
7798 set_mem_alias_set (addr
, get_frame_alias_set ());
7800 /* Special-case single register. */
7804 insn
= gen_movdi (addr
, gen_rtx_REG (Pmode
, first
));
7806 insn
= gen_movsi (addr
, gen_rtx_REG (Pmode
, first
));
7808 if (!global_not_special_regno_p (first
))
7809 RTX_FRAME_RELATED_P (insn
) = 1;
7814 insn
= gen_store_multiple (addr
,
7815 gen_rtx_REG (Pmode
, first
),
7816 GEN_INT (last
- first
+ 1));
7818 if (first
<= 6 && cfun
->stdarg
)
7819 for (i
= 0; i
< XVECLEN (PATTERN (insn
), 0); i
++)
7821 rtx mem
= XEXP (XVECEXP (PATTERN (insn
), 0, i
), 0);
7824 set_mem_alias_set (mem
, get_varargs_alias_set ());
7827 /* We need to set the FRAME_RELATED flag on all SETs
7828 inside the store-multiple pattern.
7830 However, we must not emit DWARF records for registers 2..5
7831 if they are stored for use by variable arguments ...
7833 ??? Unfortunately, it is not enough to simply not the
7834 FRAME_RELATED flags for those SETs, because the first SET
7835 of the PARALLEL is always treated as if it had the flag
7836 set, even if it does not. Therefore we emit a new pattern
7837 without those registers as REG_FRAME_RELATED_EXPR note. */
7839 if (first
>= 6 && !global_not_special_regno_p (first
))
7841 rtx pat
= PATTERN (insn
);
7843 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
7844 if (GET_CODE (XVECEXP (pat
, 0, i
)) == SET
7845 && !global_not_special_regno_p (REGNO (SET_SRC (XVECEXP (pat
,
7847 RTX_FRAME_RELATED_P (XVECEXP (pat
, 0, i
)) = 1;
7849 RTX_FRAME_RELATED_P (insn
) = 1;
7855 for (start
= first
>= 6 ? first
: 6; start
<= last
; start
++)
7856 if (!global_not_special_regno_p (start
))
7862 addr
= plus_constant (Pmode
, base
,
7863 offset
+ (start
- first
) * UNITS_PER_LONG
);
7864 note
= gen_store_multiple (gen_rtx_MEM (Pmode
, addr
),
7865 gen_rtx_REG (Pmode
, start
),
7866 GEN_INT (last
- start
+ 1));
7867 note
= PATTERN (note
);
7869 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
, note
);
7871 for (i
= 0; i
< XVECLEN (note
, 0); i
++)
7872 if (GET_CODE (XVECEXP (note
, 0, i
)) == SET
7873 && !global_not_special_regno_p (REGNO (SET_SRC (XVECEXP (note
,
7875 RTX_FRAME_RELATED_P (XVECEXP (note
, 0, i
)) = 1;
7877 RTX_FRAME_RELATED_P (insn
) = 1;
7883 /* Generate insn to restore registers FIRST to LAST from
7884 the register save area located at offset OFFSET
7885 relative to register BASE. */
7888 restore_gprs (rtx base
, int offset
, int first
, int last
)
7892 addr
= plus_constant (Pmode
, base
, offset
);
7893 addr
= gen_rtx_MEM (Pmode
, addr
);
7894 set_mem_alias_set (addr
, get_frame_alias_set ());
7896 /* Special-case single register. */
7900 insn
= gen_movdi (gen_rtx_REG (Pmode
, first
), addr
);
7902 insn
= gen_movsi (gen_rtx_REG (Pmode
, first
), addr
);
7907 insn
= gen_load_multiple (gen_rtx_REG (Pmode
, first
),
7909 GEN_INT (last
- first
+ 1));
7913 /* Return insn sequence to load the GOT register. */
7915 static GTY(()) rtx got_symbol
;
7917 s390_load_got (void)
7921 /* We cannot use pic_offset_table_rtx here since we use this
7922 function also for non-pic if __tls_get_offset is called and in
7923 that case PIC_OFFSET_TABLE_REGNUM as well as pic_offset_table_rtx
7925 rtx got_rtx
= gen_rtx_REG (Pmode
, 12);
7929 got_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "_GLOBAL_OFFSET_TABLE_");
7930 SYMBOL_REF_FLAGS (got_symbol
) = SYMBOL_FLAG_LOCAL
;
7935 if (TARGET_CPU_ZARCH
)
7937 emit_move_insn (got_rtx
, got_symbol
);
7943 offset
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, got_symbol
),
7944 UNSPEC_LTREL_OFFSET
);
7945 offset
= gen_rtx_CONST (Pmode
, offset
);
7946 offset
= force_const_mem (Pmode
, offset
);
7948 emit_move_insn (got_rtx
, offset
);
7950 offset
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, XEXP (offset
, 0)),
7952 offset
= gen_rtx_PLUS (Pmode
, got_rtx
, offset
);
7954 emit_move_insn (got_rtx
, offset
);
7957 insns
= get_insns ();
7962 /* This ties together stack memory (MEM with an alias set of frame_alias_set)
7963 and the change to the stack pointer. */
7966 s390_emit_stack_tie (void)
7968 rtx mem
= gen_frame_mem (BLKmode
,
7969 gen_rtx_REG (Pmode
, STACK_POINTER_REGNUM
));
7971 emit_insn (gen_stack_tie (mem
));
7974 /* Expand the prologue into a bunch of separate insns. */
7977 s390_emit_prologue (void)
7985 /* Complete frame layout. */
7987 s390_update_frame_layout ();
7989 /* Annotate all constant pool references to let the scheduler know
7990 they implicitly use the base register. */
7992 push_topmost_sequence ();
7994 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
7997 annotate_constant_pool_refs (&PATTERN (insn
));
7998 df_insn_rescan (insn
);
8001 pop_topmost_sequence ();
8003 /* Choose best register to use for temp use within prologue.
8004 See below for why TPF must use the register 1. */
8006 if (!has_hard_reg_initial_val (Pmode
, RETURN_REGNUM
)
8008 && !TARGET_TPF_PROFILING
)
8009 temp_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
8011 temp_reg
= gen_rtx_REG (Pmode
, 1);
8013 /* Save call saved gprs. */
8014 if (cfun_frame_layout
.first_save_gpr
!= -1)
8016 insn
= save_gprs (stack_pointer_rtx
,
8017 cfun_frame_layout
.gprs_offset
+
8018 UNITS_PER_LONG
* (cfun_frame_layout
.first_save_gpr
8019 - cfun_frame_layout
.first_save_gpr_slot
),
8020 cfun_frame_layout
.first_save_gpr
,
8021 cfun_frame_layout
.last_save_gpr
);
8025 /* Dummy insn to mark literal pool slot. */
8027 if (cfun
->machine
->base_reg
)
8028 emit_insn (gen_main_pool (cfun
->machine
->base_reg
));
8030 offset
= cfun_frame_layout
.f0_offset
;
8032 /* Save f0 and f2. */
8033 for (i
= 0; i
< 2; i
++)
8035 if (cfun_fpr_bit_p (i
))
8037 save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8040 else if (!TARGET_PACKED_STACK
)
8044 /* Save f4 and f6. */
8045 offset
= cfun_frame_layout
.f4_offset
;
8046 for (i
= 2; i
< 4; i
++)
8048 if (cfun_fpr_bit_p (i
))
8050 insn
= save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8053 /* If f4 and f6 are call clobbered they are saved due to stdargs and
8054 therefore are not frame related. */
8055 if (!call_really_used_regs
[i
+ 16])
8056 RTX_FRAME_RELATED_P (insn
) = 1;
8058 else if (!TARGET_PACKED_STACK
)
8062 if (TARGET_PACKED_STACK
8063 && cfun_save_high_fprs_p
8064 && cfun_frame_layout
.f8_offset
+ cfun_frame_layout
.high_fprs
* 8 > 0)
8066 offset
= (cfun_frame_layout
.f8_offset
8067 + (cfun_frame_layout
.high_fprs
- 1) * 8);
8069 for (i
= 15; i
> 7 && offset
>= 0; i
--)
8070 if (cfun_fpr_bit_p (i
))
8072 insn
= save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8074 RTX_FRAME_RELATED_P (insn
) = 1;
8077 if (offset
>= cfun_frame_layout
.f8_offset
)
8081 if (!TARGET_PACKED_STACK
)
8082 next_fpr
= cfun_save_high_fprs_p
? 31 : 0;
8084 if (flag_stack_usage_info
)
8085 current_function_static_stack_size
= cfun_frame_layout
.frame_size
;
8087 /* Decrement stack pointer. */
8089 if (cfun_frame_layout
.frame_size
> 0)
8091 rtx frame_off
= GEN_INT (-cfun_frame_layout
.frame_size
);
8094 if (s390_stack_size
)
8096 HOST_WIDE_INT stack_guard
;
8098 if (s390_stack_guard
)
8099 stack_guard
= s390_stack_guard
;
8102 /* If no value for stack guard is provided the smallest power of 2
8103 larger than the current frame size is chosen. */
8105 while (stack_guard
< cfun_frame_layout
.frame_size
)
8109 if (cfun_frame_layout
.frame_size
>= s390_stack_size
)
8111 warning (0, "frame size of function %qs is %wd"
8112 " bytes exceeding user provided stack limit of "
8114 "An unconditional trap is added.",
8115 current_function_name(), cfun_frame_layout
.frame_size
,
8117 emit_insn (gen_trap ());
8121 /* stack_guard has to be smaller than s390_stack_size.
8122 Otherwise we would emit an AND with zero which would
8123 not match the test under mask pattern. */
8124 if (stack_guard
>= s390_stack_size
)
8126 warning (0, "frame size of function %qs is %wd"
8127 " bytes which is more than half the stack size. "
8128 "The dynamic check would not be reliable. "
8129 "No check emitted for this function.",
8130 current_function_name(),
8131 cfun_frame_layout
.frame_size
);
8135 HOST_WIDE_INT stack_check_mask
= ((s390_stack_size
- 1)
8136 & ~(stack_guard
- 1));
8138 rtx t
= gen_rtx_AND (Pmode
, stack_pointer_rtx
,
8139 GEN_INT (stack_check_mask
));
8141 emit_insn (gen_ctrapdi4 (gen_rtx_EQ (VOIDmode
,
8143 t
, const0_rtx
, const0_rtx
));
8145 emit_insn (gen_ctrapsi4 (gen_rtx_EQ (VOIDmode
,
8147 t
, const0_rtx
, const0_rtx
));
8152 if (s390_warn_framesize
> 0
8153 && cfun_frame_layout
.frame_size
>= s390_warn_framesize
)
8154 warning (0, "frame size of %qs is %wd bytes",
8155 current_function_name (), cfun_frame_layout
.frame_size
);
8157 if (s390_warn_dynamicstack_p
&& cfun
->calls_alloca
)
8158 warning (0, "%qs uses dynamic stack allocation", current_function_name ());
8160 /* Save incoming stack pointer into temp reg. */
8161 if (TARGET_BACKCHAIN
|| next_fpr
)
8162 insn
= emit_insn (gen_move_insn (temp_reg
, stack_pointer_rtx
));
8164 /* Subtract frame size from stack pointer. */
8166 if (DISP_IN_RANGE (INTVAL (frame_off
)))
8168 insn
= gen_rtx_SET (VOIDmode
, stack_pointer_rtx
,
8169 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
8171 insn
= emit_insn (insn
);
8175 if (!CONST_OK_FOR_K (INTVAL (frame_off
)))
8176 frame_off
= force_const_mem (Pmode
, frame_off
);
8178 insn
= emit_insn (gen_add2_insn (stack_pointer_rtx
, frame_off
));
8179 annotate_constant_pool_refs (&PATTERN (insn
));
8182 RTX_FRAME_RELATED_P (insn
) = 1;
8183 real_frame_off
= GEN_INT (-cfun_frame_layout
.frame_size
);
8184 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
8185 gen_rtx_SET (VOIDmode
, stack_pointer_rtx
,
8186 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
8189 /* Set backchain. */
8191 if (TARGET_BACKCHAIN
)
8193 if (cfun_frame_layout
.backchain_offset
)
8194 addr
= gen_rtx_MEM (Pmode
,
8195 plus_constant (Pmode
, stack_pointer_rtx
,
8196 cfun_frame_layout
.backchain_offset
));
8198 addr
= gen_rtx_MEM (Pmode
, stack_pointer_rtx
);
8199 set_mem_alias_set (addr
, get_frame_alias_set ());
8200 insn
= emit_insn (gen_move_insn (addr
, temp_reg
));
8203 /* If we support non-call exceptions (e.g. for Java),
8204 we need to make sure the backchain pointer is set up
8205 before any possibly trapping memory access. */
8206 if (TARGET_BACKCHAIN
&& cfun
->can_throw_non_call_exceptions
)
8208 addr
= gen_rtx_MEM (BLKmode
, gen_rtx_SCRATCH (VOIDmode
));
8209 emit_clobber (addr
);
8213 /* Save fprs 8 - 15 (64 bit ABI). */
8215 if (cfun_save_high_fprs_p
&& next_fpr
)
8217 /* If the stack might be accessed through a different register
8218 we have to make sure that the stack pointer decrement is not
8219 moved below the use of the stack slots. */
8220 s390_emit_stack_tie ();
8222 insn
= emit_insn (gen_add2_insn (temp_reg
,
8223 GEN_INT (cfun_frame_layout
.f8_offset
)));
8227 for (i
= 24; i
<= next_fpr
; i
++)
8228 if (cfun_fpr_bit_p (i
- 16))
8230 rtx addr
= plus_constant (Pmode
, stack_pointer_rtx
,
8231 cfun_frame_layout
.frame_size
8232 + cfun_frame_layout
.f8_offset
8235 insn
= save_fpr (temp_reg
, offset
, i
);
8237 RTX_FRAME_RELATED_P (insn
) = 1;
8238 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
8239 gen_rtx_SET (VOIDmode
,
8240 gen_rtx_MEM (DFmode
, addr
),
8241 gen_rtx_REG (DFmode
, i
)));
8245 /* Set frame pointer, if needed. */
8247 if (frame_pointer_needed
)
8249 insn
= emit_move_insn (hard_frame_pointer_rtx
, stack_pointer_rtx
);
8250 RTX_FRAME_RELATED_P (insn
) = 1;
8253 /* Set up got pointer, if needed. */
8255 if (flag_pic
&& df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM
))
8257 rtx insns
= s390_load_got ();
8259 for (insn
= insns
; insn
; insn
= NEXT_INSN (insn
))
8260 annotate_constant_pool_refs (&PATTERN (insn
));
8265 if (TARGET_TPF_PROFILING
)
8267 /* Generate a BAS instruction to serve as a function
8268 entry intercept to facilitate the use of tracing
8269 algorithms located at the branch target. */
8270 emit_insn (gen_prologue_tpf ());
8272 /* Emit a blockage here so that all code
8273 lies between the profiling mechanisms. */
8274 emit_insn (gen_blockage ());
8278 /* Expand the epilogue into a bunch of separate insns. */
8281 s390_emit_epilogue (bool sibcall
)
8283 rtx frame_pointer
, return_reg
, cfa_restores
= NULL_RTX
;
8284 int area_bottom
, area_top
, offset
= 0;
8289 if (TARGET_TPF_PROFILING
)
8292 /* Generate a BAS instruction to serve as a function
8293 entry intercept to facilitate the use of tracing
8294 algorithms located at the branch target. */
8296 /* Emit a blockage here so that all code
8297 lies between the profiling mechanisms. */
8298 emit_insn (gen_blockage ());
8300 emit_insn (gen_epilogue_tpf ());
8303 /* Check whether to use frame or stack pointer for restore. */
8305 frame_pointer
= (frame_pointer_needed
8306 ? hard_frame_pointer_rtx
: stack_pointer_rtx
);
8308 s390_frame_area (&area_bottom
, &area_top
);
8310 /* Check whether we can access the register save area.
8311 If not, increment the frame pointer as required. */
8313 if (area_top
<= area_bottom
)
8315 /* Nothing to restore. */
8317 else if (DISP_IN_RANGE (cfun_frame_layout
.frame_size
+ area_bottom
)
8318 && DISP_IN_RANGE (cfun_frame_layout
.frame_size
+ area_top
- 1))
8320 /* Area is in range. */
8321 offset
= cfun_frame_layout
.frame_size
;
8325 rtx insn
, frame_off
, cfa
;
8327 offset
= area_bottom
< 0 ? -area_bottom
: 0;
8328 frame_off
= GEN_INT (cfun_frame_layout
.frame_size
- offset
);
8330 cfa
= gen_rtx_SET (VOIDmode
, frame_pointer
,
8331 gen_rtx_PLUS (Pmode
, frame_pointer
, frame_off
));
8332 if (DISP_IN_RANGE (INTVAL (frame_off
)))
8334 insn
= gen_rtx_SET (VOIDmode
, frame_pointer
,
8335 gen_rtx_PLUS (Pmode
, frame_pointer
, frame_off
));
8336 insn
= emit_insn (insn
);
8340 if (!CONST_OK_FOR_K (INTVAL (frame_off
)))
8341 frame_off
= force_const_mem (Pmode
, frame_off
);
8343 insn
= emit_insn (gen_add2_insn (frame_pointer
, frame_off
));
8344 annotate_constant_pool_refs (&PATTERN (insn
));
8346 add_reg_note (insn
, REG_CFA_ADJUST_CFA
, cfa
);
8347 RTX_FRAME_RELATED_P (insn
) = 1;
8350 /* Restore call saved fprs. */
8354 if (cfun_save_high_fprs_p
)
8356 next_offset
= cfun_frame_layout
.f8_offset
;
8357 for (i
= 24; i
< 32; i
++)
8359 if (cfun_fpr_bit_p (i
- 16))
8361 restore_fpr (frame_pointer
,
8362 offset
+ next_offset
, i
);
8364 = alloc_reg_note (REG_CFA_RESTORE
,
8365 gen_rtx_REG (DFmode
, i
), cfa_restores
);
8374 next_offset
= cfun_frame_layout
.f4_offset
;
8375 for (i
= 18; i
< 20; i
++)
8377 if (cfun_fpr_bit_p (i
- 16))
8379 restore_fpr (frame_pointer
,
8380 offset
+ next_offset
, i
);
8382 = alloc_reg_note (REG_CFA_RESTORE
,
8383 gen_rtx_REG (DFmode
, i
), cfa_restores
);
8386 else if (!TARGET_PACKED_STACK
)
8392 /* Return register. */
8394 return_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
8396 /* Restore call saved gprs. */
8398 if (cfun_frame_layout
.first_restore_gpr
!= -1)
8403 /* Check for global register and save them
8404 to stack location from where they get restored. */
8406 for (i
= cfun_frame_layout
.first_restore_gpr
;
8407 i
<= cfun_frame_layout
.last_restore_gpr
;
8410 if (global_not_special_regno_p (i
))
8412 addr
= plus_constant (Pmode
, frame_pointer
,
8413 offset
+ cfun_frame_layout
.gprs_offset
8414 + (i
- cfun_frame_layout
.first_save_gpr_slot
)
8416 addr
= gen_rtx_MEM (Pmode
, addr
);
8417 set_mem_alias_set (addr
, get_frame_alias_set ());
8418 emit_move_insn (addr
, gen_rtx_REG (Pmode
, i
));
8422 = alloc_reg_note (REG_CFA_RESTORE
,
8423 gen_rtx_REG (Pmode
, i
), cfa_restores
);
8428 /* Fetch return address from stack before load multiple,
8429 this will do good for scheduling. */
8431 if (cfun_frame_layout
.save_return_addr_p
8432 || (cfun_frame_layout
.first_restore_gpr
< BASE_REGNUM
8433 && cfun_frame_layout
.last_restore_gpr
> RETURN_REGNUM
))
8435 int return_regnum
= find_unused_clobbered_reg();
8438 return_reg
= gen_rtx_REG (Pmode
, return_regnum
);
8440 addr
= plus_constant (Pmode
, frame_pointer
,
8441 offset
+ cfun_frame_layout
.gprs_offset
8443 - cfun_frame_layout
.first_save_gpr_slot
)
8445 addr
= gen_rtx_MEM (Pmode
, addr
);
8446 set_mem_alias_set (addr
, get_frame_alias_set ());
8447 emit_move_insn (return_reg
, addr
);
8451 insn
= restore_gprs (frame_pointer
,
8452 offset
+ cfun_frame_layout
.gprs_offset
8453 + (cfun_frame_layout
.first_restore_gpr
8454 - cfun_frame_layout
.first_save_gpr_slot
)
8456 cfun_frame_layout
.first_restore_gpr
,
8457 cfun_frame_layout
.last_restore_gpr
);
8458 insn
= emit_insn (insn
);
8459 REG_NOTES (insn
) = cfa_restores
;
8460 add_reg_note (insn
, REG_CFA_DEF_CFA
,
8461 plus_constant (Pmode
, stack_pointer_rtx
,
8462 STACK_POINTER_OFFSET
));
8463 RTX_FRAME_RELATED_P (insn
) = 1;
8469 /* Return to caller. */
8471 p
= rtvec_alloc (2);
8473 RTVEC_ELT (p
, 0) = ret_rtx
;
8474 RTVEC_ELT (p
, 1) = gen_rtx_USE (VOIDmode
, return_reg
);
8475 emit_jump_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
8480 /* Return the size in bytes of a function argument of
8481 type TYPE and/or mode MODE. At least one of TYPE or
8482 MODE must be specified. */
8485 s390_function_arg_size (enum machine_mode mode
, const_tree type
)
8488 return int_size_in_bytes (type
);
8490 /* No type info available for some library calls ... */
8491 if (mode
!= BLKmode
)
8492 return GET_MODE_SIZE (mode
);
8494 /* If we have neither type nor mode, abort */
8498 /* Return true if a function argument of type TYPE and mode MODE
8499 is to be passed in a floating-point register, if available. */
8502 s390_function_arg_float (enum machine_mode mode
, const_tree type
)
8504 int size
= s390_function_arg_size (mode
, type
);
8508 /* Soft-float changes the ABI: no floating-point registers are used. */
8509 if (TARGET_SOFT_FLOAT
)
8512 /* No type info available for some library calls ... */
8514 return mode
== SFmode
|| mode
== DFmode
|| mode
== SDmode
|| mode
== DDmode
;
8516 /* The ABI says that record types with a single member are treated
8517 just like that member would be. */
8518 while (TREE_CODE (type
) == RECORD_TYPE
)
8520 tree field
, single
= NULL_TREE
;
8522 for (field
= TYPE_FIELDS (type
); field
; field
= DECL_CHAIN (field
))
8524 if (TREE_CODE (field
) != FIELD_DECL
)
8527 if (single
== NULL_TREE
)
8528 single
= TREE_TYPE (field
);
8533 if (single
== NULL_TREE
)
8539 return TREE_CODE (type
) == REAL_TYPE
;
8542 /* Return true if a function argument of type TYPE and mode MODE
8543 is to be passed in an integer register, or a pair of integer
8544 registers, if available. */
8547 s390_function_arg_integer (enum machine_mode mode
, const_tree type
)
8549 int size
= s390_function_arg_size (mode
, type
);
8553 /* No type info available for some library calls ... */
8555 return GET_MODE_CLASS (mode
) == MODE_INT
8556 || (TARGET_SOFT_FLOAT
&& SCALAR_FLOAT_MODE_P (mode
));
8558 /* We accept small integral (and similar) types. */
8559 if (INTEGRAL_TYPE_P (type
)
8560 || POINTER_TYPE_P (type
)
8561 || TREE_CODE (type
) == NULLPTR_TYPE
8562 || TREE_CODE (type
) == OFFSET_TYPE
8563 || (TARGET_SOFT_FLOAT
&& TREE_CODE (type
) == REAL_TYPE
))
8566 /* We also accept structs of size 1, 2, 4, 8 that are not
8567 passed in floating-point registers. */
8568 if (AGGREGATE_TYPE_P (type
)
8569 && exact_log2 (size
) >= 0
8570 && !s390_function_arg_float (mode
, type
))
8576 /* Return 1 if a function argument of type TYPE and mode MODE
8577 is to be passed by reference. The ABI specifies that only
8578 structures of size 1, 2, 4, or 8 bytes are passed by value,
8579 all other structures (and complex numbers) are passed by
8583 s390_pass_by_reference (cumulative_args_t ca ATTRIBUTE_UNUSED
,
8584 enum machine_mode mode
, const_tree type
,
8585 bool named ATTRIBUTE_UNUSED
)
8587 int size
= s390_function_arg_size (mode
, type
);
8593 if (AGGREGATE_TYPE_P (type
) && exact_log2 (size
) < 0)
8596 if (TREE_CODE (type
) == COMPLEX_TYPE
8597 || TREE_CODE (type
) == VECTOR_TYPE
)
8604 /* Update the data in CUM to advance over an argument of mode MODE and
8605 data type TYPE. (TYPE is null for libcalls where that information
8606 may not be available.). The boolean NAMED specifies whether the
8607 argument is a named argument (as opposed to an unnamed argument
8608 matching an ellipsis). */
8611 s390_function_arg_advance (cumulative_args_t cum_v
, enum machine_mode mode
,
8612 const_tree type
, bool named ATTRIBUTE_UNUSED
)
8614 CUMULATIVE_ARGS
*cum
= get_cumulative_args (cum_v
);
8616 if (s390_function_arg_float (mode
, type
))
8620 else if (s390_function_arg_integer (mode
, type
))
8622 int size
= s390_function_arg_size (mode
, type
);
8623 cum
->gprs
+= ((size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
);
8629 /* Define where to put the arguments to a function.
8630 Value is zero to push the argument on the stack,
8631 or a hard register in which to store the argument.
8633 MODE is the argument's machine mode.
8634 TYPE is the data type of the argument (as a tree).
8635 This is null for libcalls where that information may
8637 CUM is a variable of type CUMULATIVE_ARGS which gives info about
8638 the preceding args and about the function being called.
8639 NAMED is nonzero if this argument is a named parameter
8640 (otherwise it is an extra parameter matching an ellipsis).
8642 On S/390, we use general purpose registers 2 through 6 to
8643 pass integer, pointer, and certain structure arguments, and
8644 floating point registers 0 and 2 (0, 2, 4, and 6 on 64-bit)
8645 to pass floating point arguments. All remaining arguments
8646 are pushed to the stack. */
8649 s390_function_arg (cumulative_args_t cum_v
, enum machine_mode mode
,
8650 const_tree type
, bool named ATTRIBUTE_UNUSED
)
8652 CUMULATIVE_ARGS
*cum
= get_cumulative_args (cum_v
);
8654 if (s390_function_arg_float (mode
, type
))
8656 if (cum
->fprs
+ 1 > FP_ARG_NUM_REG
)
8659 return gen_rtx_REG (mode
, cum
->fprs
+ 16);
8661 else if (s390_function_arg_integer (mode
, type
))
8663 int size
= s390_function_arg_size (mode
, type
);
8664 int n_gprs
= (size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
;
8666 if (cum
->gprs
+ n_gprs
> GP_ARG_NUM_REG
)
8668 else if (n_gprs
== 1 || UNITS_PER_WORD
== UNITS_PER_LONG
)
8669 return gen_rtx_REG (mode
, cum
->gprs
+ 2);
8670 else if (n_gprs
== 2)
8672 rtvec p
= rtvec_alloc (2);
8675 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, cum
->gprs
+ 2),
8678 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, cum
->gprs
+ 3),
8681 return gen_rtx_PARALLEL (mode
, p
);
8685 /* After the real arguments, expand_call calls us once again
8686 with a void_type_node type. Whatever we return here is
8687 passed as operand 2 to the call expanders.
8689 We don't need this feature ... */
8690 else if (type
== void_type_node
)
8696 /* Return true if return values of type TYPE should be returned
8697 in a memory buffer whose address is passed by the caller as
8698 hidden first argument. */
8701 s390_return_in_memory (const_tree type
, const_tree fundecl ATTRIBUTE_UNUSED
)
8703 /* We accept small integral (and similar) types. */
8704 if (INTEGRAL_TYPE_P (type
)
8705 || POINTER_TYPE_P (type
)
8706 || TREE_CODE (type
) == OFFSET_TYPE
8707 || TREE_CODE (type
) == REAL_TYPE
)
8708 return int_size_in_bytes (type
) > 8;
8710 /* Aggregates and similar constructs are always returned
8712 if (AGGREGATE_TYPE_P (type
)
8713 || TREE_CODE (type
) == COMPLEX_TYPE
8714 || TREE_CODE (type
) == VECTOR_TYPE
)
8717 /* ??? We get called on all sorts of random stuff from
8718 aggregate_value_p. We can't abort, but it's not clear
8719 what's safe to return. Pretend it's a struct I guess. */
8723 /* Function arguments and return values are promoted to word size. */
8725 static enum machine_mode
8726 s390_promote_function_mode (const_tree type
, enum machine_mode mode
,
8728 const_tree fntype ATTRIBUTE_UNUSED
,
8729 int for_return ATTRIBUTE_UNUSED
)
8731 if (INTEGRAL_MODE_P (mode
)
8732 && GET_MODE_SIZE (mode
) < UNITS_PER_LONG
)
8734 if (type
!= NULL_TREE
&& POINTER_TYPE_P (type
))
8735 *punsignedp
= POINTERS_EXTEND_UNSIGNED
;
8742 /* Define where to return a (scalar) value of type RET_TYPE.
8743 If RET_TYPE is null, define where to return a (scalar)
8744 value of mode MODE from a libcall. */
8747 s390_function_and_libcall_value (enum machine_mode mode
,
8748 const_tree ret_type
,
8749 const_tree fntype_or_decl
,
8750 bool outgoing ATTRIBUTE_UNUSED
)
8752 /* For normal functions perform the promotion as
8753 promote_function_mode would do. */
8756 int unsignedp
= TYPE_UNSIGNED (ret_type
);
8757 mode
= promote_function_mode (ret_type
, mode
, &unsignedp
,
8761 gcc_assert (GET_MODE_CLASS (mode
) == MODE_INT
|| SCALAR_FLOAT_MODE_P (mode
));
8762 gcc_assert (GET_MODE_SIZE (mode
) <= 8);
8764 if (TARGET_HARD_FLOAT
&& SCALAR_FLOAT_MODE_P (mode
))
8765 return gen_rtx_REG (mode
, 16);
8766 else if (GET_MODE_SIZE (mode
) <= UNITS_PER_LONG
8767 || UNITS_PER_LONG
== UNITS_PER_WORD
)
8768 return gen_rtx_REG (mode
, 2);
8769 else if (GET_MODE_SIZE (mode
) == 2 * UNITS_PER_LONG
)
8771 /* This case is triggered when returning a 64 bit value with
8772 -m31 -mzarch. Although the value would fit into a single
8773 register it has to be forced into a 32 bit register pair in
8774 order to match the ABI. */
8775 rtvec p
= rtvec_alloc (2);
8778 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, 2), const0_rtx
);
8780 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, 3), GEN_INT (4));
8782 return gen_rtx_PARALLEL (mode
, p
);
8788 /* Define where to return a scalar return value of type RET_TYPE. */
8791 s390_function_value (const_tree ret_type
, const_tree fn_decl_or_type
,
8794 return s390_function_and_libcall_value (TYPE_MODE (ret_type
), ret_type
,
8795 fn_decl_or_type
, outgoing
);
8798 /* Define where to return a scalar libcall return value of mode
8802 s390_libcall_value (enum machine_mode mode
, const_rtx fun ATTRIBUTE_UNUSED
)
8804 return s390_function_and_libcall_value (mode
, NULL_TREE
,
8809 /* Create and return the va_list datatype.
8811 On S/390, va_list is an array type equivalent to
8813 typedef struct __va_list_tag
8817 void *__overflow_arg_area;
8818 void *__reg_save_area;
8821 where __gpr and __fpr hold the number of general purpose
8822 or floating point arguments used up to now, respectively,
8823 __overflow_arg_area points to the stack location of the
8824 next argument passed on the stack, and __reg_save_area
8825 always points to the start of the register area in the
8826 call frame of the current function. The function prologue
8827 saves all registers used for argument passing into this
8828 area if the function uses variable arguments. */
8831 s390_build_builtin_va_list (void)
8833 tree f_gpr
, f_fpr
, f_ovf
, f_sav
, record
, type_decl
;
8835 record
= lang_hooks
.types
.make_type (RECORD_TYPE
);
8838 build_decl (BUILTINS_LOCATION
,
8839 TYPE_DECL
, get_identifier ("__va_list_tag"), record
);
8841 f_gpr
= build_decl (BUILTINS_LOCATION
,
8842 FIELD_DECL
, get_identifier ("__gpr"),
8843 long_integer_type_node
);
8844 f_fpr
= build_decl (BUILTINS_LOCATION
,
8845 FIELD_DECL
, get_identifier ("__fpr"),
8846 long_integer_type_node
);
8847 f_ovf
= build_decl (BUILTINS_LOCATION
,
8848 FIELD_DECL
, get_identifier ("__overflow_arg_area"),
8850 f_sav
= build_decl (BUILTINS_LOCATION
,
8851 FIELD_DECL
, get_identifier ("__reg_save_area"),
8854 va_list_gpr_counter_field
= f_gpr
;
8855 va_list_fpr_counter_field
= f_fpr
;
8857 DECL_FIELD_CONTEXT (f_gpr
) = record
;
8858 DECL_FIELD_CONTEXT (f_fpr
) = record
;
8859 DECL_FIELD_CONTEXT (f_ovf
) = record
;
8860 DECL_FIELD_CONTEXT (f_sav
) = record
;
8862 TYPE_STUB_DECL (record
) = type_decl
;
8863 TYPE_NAME (record
) = type_decl
;
8864 TYPE_FIELDS (record
) = f_gpr
;
8865 DECL_CHAIN (f_gpr
) = f_fpr
;
8866 DECL_CHAIN (f_fpr
) = f_ovf
;
8867 DECL_CHAIN (f_ovf
) = f_sav
;
8869 layout_type (record
);
8871 /* The correct type is an array type of one element. */
8872 return build_array_type (record
, build_index_type (size_zero_node
));
8875 /* Implement va_start by filling the va_list structure VALIST.
8876 STDARG_P is always true, and ignored.
8877 NEXTARG points to the first anonymous stack argument.
8879 The following global variables are used to initialize
8880 the va_list structure:
8883 holds number of gprs and fprs used for named arguments.
8884 crtl->args.arg_offset_rtx:
8885 holds the offset of the first anonymous stack argument
8886 (relative to the virtual arg pointer). */
8889 s390_va_start (tree valist
, rtx nextarg ATTRIBUTE_UNUSED
)
8891 HOST_WIDE_INT n_gpr
, n_fpr
;
8893 tree f_gpr
, f_fpr
, f_ovf
, f_sav
;
8894 tree gpr
, fpr
, ovf
, sav
, t
;
8896 f_gpr
= TYPE_FIELDS (TREE_TYPE (va_list_type_node
));
8897 f_fpr
= DECL_CHAIN (f_gpr
);
8898 f_ovf
= DECL_CHAIN (f_fpr
);
8899 f_sav
= DECL_CHAIN (f_ovf
);
8901 valist
= build_simple_mem_ref (valist
);
8902 gpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_gpr
), valist
, f_gpr
, NULL_TREE
);
8903 fpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_fpr
), valist
, f_fpr
, NULL_TREE
);
8904 ovf
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovf
), valist
, f_ovf
, NULL_TREE
);
8905 sav
= build3 (COMPONENT_REF
, TREE_TYPE (f_sav
), valist
, f_sav
, NULL_TREE
);
8907 /* Count number of gp and fp argument registers used. */
8909 n_gpr
= crtl
->args
.info
.gprs
;
8910 n_fpr
= crtl
->args
.info
.fprs
;
8912 if (cfun
->va_list_gpr_size
)
8914 t
= build2 (MODIFY_EXPR
, TREE_TYPE (gpr
), gpr
,
8915 build_int_cst (NULL_TREE
, n_gpr
));
8916 TREE_SIDE_EFFECTS (t
) = 1;
8917 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8920 if (cfun
->va_list_fpr_size
)
8922 t
= build2 (MODIFY_EXPR
, TREE_TYPE (fpr
), fpr
,
8923 build_int_cst (NULL_TREE
, n_fpr
));
8924 TREE_SIDE_EFFECTS (t
) = 1;
8925 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8928 /* Find the overflow area. */
8929 if (n_gpr
+ cfun
->va_list_gpr_size
> GP_ARG_NUM_REG
8930 || n_fpr
+ cfun
->va_list_fpr_size
> FP_ARG_NUM_REG
)
8932 t
= make_tree (TREE_TYPE (ovf
), virtual_incoming_args_rtx
);
8934 off
= INTVAL (crtl
->args
.arg_offset_rtx
);
8935 off
= off
< 0 ? 0 : off
;
8936 if (TARGET_DEBUG_ARG
)
8937 fprintf (stderr
, "va_start: n_gpr = %d, n_fpr = %d off %d\n",
8938 (int)n_gpr
, (int)n_fpr
, off
);
8940 t
= fold_build_pointer_plus_hwi (t
, off
);
8942 t
= build2 (MODIFY_EXPR
, TREE_TYPE (ovf
), ovf
, t
);
8943 TREE_SIDE_EFFECTS (t
) = 1;
8944 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8947 /* Find the register save area. */
8948 if ((cfun
->va_list_gpr_size
&& n_gpr
< GP_ARG_NUM_REG
)
8949 || (cfun
->va_list_fpr_size
&& n_fpr
< FP_ARG_NUM_REG
))
8951 t
= make_tree (TREE_TYPE (sav
), return_address_pointer_rtx
);
8952 t
= fold_build_pointer_plus_hwi (t
, -RETURN_REGNUM
* UNITS_PER_LONG
);
8954 t
= build2 (MODIFY_EXPR
, TREE_TYPE (sav
), sav
, t
);
8955 TREE_SIDE_EFFECTS (t
) = 1;
8956 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8960 /* Implement va_arg by updating the va_list structure
8961 VALIST as required to retrieve an argument of type
8962 TYPE, and returning that argument.
8964 Generates code equivalent to:
8966 if (integral value) {
8967 if (size <= 4 && args.gpr < 5 ||
8968 size > 4 && args.gpr < 4 )
8969 ret = args.reg_save_area[args.gpr+8]
8971 ret = *args.overflow_arg_area++;
8972 } else if (float value) {
8974 ret = args.reg_save_area[args.fpr+64]
8976 ret = *args.overflow_arg_area++;
8977 } else if (aggregate value) {
8979 ret = *args.reg_save_area[args.gpr]
8981 ret = **args.overflow_arg_area++;
8985 s390_gimplify_va_arg (tree valist
, tree type
, gimple_seq
*pre_p
,
8986 gimple_seq
*post_p ATTRIBUTE_UNUSED
)
8988 tree f_gpr
, f_fpr
, f_ovf
, f_sav
;
8989 tree gpr
, fpr
, ovf
, sav
, reg
, t
, u
;
8990 int indirect_p
, size
, n_reg
, sav_ofs
, sav_scale
, max_reg
;
8991 tree lab_false
, lab_over
, addr
;
8993 f_gpr
= TYPE_FIELDS (TREE_TYPE (va_list_type_node
));
8994 f_fpr
= DECL_CHAIN (f_gpr
);
8995 f_ovf
= DECL_CHAIN (f_fpr
);
8996 f_sav
= DECL_CHAIN (f_ovf
);
8998 valist
= build_va_arg_indirect_ref (valist
);
8999 gpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_gpr
), valist
, f_gpr
, NULL_TREE
);
9000 fpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_fpr
), valist
, f_fpr
, NULL_TREE
);
9001 sav
= build3 (COMPONENT_REF
, TREE_TYPE (f_sav
), valist
, f_sav
, NULL_TREE
);
9003 /* The tree for args* cannot be shared between gpr/fpr and ovf since
9004 both appear on a lhs. */
9005 valist
= unshare_expr (valist
);
9006 ovf
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovf
), valist
, f_ovf
, NULL_TREE
);
9008 size
= int_size_in_bytes (type
);
9010 if (pass_by_reference (NULL
, TYPE_MODE (type
), type
, false))
9012 if (TARGET_DEBUG_ARG
)
9014 fprintf (stderr
, "va_arg: aggregate type");
9018 /* Aggregates are passed by reference. */
9023 /* kernel stack layout on 31 bit: It is assumed here that no padding
9024 will be added by s390_frame_info because for va_args always an even
9025 number of gprs has to be saved r15-r2 = 14 regs. */
9026 sav_ofs
= 2 * UNITS_PER_LONG
;
9027 sav_scale
= UNITS_PER_LONG
;
9028 size
= UNITS_PER_LONG
;
9029 max_reg
= GP_ARG_NUM_REG
- n_reg
;
9031 else if (s390_function_arg_float (TYPE_MODE (type
), type
))
9033 if (TARGET_DEBUG_ARG
)
9035 fprintf (stderr
, "va_arg: float type");
9039 /* FP args go in FP registers, if present. */
9043 sav_ofs
= 16 * UNITS_PER_LONG
;
9045 max_reg
= FP_ARG_NUM_REG
- n_reg
;
9049 if (TARGET_DEBUG_ARG
)
9051 fprintf (stderr
, "va_arg: other type");
9055 /* Otherwise into GP registers. */
9058 n_reg
= (size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
;
9060 /* kernel stack layout on 31 bit: It is assumed here that no padding
9061 will be added by s390_frame_info because for va_args always an even
9062 number of gprs has to be saved r15-r2 = 14 regs. */
9063 sav_ofs
= 2 * UNITS_PER_LONG
;
9065 if (size
< UNITS_PER_LONG
)
9066 sav_ofs
+= UNITS_PER_LONG
- size
;
9068 sav_scale
= UNITS_PER_LONG
;
9069 max_reg
= GP_ARG_NUM_REG
- n_reg
;
9072 /* Pull the value out of the saved registers ... */
9074 lab_false
= create_artificial_label (UNKNOWN_LOCATION
);
9075 lab_over
= create_artificial_label (UNKNOWN_LOCATION
);
9076 addr
= create_tmp_var (ptr_type_node
, "addr");
9078 t
= fold_convert (TREE_TYPE (reg
), size_int (max_reg
));
9079 t
= build2 (GT_EXPR
, boolean_type_node
, reg
, t
);
9080 u
= build1 (GOTO_EXPR
, void_type_node
, lab_false
);
9081 t
= build3 (COND_EXPR
, void_type_node
, t
, u
, NULL_TREE
);
9082 gimplify_and_add (t
, pre_p
);
9084 t
= fold_build_pointer_plus_hwi (sav
, sav_ofs
);
9085 u
= build2 (MULT_EXPR
, TREE_TYPE (reg
), reg
,
9086 fold_convert (TREE_TYPE (reg
), size_int (sav_scale
)));
9087 t
= fold_build_pointer_plus (t
, u
);
9089 gimplify_assign (addr
, t
, pre_p
);
9091 gimple_seq_add_stmt (pre_p
, gimple_build_goto (lab_over
));
9093 gimple_seq_add_stmt (pre_p
, gimple_build_label (lab_false
));
9096 /* ... Otherwise out of the overflow area. */
9099 if (size
< UNITS_PER_LONG
)
9100 t
= fold_build_pointer_plus_hwi (t
, UNITS_PER_LONG
- size
);
9102 gimplify_expr (&t
, pre_p
, NULL
, is_gimple_val
, fb_rvalue
);
9104 gimplify_assign (addr
, t
, pre_p
);
9106 t
= fold_build_pointer_plus_hwi (t
, size
);
9107 gimplify_assign (ovf
, t
, pre_p
);
9109 gimple_seq_add_stmt (pre_p
, gimple_build_label (lab_over
));
9112 /* Increment register save count. */
9114 u
= build2 (PREINCREMENT_EXPR
, TREE_TYPE (reg
), reg
,
9115 fold_convert (TREE_TYPE (reg
), size_int (n_reg
)));
9116 gimplify_and_add (u
, pre_p
);
9120 t
= build_pointer_type_for_mode (build_pointer_type (type
),
9122 addr
= fold_convert (t
, addr
);
9123 addr
= build_va_arg_indirect_ref (addr
);
9127 t
= build_pointer_type_for_mode (type
, ptr_mode
, true);
9128 addr
= fold_convert (t
, addr
);
9131 return build_va_arg_indirect_ref (addr
);
9139 S390_BUILTIN_THREAD_POINTER
,
9140 S390_BUILTIN_SET_THREAD_POINTER
,
9145 static enum insn_code
const code_for_builtin_64
[S390_BUILTIN_max
] = {
9150 static enum insn_code
const code_for_builtin_31
[S390_BUILTIN_max
] = {
9156 s390_init_builtins (void)
9160 ftype
= build_function_type_list (ptr_type_node
, NULL_TREE
);
9161 add_builtin_function ("__builtin_thread_pointer", ftype
,
9162 S390_BUILTIN_THREAD_POINTER
, BUILT_IN_MD
,
9165 ftype
= build_function_type_list (void_type_node
, ptr_type_node
, NULL_TREE
);
9166 add_builtin_function ("__builtin_set_thread_pointer", ftype
,
9167 S390_BUILTIN_SET_THREAD_POINTER
, BUILT_IN_MD
,
9171 /* Expand an expression EXP that calls a built-in function,
9172 with result going to TARGET if that's convenient
9173 (and in mode MODE if that's convenient).
9174 SUBTARGET may be used as the target for computing one of EXP's operands.
9175 IGNORE is nonzero if the value is to be ignored. */
9178 s390_expand_builtin (tree exp
, rtx target
, rtx subtarget ATTRIBUTE_UNUSED
,
9179 enum machine_mode mode ATTRIBUTE_UNUSED
,
9180 int ignore ATTRIBUTE_UNUSED
)
9184 enum insn_code
const *code_for_builtin
=
9185 TARGET_64BIT
? code_for_builtin_64
: code_for_builtin_31
;
9187 tree fndecl
= TREE_OPERAND (CALL_EXPR_FN (exp
), 0);
9188 unsigned int fcode
= DECL_FUNCTION_CODE (fndecl
);
9189 enum insn_code icode
;
9190 rtx op
[MAX_ARGS
], pat
;
9194 call_expr_arg_iterator iter
;
9196 if (fcode
>= S390_BUILTIN_max
)
9197 internal_error ("bad builtin fcode");
9198 icode
= code_for_builtin
[fcode
];
9200 internal_error ("bad builtin fcode");
9202 nonvoid
= TREE_TYPE (TREE_TYPE (fndecl
)) != void_type_node
;
9205 FOR_EACH_CALL_EXPR_ARG (arg
, iter
, exp
)
9207 const struct insn_operand_data
*insn_op
;
9209 if (arg
== error_mark_node
)
9211 if (arity
> MAX_ARGS
)
9214 insn_op
= &insn_data
[icode
].operand
[arity
+ nonvoid
];
9216 op
[arity
] = expand_expr (arg
, NULL_RTX
, insn_op
->mode
, EXPAND_NORMAL
);
9218 if (!(*insn_op
->predicate
) (op
[arity
], insn_op
->mode
))
9219 op
[arity
] = copy_to_mode_reg (insn_op
->mode
, op
[arity
]);
9225 enum machine_mode tmode
= insn_data
[icode
].operand
[0].mode
;
9227 || GET_MODE (target
) != tmode
9228 || !(*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
9229 target
= gen_reg_rtx (tmode
);
9235 pat
= GEN_FCN (icode
) (target
);
9239 pat
= GEN_FCN (icode
) (target
, op
[0]);
9241 pat
= GEN_FCN (icode
) (op
[0]);
9244 pat
= GEN_FCN (icode
) (target
, op
[0], op
[1]);
9260 /* Output assembly code for the trampoline template to
9263 On S/390, we use gpr 1 internally in the trampoline code;
9264 gpr 0 is used to hold the static chain. */
9267 s390_asm_trampoline_template (FILE *file
)
9270 op
[0] = gen_rtx_REG (Pmode
, 0);
9271 op
[1] = gen_rtx_REG (Pmode
, 1);
9275 output_asm_insn ("basr\t%1,0", op
); /* 2 byte */
9276 output_asm_insn ("lmg\t%0,%1,14(%1)", op
); /* 6 byte */
9277 output_asm_insn ("br\t%1", op
); /* 2 byte */
9278 ASM_OUTPUT_SKIP (file
, (HOST_WIDE_INT
)(TRAMPOLINE_SIZE
- 10));
9282 output_asm_insn ("basr\t%1,0", op
); /* 2 byte */
9283 output_asm_insn ("lm\t%0,%1,6(%1)", op
); /* 4 byte */
9284 output_asm_insn ("br\t%1", op
); /* 2 byte */
9285 ASM_OUTPUT_SKIP (file
, (HOST_WIDE_INT
)(TRAMPOLINE_SIZE
- 8));
9289 /* Emit RTL insns to initialize the variable parts of a trampoline.
9290 FNADDR is an RTX for the address of the function's pure code.
9291 CXT is an RTX for the static chain value for the function. */
9294 s390_trampoline_init (rtx m_tramp
, tree fndecl
, rtx cxt
)
9296 rtx fnaddr
= XEXP (DECL_RTL (fndecl
), 0);
9299 emit_block_move (m_tramp
, assemble_trampoline_template (),
9300 GEN_INT (2 * UNITS_PER_LONG
), BLOCK_OP_NORMAL
);
9302 mem
= adjust_address (m_tramp
, Pmode
, 2 * UNITS_PER_LONG
);
9303 emit_move_insn (mem
, cxt
);
9304 mem
= adjust_address (m_tramp
, Pmode
, 3 * UNITS_PER_LONG
);
9305 emit_move_insn (mem
, fnaddr
);
9308 /* Output assembler code to FILE to increment profiler label # LABELNO
9309 for profiling a function entry. */
9312 s390_function_profiler (FILE *file
, int labelno
)
9317 ASM_GENERATE_INTERNAL_LABEL (label
, "LP", labelno
);
9319 fprintf (file
, "# function profiler \n");
9321 op
[0] = gen_rtx_REG (Pmode
, RETURN_REGNUM
);
9322 op
[1] = gen_rtx_REG (Pmode
, STACK_POINTER_REGNUM
);
9323 op
[1] = gen_rtx_MEM (Pmode
, plus_constant (Pmode
, op
[1], UNITS_PER_LONG
));
9325 op
[2] = gen_rtx_REG (Pmode
, 1);
9326 op
[3] = gen_rtx_SYMBOL_REF (Pmode
, label
);
9327 SYMBOL_REF_FLAGS (op
[3]) = SYMBOL_FLAG_LOCAL
;
9329 op
[4] = gen_rtx_SYMBOL_REF (Pmode
, "_mcount");
9332 op
[4] = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op
[4]), UNSPEC_PLT
);
9333 op
[4] = gen_rtx_CONST (Pmode
, op
[4]);
9338 output_asm_insn ("stg\t%0,%1", op
);
9339 output_asm_insn ("larl\t%2,%3", op
);
9340 output_asm_insn ("brasl\t%0,%4", op
);
9341 output_asm_insn ("lg\t%0,%1", op
);
9345 op
[6] = gen_label_rtx ();
9347 output_asm_insn ("st\t%0,%1", op
);
9348 output_asm_insn ("bras\t%2,%l6", op
);
9349 output_asm_insn (".long\t%4", op
);
9350 output_asm_insn (".long\t%3", op
);
9351 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[6]));
9352 output_asm_insn ("l\t%0,0(%2)", op
);
9353 output_asm_insn ("l\t%2,4(%2)", op
);
9354 output_asm_insn ("basr\t%0,%0", op
);
9355 output_asm_insn ("l\t%0,%1", op
);
9359 op
[5] = gen_label_rtx ();
9360 op
[6] = gen_label_rtx ();
9362 output_asm_insn ("st\t%0,%1", op
);
9363 output_asm_insn ("bras\t%2,%l6", op
);
9364 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[5]));
9365 output_asm_insn (".long\t%4-%l5", op
);
9366 output_asm_insn (".long\t%3-%l5", op
);
9367 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[6]));
9368 output_asm_insn ("lr\t%0,%2", op
);
9369 output_asm_insn ("a\t%0,0(%2)", op
);
9370 output_asm_insn ("a\t%2,4(%2)", op
);
9371 output_asm_insn ("basr\t%0,%0", op
);
9372 output_asm_insn ("l\t%0,%1", op
);
9376 /* Encode symbol attributes (local vs. global, tls model) of a SYMBOL_REF
9377 into its SYMBOL_REF_FLAGS. */
9380 s390_encode_section_info (tree decl
, rtx rtl
, int first
)
9382 default_encode_section_info (decl
, rtl
, first
);
9384 if (TREE_CODE (decl
) == VAR_DECL
)
9386 /* If a variable has a forced alignment to < 2 bytes, mark it
9387 with SYMBOL_FLAG_ALIGN1 to prevent it from being used as LARL
9389 if (DECL_USER_ALIGN (decl
) && DECL_ALIGN (decl
) < 16)
9390 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_ALIGN1
;
9391 if (!DECL_SIZE (decl
)
9392 || !DECL_ALIGN (decl
)
9393 || !host_integerp (DECL_SIZE (decl
), 0)
9394 || (DECL_ALIGN (decl
) <= 64
9395 && DECL_ALIGN (decl
) != tree_low_cst (DECL_SIZE (decl
), 0)))
9396 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_NOT_NATURALLY_ALIGNED
;
9399 /* Literal pool references don't have a decl so they are handled
9400 differently here. We rely on the information in the MEM_ALIGN
9401 entry to decide upon natural alignment. */
9403 && GET_CODE (XEXP (rtl
, 0)) == SYMBOL_REF
9404 && TREE_CONSTANT_POOL_ADDRESS_P (XEXP (rtl
, 0))
9405 && (MEM_ALIGN (rtl
) == 0
9406 || GET_MODE_BITSIZE (GET_MODE (rtl
)) == 0
9407 || MEM_ALIGN (rtl
) < GET_MODE_BITSIZE (GET_MODE (rtl
))))
9408 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_NOT_NATURALLY_ALIGNED
;
9411 /* Output thunk to FILE that implements a C++ virtual function call (with
9412 multiple inheritance) to FUNCTION. The thunk adjusts the this pointer
9413 by DELTA, and unless VCALL_OFFSET is zero, applies an additional adjustment
9414 stored at VCALL_OFFSET in the vtable whose address is located at offset 0
9415 relative to the resulting this pointer. */
9418 s390_output_mi_thunk (FILE *file
, tree thunk ATTRIBUTE_UNUSED
,
9419 HOST_WIDE_INT delta
, HOST_WIDE_INT vcall_offset
,
9425 /* Make sure unwind info is emitted for the thunk if needed. */
9426 final_start_function (emit_barrier (), file
, 1);
9428 /* Operand 0 is the target function. */
9429 op
[0] = XEXP (DECL_RTL (function
), 0);
9430 if (flag_pic
&& !SYMBOL_REF_LOCAL_P (op
[0]))
9433 op
[0] = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op
[0]),
9434 TARGET_64BIT
? UNSPEC_PLT
: UNSPEC_GOT
);
9435 op
[0] = gen_rtx_CONST (Pmode
, op
[0]);
9438 /* Operand 1 is the 'this' pointer. */
9439 if (aggregate_value_p (TREE_TYPE (TREE_TYPE (function
)), function
))
9440 op
[1] = gen_rtx_REG (Pmode
, 3);
9442 op
[1] = gen_rtx_REG (Pmode
, 2);
9444 /* Operand 2 is the delta. */
9445 op
[2] = GEN_INT (delta
);
9447 /* Operand 3 is the vcall_offset. */
9448 op
[3] = GEN_INT (vcall_offset
);
9450 /* Operand 4 is the temporary register. */
9451 op
[4] = gen_rtx_REG (Pmode
, 1);
9453 /* Operands 5 to 8 can be used as labels. */
9459 /* Operand 9 can be used for temporary register. */
9462 /* Generate code. */
9465 /* Setup literal pool pointer if required. */
9466 if ((!DISP_IN_RANGE (delta
)
9467 && !CONST_OK_FOR_K (delta
)
9468 && !CONST_OK_FOR_Os (delta
))
9469 || (!DISP_IN_RANGE (vcall_offset
)
9470 && !CONST_OK_FOR_K (vcall_offset
)
9471 && !CONST_OK_FOR_Os (vcall_offset
)))
9473 op
[5] = gen_label_rtx ();
9474 output_asm_insn ("larl\t%4,%5", op
);
9477 /* Add DELTA to this pointer. */
9480 if (CONST_OK_FOR_J (delta
))
9481 output_asm_insn ("la\t%1,%2(%1)", op
);
9482 else if (DISP_IN_RANGE (delta
))
9483 output_asm_insn ("lay\t%1,%2(%1)", op
);
9484 else if (CONST_OK_FOR_K (delta
))
9485 output_asm_insn ("aghi\t%1,%2", op
);
9486 else if (CONST_OK_FOR_Os (delta
))
9487 output_asm_insn ("agfi\t%1,%2", op
);
9490 op
[6] = gen_label_rtx ();
9491 output_asm_insn ("agf\t%1,%6-%5(%4)", op
);
9495 /* Perform vcall adjustment. */
9498 if (DISP_IN_RANGE (vcall_offset
))
9500 output_asm_insn ("lg\t%4,0(%1)", op
);
9501 output_asm_insn ("ag\t%1,%3(%4)", op
);
9503 else if (CONST_OK_FOR_K (vcall_offset
))
9505 output_asm_insn ("lghi\t%4,%3", op
);
9506 output_asm_insn ("ag\t%4,0(%1)", op
);
9507 output_asm_insn ("ag\t%1,0(%4)", op
);
9509 else if (CONST_OK_FOR_Os (vcall_offset
))
9511 output_asm_insn ("lgfi\t%4,%3", op
);
9512 output_asm_insn ("ag\t%4,0(%1)", op
);
9513 output_asm_insn ("ag\t%1,0(%4)", op
);
9517 op
[7] = gen_label_rtx ();
9518 output_asm_insn ("llgf\t%4,%7-%5(%4)", op
);
9519 output_asm_insn ("ag\t%4,0(%1)", op
);
9520 output_asm_insn ("ag\t%1,0(%4)", op
);
9524 /* Jump to target. */
9525 output_asm_insn ("jg\t%0", op
);
9527 /* Output literal pool if required. */
9530 output_asm_insn (".align\t4", op
);
9531 targetm
.asm_out
.internal_label (file
, "L",
9532 CODE_LABEL_NUMBER (op
[5]));
9536 targetm
.asm_out
.internal_label (file
, "L",
9537 CODE_LABEL_NUMBER (op
[6]));
9538 output_asm_insn (".long\t%2", op
);
9542 targetm
.asm_out
.internal_label (file
, "L",
9543 CODE_LABEL_NUMBER (op
[7]));
9544 output_asm_insn (".long\t%3", op
);
9549 /* Setup base pointer if required. */
9551 || (!DISP_IN_RANGE (delta
)
9552 && !CONST_OK_FOR_K (delta
)
9553 && !CONST_OK_FOR_Os (delta
))
9554 || (!DISP_IN_RANGE (delta
)
9555 && !CONST_OK_FOR_K (vcall_offset
)
9556 && !CONST_OK_FOR_Os (vcall_offset
)))
9558 op
[5] = gen_label_rtx ();
9559 output_asm_insn ("basr\t%4,0", op
);
9560 targetm
.asm_out
.internal_label (file
, "L",
9561 CODE_LABEL_NUMBER (op
[5]));
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 ("ahi\t%1,%2", op
);
9573 else if (CONST_OK_FOR_Os (delta
))
9574 output_asm_insn ("afi\t%1,%2", op
);
9577 op
[6] = gen_label_rtx ();
9578 output_asm_insn ("a\t%1,%6-%5(%4)", op
);
9582 /* Perform vcall adjustment. */
9585 if (CONST_OK_FOR_J (vcall_offset
))
9587 output_asm_insn ("l\t%4,0(%1)", op
);
9588 output_asm_insn ("a\t%1,%3(%4)", op
);
9590 else if (DISP_IN_RANGE (vcall_offset
))
9592 output_asm_insn ("l\t%4,0(%1)", op
);
9593 output_asm_insn ("ay\t%1,%3(%4)", op
);
9595 else if (CONST_OK_FOR_K (vcall_offset
))
9597 output_asm_insn ("lhi\t%4,%3", op
);
9598 output_asm_insn ("a\t%4,0(%1)", op
);
9599 output_asm_insn ("a\t%1,0(%4)", op
);
9601 else if (CONST_OK_FOR_Os (vcall_offset
))
9603 output_asm_insn ("iilf\t%4,%3", op
);
9604 output_asm_insn ("a\t%4,0(%1)", op
);
9605 output_asm_insn ("a\t%1,0(%4)", op
);
9609 op
[7] = gen_label_rtx ();
9610 output_asm_insn ("l\t%4,%7-%5(%4)", op
);
9611 output_asm_insn ("a\t%4,0(%1)", op
);
9612 output_asm_insn ("a\t%1,0(%4)", op
);
9615 /* We had to clobber the base pointer register.
9616 Re-setup the base pointer (with a different base). */
9617 op
[5] = gen_label_rtx ();
9618 output_asm_insn ("basr\t%4,0", op
);
9619 targetm
.asm_out
.internal_label (file
, "L",
9620 CODE_LABEL_NUMBER (op
[5]));
9623 /* Jump to target. */
9624 op
[8] = gen_label_rtx ();
9627 output_asm_insn ("l\t%4,%8-%5(%4)", op
);
9629 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9630 /* We cannot call through .plt, since .plt requires %r12 loaded. */
9631 else if (flag_pic
== 1)
9633 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9634 output_asm_insn ("l\t%4,%0(%4)", op
);
9636 else if (flag_pic
== 2)
9638 op
[9] = gen_rtx_REG (Pmode
, 0);
9639 output_asm_insn ("l\t%9,%8-4-%5(%4)", op
);
9640 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9641 output_asm_insn ("ar\t%4,%9", op
);
9642 output_asm_insn ("l\t%4,0(%4)", op
);
9645 output_asm_insn ("br\t%4", op
);
9647 /* Output literal pool. */
9648 output_asm_insn (".align\t4", op
);
9650 if (nonlocal
&& flag_pic
== 2)
9651 output_asm_insn (".long\t%0", op
);
9654 op
[0] = gen_rtx_SYMBOL_REF (Pmode
, "_GLOBAL_OFFSET_TABLE_");
9655 SYMBOL_REF_FLAGS (op
[0]) = SYMBOL_FLAG_LOCAL
;
9658 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[8]));
9660 output_asm_insn (".long\t%0", op
);
9662 output_asm_insn (".long\t%0-%5", op
);
9666 targetm
.asm_out
.internal_label (file
, "L",
9667 CODE_LABEL_NUMBER (op
[6]));
9668 output_asm_insn (".long\t%2", op
);
9672 targetm
.asm_out
.internal_label (file
, "L",
9673 CODE_LABEL_NUMBER (op
[7]));
9674 output_asm_insn (".long\t%3", op
);
9677 final_end_function ();
9681 s390_valid_pointer_mode (enum machine_mode mode
)
9683 return (mode
== SImode
|| (TARGET_64BIT
&& mode
== DImode
));
9686 /* Checks whether the given CALL_EXPR would use a caller
9687 saved register. This is used to decide whether sibling call
9688 optimization could be performed on the respective function
9692 s390_call_saved_register_used (tree call_expr
)
9694 CUMULATIVE_ARGS cum_v
;
9695 cumulative_args_t cum
;
9697 enum machine_mode mode
;
9702 INIT_CUMULATIVE_ARGS (cum_v
, NULL
, NULL
, 0, 0);
9703 cum
= pack_cumulative_args (&cum_v
);
9705 for (i
= 0; i
< call_expr_nargs (call_expr
); i
++)
9707 parameter
= CALL_EXPR_ARG (call_expr
, i
);
9708 gcc_assert (parameter
);
9710 /* For an undeclared variable passed as parameter we will get
9711 an ERROR_MARK node here. */
9712 if (TREE_CODE (parameter
) == ERROR_MARK
)
9715 type
= TREE_TYPE (parameter
);
9718 mode
= TYPE_MODE (type
);
9721 if (pass_by_reference (&cum_v
, mode
, type
, true))
9724 type
= build_pointer_type (type
);
9727 parm_rtx
= s390_function_arg (cum
, mode
, type
, 0);
9729 s390_function_arg_advance (cum
, mode
, type
, 0);
9734 if (REG_P (parm_rtx
))
9737 reg
< HARD_REGNO_NREGS (REGNO (parm_rtx
), GET_MODE (parm_rtx
));
9739 if (!call_used_regs
[reg
+ REGNO (parm_rtx
)])
9743 if (GET_CODE (parm_rtx
) == PARALLEL
)
9747 for (i
= 0; i
< XVECLEN (parm_rtx
, 0); i
++)
9749 rtx r
= XEXP (XVECEXP (parm_rtx
, 0, i
), 0);
9751 gcc_assert (REG_P (r
));
9754 reg
< HARD_REGNO_NREGS (REGNO (r
), GET_MODE (r
));
9756 if (!call_used_regs
[reg
+ REGNO (r
)])
9765 /* Return true if the given call expression can be
9766 turned into a sibling call.
9767 DECL holds the declaration of the function to be called whereas
9768 EXP is the call expression itself. */
9771 s390_function_ok_for_sibcall (tree decl
, tree exp
)
9773 /* The TPF epilogue uses register 1. */
9774 if (TARGET_TPF_PROFILING
)
9777 /* The 31 bit PLT code uses register 12 (GOT pointer - caller saved)
9778 which would have to be restored before the sibcall. */
9779 if (!TARGET_64BIT
&& flag_pic
&& decl
&& !targetm
.binds_local_p (decl
))
9782 /* Register 6 on s390 is available as an argument register but unfortunately
9783 "caller saved". This makes functions needing this register for arguments
9784 not suitable for sibcalls. */
9785 return !s390_call_saved_register_used (exp
);
9788 /* Return the fixed registers used for condition codes. */
9791 s390_fixed_condition_code_regs (unsigned int *p1
, unsigned int *p2
)
9794 *p2
= INVALID_REGNUM
;
9799 /* This function is used by the call expanders of the machine description.
9800 It emits the call insn itself together with the necessary operations
9801 to adjust the target address and returns the emitted insn.
9802 ADDR_LOCATION is the target address rtx
9803 TLS_CALL the location of the thread-local symbol
9804 RESULT_REG the register where the result of the call should be stored
9805 RETADDR_REG the register where the return address should be stored
9806 If this parameter is NULL_RTX the call is considered
9807 to be a sibling call. */
9810 s390_emit_call (rtx addr_location
, rtx tls_call
, rtx result_reg
,
9813 bool plt_call
= false;
9819 /* Direct function calls need special treatment. */
9820 if (GET_CODE (addr_location
) == SYMBOL_REF
)
9822 /* When calling a global routine in PIC mode, we must
9823 replace the symbol itself with the PLT stub. */
9824 if (flag_pic
&& !SYMBOL_REF_LOCAL_P (addr_location
))
9826 if (retaddr_reg
!= NULL_RTX
)
9828 addr_location
= gen_rtx_UNSPEC (Pmode
,
9829 gen_rtvec (1, addr_location
),
9831 addr_location
= gen_rtx_CONST (Pmode
, addr_location
);
9835 /* For -fpic code the PLT entries might use r12 which is
9836 call-saved. Therefore we cannot do a sibcall when
9837 calling directly using a symbol ref. When reaching
9838 this point we decided (in s390_function_ok_for_sibcall)
9839 to do a sibcall for a function pointer but one of the
9840 optimizers was able to get rid of the function pointer
9841 by propagating the symbol ref into the call. This
9842 optimization is illegal for S/390 so we turn the direct
9843 call into a indirect call again. */
9844 addr_location
= force_reg (Pmode
, addr_location
);
9847 /* Unless we can use the bras(l) insn, force the
9848 routine address into a register. */
9849 if (!TARGET_SMALL_EXEC
&& !TARGET_CPU_ZARCH
)
9852 addr_location
= legitimize_pic_address (addr_location
, 0);
9854 addr_location
= force_reg (Pmode
, addr_location
);
9858 /* If it is already an indirect call or the code above moved the
9859 SYMBOL_REF to somewhere else make sure the address can be found in
9861 if (retaddr_reg
== NULL_RTX
9862 && GET_CODE (addr_location
) != SYMBOL_REF
9865 emit_move_insn (gen_rtx_REG (Pmode
, SIBCALL_REGNUM
), addr_location
);
9866 addr_location
= gen_rtx_REG (Pmode
, SIBCALL_REGNUM
);
9869 addr_location
= gen_rtx_MEM (QImode
, addr_location
);
9870 call
= gen_rtx_CALL (VOIDmode
, addr_location
, const0_rtx
);
9872 if (result_reg
!= NULL_RTX
)
9873 call
= gen_rtx_SET (VOIDmode
, result_reg
, call
);
9875 if (retaddr_reg
!= NULL_RTX
)
9877 clobber
= gen_rtx_CLOBBER (VOIDmode
, retaddr_reg
);
9879 if (tls_call
!= NULL_RTX
)
9880 vec
= gen_rtvec (3, call
, clobber
,
9881 gen_rtx_USE (VOIDmode
, tls_call
));
9883 vec
= gen_rtvec (2, call
, clobber
);
9885 call
= gen_rtx_PARALLEL (VOIDmode
, vec
);
9888 insn
= emit_call_insn (call
);
9890 /* 31-bit PLT stubs and tls calls use the GOT register implicitly. */
9891 if ((!TARGET_64BIT
&& plt_call
) || tls_call
!= NULL_RTX
)
9893 /* s390_function_ok_for_sibcall should
9894 have denied sibcalls in this case. */
9895 gcc_assert (retaddr_reg
!= NULL_RTX
);
9896 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), gen_rtx_REG (Pmode
, 12));
9901 /* Implement TARGET_CONDITIONAL_REGISTER_USAGE. */
9904 s390_conditional_register_usage (void)
9910 fixed_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
9911 call_used_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
9913 if (TARGET_CPU_ZARCH
)
9915 fixed_regs
[BASE_REGNUM
] = 0;
9916 call_used_regs
[BASE_REGNUM
] = 0;
9917 fixed_regs
[RETURN_REGNUM
] = 0;
9918 call_used_regs
[RETURN_REGNUM
] = 0;
9922 for (i
= 24; i
< 32; i
++)
9923 call_used_regs
[i
] = call_really_used_regs
[i
] = 0;
9927 for (i
= 18; i
< 20; i
++)
9928 call_used_regs
[i
] = call_really_used_regs
[i
] = 0;
9931 if (TARGET_SOFT_FLOAT
)
9933 for (i
= 16; i
< 32; i
++)
9934 call_used_regs
[i
] = fixed_regs
[i
] = 1;
9938 /* Corresponding function to eh_return expander. */
9940 static GTY(()) rtx s390_tpf_eh_return_symbol
;
9942 s390_emit_tpf_eh_return (rtx target
)
9946 if (!s390_tpf_eh_return_symbol
)
9947 s390_tpf_eh_return_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "__tpf_eh_return");
9949 reg
= gen_rtx_REG (Pmode
, 2);
9951 emit_move_insn (reg
, target
);
9952 insn
= s390_emit_call (s390_tpf_eh_return_symbol
, NULL_RTX
, reg
,
9953 gen_rtx_REG (Pmode
, RETURN_REGNUM
));
9954 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), reg
);
9956 emit_move_insn (EH_RETURN_HANDLER_RTX
, reg
);
9959 /* Rework the prologue/epilogue to avoid saving/restoring
9960 registers unnecessarily. */
9963 s390_optimize_prologue (void)
9965 rtx insn
, new_insn
, next_insn
;
9967 /* Do a final recompute of the frame-related data. */
9969 s390_update_frame_layout ();
9971 /* If all special registers are in fact used, there's nothing we
9972 can do, so no point in walking the insn list. */
9974 if (cfun_frame_layout
.first_save_gpr
<= BASE_REGNUM
9975 && cfun_frame_layout
.last_save_gpr
>= BASE_REGNUM
9976 && (TARGET_CPU_ZARCH
9977 || (cfun_frame_layout
.first_save_gpr
<= RETURN_REGNUM
9978 && cfun_frame_layout
.last_save_gpr
>= RETURN_REGNUM
)))
9981 /* Search for prologue/epilogue insns and replace them. */
9983 for (insn
= get_insns (); insn
; insn
= next_insn
)
9985 int first
, last
, off
;
9986 rtx set
, base
, offset
;
9988 next_insn
= NEXT_INSN (insn
);
9990 if (GET_CODE (insn
) != INSN
)
9993 if (GET_CODE (PATTERN (insn
)) == PARALLEL
9994 && store_multiple_operation (PATTERN (insn
), VOIDmode
))
9996 set
= XVECEXP (PATTERN (insn
), 0, 0);
9997 first
= REGNO (SET_SRC (set
));
9998 last
= first
+ XVECLEN (PATTERN (insn
), 0) - 1;
9999 offset
= const0_rtx
;
10000 base
= eliminate_constant_term (XEXP (SET_DEST (set
), 0), &offset
);
10001 off
= INTVAL (offset
);
10003 if (GET_CODE (base
) != REG
|| off
< 0)
10005 if (cfun_frame_layout
.first_save_gpr
!= -1
10006 && (cfun_frame_layout
.first_save_gpr
< first
10007 || cfun_frame_layout
.last_save_gpr
> last
))
10009 if (REGNO (base
) != STACK_POINTER_REGNUM
10010 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10012 if (first
> BASE_REGNUM
|| last
< BASE_REGNUM
)
10015 if (cfun_frame_layout
.first_save_gpr
!= -1)
10017 new_insn
= save_gprs (base
,
10018 off
+ (cfun_frame_layout
.first_save_gpr
10019 - first
) * UNITS_PER_LONG
,
10020 cfun_frame_layout
.first_save_gpr
,
10021 cfun_frame_layout
.last_save_gpr
);
10022 new_insn
= emit_insn_before (new_insn
, insn
);
10023 INSN_ADDRESSES_NEW (new_insn
, -1);
10026 remove_insn (insn
);
10030 if (cfun_frame_layout
.first_save_gpr
== -1
10031 && GET_CODE (PATTERN (insn
)) == SET
10032 && GET_CODE (SET_SRC (PATTERN (insn
))) == REG
10033 && (REGNO (SET_SRC (PATTERN (insn
))) == BASE_REGNUM
10034 || (!TARGET_CPU_ZARCH
10035 && REGNO (SET_SRC (PATTERN (insn
))) == RETURN_REGNUM
))
10036 && GET_CODE (SET_DEST (PATTERN (insn
))) == MEM
)
10038 set
= PATTERN (insn
);
10039 first
= REGNO (SET_SRC (set
));
10040 offset
= const0_rtx
;
10041 base
= eliminate_constant_term (XEXP (SET_DEST (set
), 0), &offset
);
10042 off
= INTVAL (offset
);
10044 if (GET_CODE (base
) != REG
|| off
< 0)
10046 if (REGNO (base
) != STACK_POINTER_REGNUM
10047 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10050 remove_insn (insn
);
10054 if (GET_CODE (PATTERN (insn
)) == PARALLEL
10055 && load_multiple_operation (PATTERN (insn
), VOIDmode
))
10057 set
= XVECEXP (PATTERN (insn
), 0, 0);
10058 first
= REGNO (SET_DEST (set
));
10059 last
= first
+ XVECLEN (PATTERN (insn
), 0) - 1;
10060 offset
= const0_rtx
;
10061 base
= eliminate_constant_term (XEXP (SET_SRC (set
), 0), &offset
);
10062 off
= INTVAL (offset
);
10064 if (GET_CODE (base
) != REG
|| off
< 0)
10066 if (cfun_frame_layout
.first_restore_gpr
!= -1
10067 && (cfun_frame_layout
.first_restore_gpr
< first
10068 || cfun_frame_layout
.last_restore_gpr
> last
))
10070 if (REGNO (base
) != STACK_POINTER_REGNUM
10071 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10073 if (first
> BASE_REGNUM
|| last
< BASE_REGNUM
)
10076 if (cfun_frame_layout
.first_restore_gpr
!= -1)
10078 new_insn
= restore_gprs (base
,
10079 off
+ (cfun_frame_layout
.first_restore_gpr
10080 - first
) * UNITS_PER_LONG
,
10081 cfun_frame_layout
.first_restore_gpr
,
10082 cfun_frame_layout
.last_restore_gpr
);
10083 new_insn
= emit_insn_before (new_insn
, insn
);
10084 INSN_ADDRESSES_NEW (new_insn
, -1);
10087 remove_insn (insn
);
10091 if (cfun_frame_layout
.first_restore_gpr
== -1
10092 && GET_CODE (PATTERN (insn
)) == SET
10093 && GET_CODE (SET_DEST (PATTERN (insn
))) == REG
10094 && (REGNO (SET_DEST (PATTERN (insn
))) == BASE_REGNUM
10095 || (!TARGET_CPU_ZARCH
10096 && REGNO (SET_DEST (PATTERN (insn
))) == RETURN_REGNUM
))
10097 && GET_CODE (SET_SRC (PATTERN (insn
))) == MEM
)
10099 set
= PATTERN (insn
);
10100 first
= REGNO (SET_DEST (set
));
10101 offset
= const0_rtx
;
10102 base
= eliminate_constant_term (XEXP (SET_SRC (set
), 0), &offset
);
10103 off
= INTVAL (offset
);
10105 if (GET_CODE (base
) != REG
|| off
< 0)
10107 if (REGNO (base
) != STACK_POINTER_REGNUM
10108 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10111 remove_insn (insn
);
10117 /* On z10 and later the dynamic branch prediction must see the
10118 backward jump within a certain windows. If not it falls back to
10119 the static prediction. This function rearranges the loop backward
10120 branch in a way which makes the static prediction always correct.
10121 The function returns true if it added an instruction. */
10123 s390_fix_long_loop_prediction (rtx insn
)
10125 rtx set
= single_set (insn
);
10126 rtx code_label
, label_ref
, new_label
;
10132 /* This will exclude branch on count and branch on index patterns
10133 since these are correctly statically predicted. */
10135 || SET_DEST (set
) != pc_rtx
10136 || GET_CODE (SET_SRC(set
)) != IF_THEN_ELSE
)
10139 label_ref
= (GET_CODE (XEXP (SET_SRC (set
), 1)) == LABEL_REF
?
10140 XEXP (SET_SRC (set
), 1) : XEXP (SET_SRC (set
), 2));
10142 gcc_assert (GET_CODE (label_ref
) == LABEL_REF
);
10144 code_label
= XEXP (label_ref
, 0);
10146 if (INSN_ADDRESSES (INSN_UID (code_label
)) == -1
10147 || INSN_ADDRESSES (INSN_UID (insn
)) == -1
10148 || (INSN_ADDRESSES (INSN_UID (insn
))
10149 - INSN_ADDRESSES (INSN_UID (code_label
)) < PREDICT_DISTANCE
))
10152 for (distance
= 0, cur_insn
= PREV_INSN (insn
);
10153 distance
< PREDICT_DISTANCE
- 6;
10154 distance
+= get_attr_length (cur_insn
), cur_insn
= PREV_INSN (cur_insn
))
10155 if (!cur_insn
|| JUMP_P (cur_insn
) || LABEL_P (cur_insn
))
10158 new_label
= gen_label_rtx ();
10159 uncond_jump
= emit_jump_insn_after (
10160 gen_rtx_SET (VOIDmode
, pc_rtx
,
10161 gen_rtx_LABEL_REF (VOIDmode
, code_label
)),
10163 emit_label_after (new_label
, uncond_jump
);
10165 tmp
= XEXP (SET_SRC (set
), 1);
10166 XEXP (SET_SRC (set
), 1) = XEXP (SET_SRC (set
), 2);
10167 XEXP (SET_SRC (set
), 2) = tmp
;
10168 INSN_CODE (insn
) = -1;
10170 XEXP (label_ref
, 0) = new_label
;
10171 JUMP_LABEL (insn
) = new_label
;
10172 JUMP_LABEL (uncond_jump
) = code_label
;
10177 /* Returns 1 if INSN reads the value of REG for purposes not related
10178 to addressing of memory, and 0 otherwise. */
10180 s390_non_addr_reg_read_p (rtx reg
, rtx insn
)
10182 return reg_referenced_p (reg
, PATTERN (insn
))
10183 && !reg_used_in_mem_p (REGNO (reg
), PATTERN (insn
));
10186 /* Starting from INSN find_cond_jump looks downwards in the insn
10187 stream for a single jump insn which is the last user of the
10188 condition code set in INSN. */
10190 find_cond_jump (rtx insn
)
10192 for (; insn
; insn
= NEXT_INSN (insn
))
10196 if (LABEL_P (insn
))
10199 if (!JUMP_P (insn
))
10201 if (reg_mentioned_p (gen_rtx_REG (CCmode
, CC_REGNUM
), insn
))
10206 /* This will be triggered by a return. */
10207 if (GET_CODE (PATTERN (insn
)) != SET
)
10210 gcc_assert (SET_DEST (PATTERN (insn
)) == pc_rtx
);
10211 ite
= SET_SRC (PATTERN (insn
));
10213 if (GET_CODE (ite
) != IF_THEN_ELSE
)
10216 cc
= XEXP (XEXP (ite
, 0), 0);
10217 if (!REG_P (cc
) || !CC_REGNO_P (REGNO (cc
)))
10220 if (find_reg_note (insn
, REG_DEAD
, cc
))
10228 /* Swap the condition in COND and the operands in OP0 and OP1 so that
10229 the semantics does not change. If NULL_RTX is passed as COND the
10230 function tries to find the conditional jump starting with INSN. */
10232 s390_swap_cmp (rtx cond
, rtx
*op0
, rtx
*op1
, rtx insn
)
10236 if (cond
== NULL_RTX
)
10238 rtx jump
= find_cond_jump (NEXT_INSN (insn
));
10239 jump
= jump
? single_set (jump
) : NULL_RTX
;
10241 if (jump
== NULL_RTX
)
10244 cond
= XEXP (XEXP (jump
, 1), 0);
10249 PUT_CODE (cond
, swap_condition (GET_CODE (cond
)));
10252 /* On z10, instructions of the compare-and-branch family have the
10253 property to access the register occurring as second operand with
10254 its bits complemented. If such a compare is grouped with a second
10255 instruction that accesses the same register non-complemented, and
10256 if that register's value is delivered via a bypass, then the
10257 pipeline recycles, thereby causing significant performance decline.
10258 This function locates such situations and exchanges the two
10259 operands of the compare. The function return true whenever it
10262 s390_z10_optimize_cmp (rtx insn
)
10264 rtx prev_insn
, next_insn
;
10265 bool insn_added_p
= false;
10266 rtx cond
, *op0
, *op1
;
10268 if (GET_CODE (PATTERN (insn
)) == PARALLEL
)
10270 /* Handle compare and branch and branch on count
10272 rtx pattern
= single_set (insn
);
10275 || SET_DEST (pattern
) != pc_rtx
10276 || GET_CODE (SET_SRC (pattern
)) != IF_THEN_ELSE
)
10279 cond
= XEXP (SET_SRC (pattern
), 0);
10280 op0
= &XEXP (cond
, 0);
10281 op1
= &XEXP (cond
, 1);
10283 else if (GET_CODE (PATTERN (insn
)) == SET
)
10287 /* Handle normal compare instructions. */
10288 src
= SET_SRC (PATTERN (insn
));
10289 dest
= SET_DEST (PATTERN (insn
));
10292 || !CC_REGNO_P (REGNO (dest
))
10293 || GET_CODE (src
) != COMPARE
)
10296 /* s390_swap_cmp will try to find the conditional
10297 jump when passing NULL_RTX as condition. */
10299 op0
= &XEXP (src
, 0);
10300 op1
= &XEXP (src
, 1);
10305 if (!REG_P (*op0
) || !REG_P (*op1
))
10308 if (GET_MODE_CLASS (GET_MODE (*op0
)) != MODE_INT
)
10311 /* Swap the COMPARE arguments and its mask if there is a
10312 conflicting access in the previous insn. */
10313 prev_insn
= prev_active_insn (insn
);
10314 if (prev_insn
!= NULL_RTX
&& INSN_P (prev_insn
)
10315 && reg_referenced_p (*op1
, PATTERN (prev_insn
)))
10316 s390_swap_cmp (cond
, op0
, op1
, insn
);
10318 /* Check if there is a conflict with the next insn. If there
10319 was no conflict with the previous insn, then swap the
10320 COMPARE arguments and its mask. If we already swapped
10321 the operands, or if swapping them would cause a conflict
10322 with the previous insn, issue a NOP after the COMPARE in
10323 order to separate the two instuctions. */
10324 next_insn
= next_active_insn (insn
);
10325 if (next_insn
!= NULL_RTX
&& INSN_P (next_insn
)
10326 && s390_non_addr_reg_read_p (*op1
, next_insn
))
10328 if (prev_insn
!= NULL_RTX
&& INSN_P (prev_insn
)
10329 && s390_non_addr_reg_read_p (*op0
, prev_insn
))
10331 if (REGNO (*op1
) == 0)
10332 emit_insn_after (gen_nop1 (), insn
);
10334 emit_insn_after (gen_nop (), insn
);
10335 insn_added_p
= true;
10338 s390_swap_cmp (cond
, op0
, op1
, insn
);
10340 return insn_added_p
;
10343 /* Perform machine-dependent processing. */
10348 bool pool_overflow
= false;
10350 /* Make sure all splits have been performed; splits after
10351 machine_dependent_reorg might confuse insn length counts. */
10352 split_all_insns_noflow ();
10354 /* Install the main literal pool and the associated base
10355 register load insns.
10357 In addition, there are two problematic situations we need
10360 - the literal pool might be > 4096 bytes in size, so that
10361 some of its elements cannot be directly accessed
10363 - a branch target might be > 64K away from the branch, so that
10364 it is not possible to use a PC-relative instruction.
10366 To fix those, we split the single literal pool into multiple
10367 pool chunks, reloading the pool base register at various
10368 points throughout the function to ensure it always points to
10369 the pool chunk the following code expects, and / or replace
10370 PC-relative branches by absolute branches.
10372 However, the two problems are interdependent: splitting the
10373 literal pool can move a branch further away from its target,
10374 causing the 64K limit to overflow, and on the other hand,
10375 replacing a PC-relative branch by an absolute branch means
10376 we need to put the branch target address into the literal
10377 pool, possibly causing it to overflow.
10379 So, we loop trying to fix up both problems until we manage
10380 to satisfy both conditions at the same time. Note that the
10381 loop is guaranteed to terminate as every pass of the loop
10382 strictly decreases the total number of PC-relative branches
10383 in the function. (This is not completely true as there
10384 might be branch-over-pool insns introduced by chunkify_start.
10385 Those never need to be split however.) */
10389 struct constant_pool
*pool
= NULL
;
10391 /* Collect the literal pool. */
10392 if (!pool_overflow
)
10394 pool
= s390_mainpool_start ();
10396 pool_overflow
= true;
10399 /* If literal pool overflowed, start to chunkify it. */
10401 pool
= s390_chunkify_start ();
10403 /* Split out-of-range branches. If this has created new
10404 literal pool entries, cancel current chunk list and
10405 recompute it. zSeries machines have large branch
10406 instructions, so we never need to split a branch. */
10407 if (!TARGET_CPU_ZARCH
&& s390_split_branches ())
10410 s390_chunkify_cancel (pool
);
10412 s390_mainpool_cancel (pool
);
10417 /* If we made it up to here, both conditions are satisfied.
10418 Finish up literal pool related changes. */
10420 s390_chunkify_finish (pool
);
10422 s390_mainpool_finish (pool
);
10424 /* We're done splitting branches. */
10425 cfun
->machine
->split_branches_pending_p
= false;
10429 /* Generate out-of-pool execute target insns. */
10430 if (TARGET_CPU_ZARCH
)
10432 rtx insn
, label
, target
;
10434 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
10436 label
= s390_execute_label (insn
);
10440 gcc_assert (label
!= const0_rtx
);
10442 target
= emit_label (XEXP (label
, 0));
10443 INSN_ADDRESSES_NEW (target
, -1);
10445 target
= emit_insn (s390_execute_target (insn
));
10446 INSN_ADDRESSES_NEW (target
, -1);
10450 /* Try to optimize prologue and epilogue further. */
10451 s390_optimize_prologue ();
10453 /* Walk over the insns and do some >=z10 specific changes. */
10454 if (s390_tune
== PROCESSOR_2097_Z10
10455 || s390_tune
== PROCESSOR_2817_Z196
)
10458 bool insn_added_p
= false;
10460 /* The insn lengths and addresses have to be up to date for the
10461 following manipulations. */
10462 shorten_branches (get_insns ());
10464 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
10466 if (!INSN_P (insn
) || INSN_CODE (insn
) <= 0)
10470 insn_added_p
|= s390_fix_long_loop_prediction (insn
);
10472 if ((GET_CODE (PATTERN (insn
)) == PARALLEL
10473 || GET_CODE (PATTERN (insn
)) == SET
)
10474 && s390_tune
== PROCESSOR_2097_Z10
)
10475 insn_added_p
|= s390_z10_optimize_cmp (insn
);
10478 /* Adjust branches if we added new instructions. */
10480 shorten_branches (get_insns ());
10484 /* Return true if INSN is a fp load insn writing register REGNO. */
10486 s390_fpload_toreg (rtx insn
, unsigned int regno
)
10489 enum attr_type flag
= s390_safe_attr_type (insn
);
10491 if (flag
!= TYPE_FLOADSF
&& flag
!= TYPE_FLOADDF
)
10494 set
= single_set (insn
);
10496 if (set
== NULL_RTX
)
10499 if (!REG_P (SET_DEST (set
)) || !MEM_P (SET_SRC (set
)))
10502 if (REGNO (SET_DEST (set
)) != regno
)
10508 /* This value describes the distance to be avoided between an
10509 aritmetic fp instruction and an fp load writing the same register.
10510 Z10_EARLYLOAD_DISTANCE - 1 as well as Z10_EARLYLOAD_DISTANCE + 1 is
10511 fine but the exact value has to be avoided. Otherwise the FP
10512 pipeline will throw an exception causing a major penalty. */
10513 #define Z10_EARLYLOAD_DISTANCE 7
10515 /* Rearrange the ready list in order to avoid the situation described
10516 for Z10_EARLYLOAD_DISTANCE. A problematic load instruction is
10517 moved to the very end of the ready list. */
10519 s390_z10_prevent_earlyload_conflicts (rtx
*ready
, int *nready_p
)
10521 unsigned int regno
;
10522 int nready
= *nready_p
;
10527 enum attr_type flag
;
10530 /* Skip DISTANCE - 1 active insns. */
10531 for (insn
= last_scheduled_insn
, distance
= Z10_EARLYLOAD_DISTANCE
- 1;
10532 distance
> 0 && insn
!= NULL_RTX
;
10533 distance
--, insn
= prev_active_insn (insn
))
10534 if (CALL_P (insn
) || JUMP_P (insn
))
10537 if (insn
== NULL_RTX
)
10540 set
= single_set (insn
);
10542 if (set
== NULL_RTX
|| !REG_P (SET_DEST (set
))
10543 || GET_MODE_CLASS (GET_MODE (SET_DEST (set
))) != MODE_FLOAT
)
10546 flag
= s390_safe_attr_type (insn
);
10548 if (flag
== TYPE_FLOADSF
|| flag
== TYPE_FLOADDF
)
10551 regno
= REGNO (SET_DEST (set
));
10554 while (!s390_fpload_toreg (ready
[i
], regno
) && i
> 0)
10561 memmove (&ready
[1], &ready
[0], sizeof (rtx
) * i
);
10565 /* This function is called via hook TARGET_SCHED_REORDER before
10566 issuing one insn from list READY which contains *NREADYP entries.
10567 For target z10 it reorders load instructions to avoid early load
10568 conflicts in the floating point pipeline */
10570 s390_sched_reorder (FILE *file ATTRIBUTE_UNUSED
, int verbose ATTRIBUTE_UNUSED
,
10571 rtx
*ready
, int *nreadyp
, int clock ATTRIBUTE_UNUSED
)
10573 if (s390_tune
== PROCESSOR_2097_Z10
)
10574 if (reload_completed
&& *nreadyp
> 1)
10575 s390_z10_prevent_earlyload_conflicts (ready
, nreadyp
);
10577 return s390_issue_rate ();
10580 /* This function is called via hook TARGET_SCHED_VARIABLE_ISSUE after
10581 the scheduler has issued INSN. It stores the last issued insn into
10582 last_scheduled_insn in order to make it available for
10583 s390_sched_reorder. */
10585 s390_sched_variable_issue (FILE *file ATTRIBUTE_UNUSED
,
10586 int verbose ATTRIBUTE_UNUSED
,
10587 rtx insn
, int more
)
10589 last_scheduled_insn
= insn
;
10591 if (GET_CODE (PATTERN (insn
)) != USE
10592 && GET_CODE (PATTERN (insn
)) != CLOBBER
)
10599 s390_sched_init (FILE *file ATTRIBUTE_UNUSED
,
10600 int verbose ATTRIBUTE_UNUSED
,
10601 int max_ready ATTRIBUTE_UNUSED
)
10603 last_scheduled_insn
= NULL_RTX
;
10606 /* This function checks the whole of insn X for memory references. The
10607 function always returns zero because the framework it is called
10608 from would stop recursively analyzing the insn upon a return value
10609 other than zero. The real result of this function is updating
10610 counter variable MEM_COUNT. */
10612 check_dpu (rtx
*x
, unsigned *mem_count
)
10614 if (*x
!= NULL_RTX
&& MEM_P (*x
))
10619 /* This target hook implementation for TARGET_LOOP_UNROLL_ADJUST calculates
10620 a new number struct loop *loop should be unrolled if tuned for cpus with
10621 a built-in stride prefetcher.
10622 The loop is analyzed for memory accesses by calling check_dpu for
10623 each rtx of the loop. Depending on the loop_depth and the amount of
10624 memory accesses a new number <=nunroll is returned to improve the
10625 behaviour of the hardware prefetch unit. */
10627 s390_loop_unroll_adjust (unsigned nunroll
, struct loop
*loop
)
10632 unsigned mem_count
= 0;
10634 if (s390_tune
!= PROCESSOR_2097_Z10
&& s390_tune
!= PROCESSOR_2817_Z196
)
10637 /* Count the number of memory references within the loop body. */
10638 bbs
= get_loop_body (loop
);
10639 for (i
= 0; i
< loop
->num_nodes
; i
++)
10641 for (insn
= BB_HEAD (bbs
[i
]); insn
!= BB_END (bbs
[i
]); insn
= NEXT_INSN (insn
))
10642 if (INSN_P (insn
) && INSN_CODE (insn
) != -1)
10643 for_each_rtx (&insn
, (rtx_function
) check_dpu
, &mem_count
);
10647 /* Prevent division by zero, and we do not need to adjust nunroll in this case. */
10648 if (mem_count
== 0)
10651 switch (loop_depth(loop
))
10654 return MIN (nunroll
, 28 / mem_count
);
10656 return MIN (nunroll
, 22 / mem_count
);
10658 return MIN (nunroll
, 16 / mem_count
);
10662 /* Initialize GCC target structure. */
10664 #undef TARGET_ASM_ALIGNED_HI_OP
10665 #define TARGET_ASM_ALIGNED_HI_OP "\t.word\t"
10666 #undef TARGET_ASM_ALIGNED_DI_OP
10667 #define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t"
10668 #undef TARGET_ASM_INTEGER
10669 #define TARGET_ASM_INTEGER s390_assemble_integer
10671 #undef TARGET_ASM_OPEN_PAREN
10672 #define TARGET_ASM_OPEN_PAREN ""
10674 #undef TARGET_ASM_CLOSE_PAREN
10675 #define TARGET_ASM_CLOSE_PAREN ""
10677 #undef TARGET_OPTION_OVERRIDE
10678 #define TARGET_OPTION_OVERRIDE s390_option_override
10680 #undef TARGET_ENCODE_SECTION_INFO
10681 #define TARGET_ENCODE_SECTION_INFO s390_encode_section_info
10683 #undef TARGET_SCALAR_MODE_SUPPORTED_P
10684 #define TARGET_SCALAR_MODE_SUPPORTED_P s390_scalar_mode_supported_p
10687 #undef TARGET_HAVE_TLS
10688 #define TARGET_HAVE_TLS true
10690 #undef TARGET_CANNOT_FORCE_CONST_MEM
10691 #define TARGET_CANNOT_FORCE_CONST_MEM s390_cannot_force_const_mem
10693 #undef TARGET_DELEGITIMIZE_ADDRESS
10694 #define TARGET_DELEGITIMIZE_ADDRESS s390_delegitimize_address
10696 #undef TARGET_LEGITIMIZE_ADDRESS
10697 #define TARGET_LEGITIMIZE_ADDRESS s390_legitimize_address
10699 #undef TARGET_RETURN_IN_MEMORY
10700 #define TARGET_RETURN_IN_MEMORY s390_return_in_memory
10702 #undef TARGET_INIT_BUILTINS
10703 #define TARGET_INIT_BUILTINS s390_init_builtins
10704 #undef TARGET_EXPAND_BUILTIN
10705 #define TARGET_EXPAND_BUILTIN s390_expand_builtin
10707 #undef TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA
10708 #define TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA s390_output_addr_const_extra
10710 #undef TARGET_ASM_OUTPUT_MI_THUNK
10711 #define TARGET_ASM_OUTPUT_MI_THUNK s390_output_mi_thunk
10712 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
10713 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true
10715 #undef TARGET_SCHED_ADJUST_PRIORITY
10716 #define TARGET_SCHED_ADJUST_PRIORITY s390_adjust_priority
10717 #undef TARGET_SCHED_ISSUE_RATE
10718 #define TARGET_SCHED_ISSUE_RATE s390_issue_rate
10719 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
10720 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD s390_first_cycle_multipass_dfa_lookahead
10722 #undef TARGET_SCHED_VARIABLE_ISSUE
10723 #define TARGET_SCHED_VARIABLE_ISSUE s390_sched_variable_issue
10724 #undef TARGET_SCHED_REORDER
10725 #define TARGET_SCHED_REORDER s390_sched_reorder
10726 #undef TARGET_SCHED_INIT
10727 #define TARGET_SCHED_INIT s390_sched_init
10729 #undef TARGET_CANNOT_COPY_INSN_P
10730 #define TARGET_CANNOT_COPY_INSN_P s390_cannot_copy_insn_p
10731 #undef TARGET_RTX_COSTS
10732 #define TARGET_RTX_COSTS s390_rtx_costs
10733 #undef TARGET_ADDRESS_COST
10734 #define TARGET_ADDRESS_COST s390_address_cost
10735 #undef TARGET_REGISTER_MOVE_COST
10736 #define TARGET_REGISTER_MOVE_COST s390_register_move_cost
10737 #undef TARGET_MEMORY_MOVE_COST
10738 #define TARGET_MEMORY_MOVE_COST s390_memory_move_cost
10740 #undef TARGET_MACHINE_DEPENDENT_REORG
10741 #define TARGET_MACHINE_DEPENDENT_REORG s390_reorg
10743 #undef TARGET_VALID_POINTER_MODE
10744 #define TARGET_VALID_POINTER_MODE s390_valid_pointer_mode
10746 #undef TARGET_BUILD_BUILTIN_VA_LIST
10747 #define TARGET_BUILD_BUILTIN_VA_LIST s390_build_builtin_va_list
10748 #undef TARGET_EXPAND_BUILTIN_VA_START
10749 #define TARGET_EXPAND_BUILTIN_VA_START s390_va_start
10750 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
10751 #define TARGET_GIMPLIFY_VA_ARG_EXPR s390_gimplify_va_arg
10753 #undef TARGET_PROMOTE_FUNCTION_MODE
10754 #define TARGET_PROMOTE_FUNCTION_MODE s390_promote_function_mode
10755 #undef TARGET_PASS_BY_REFERENCE
10756 #define TARGET_PASS_BY_REFERENCE s390_pass_by_reference
10758 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
10759 #define TARGET_FUNCTION_OK_FOR_SIBCALL s390_function_ok_for_sibcall
10760 #undef TARGET_FUNCTION_ARG
10761 #define TARGET_FUNCTION_ARG s390_function_arg
10762 #undef TARGET_FUNCTION_ARG_ADVANCE
10763 #define TARGET_FUNCTION_ARG_ADVANCE s390_function_arg_advance
10764 #undef TARGET_FUNCTION_VALUE
10765 #define TARGET_FUNCTION_VALUE s390_function_value
10766 #undef TARGET_LIBCALL_VALUE
10767 #define TARGET_LIBCALL_VALUE s390_libcall_value
10769 #undef TARGET_FIXED_CONDITION_CODE_REGS
10770 #define TARGET_FIXED_CONDITION_CODE_REGS s390_fixed_condition_code_regs
10772 #undef TARGET_CC_MODES_COMPATIBLE
10773 #define TARGET_CC_MODES_COMPATIBLE s390_cc_modes_compatible
10775 #undef TARGET_INVALID_WITHIN_DOLOOP
10776 #define TARGET_INVALID_WITHIN_DOLOOP hook_constcharptr_const_rtx_null
10779 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
10780 #define TARGET_ASM_OUTPUT_DWARF_DTPREL s390_output_dwarf_dtprel
10783 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
10784 #undef TARGET_MANGLE_TYPE
10785 #define TARGET_MANGLE_TYPE s390_mangle_type
10788 #undef TARGET_SCALAR_MODE_SUPPORTED_P
10789 #define TARGET_SCALAR_MODE_SUPPORTED_P s390_scalar_mode_supported_p
10791 #undef TARGET_PREFERRED_RELOAD_CLASS
10792 #define TARGET_PREFERRED_RELOAD_CLASS s390_preferred_reload_class
10794 #undef TARGET_SECONDARY_RELOAD
10795 #define TARGET_SECONDARY_RELOAD s390_secondary_reload
10797 #undef TARGET_LIBGCC_CMP_RETURN_MODE
10798 #define TARGET_LIBGCC_CMP_RETURN_MODE s390_libgcc_cmp_return_mode
10800 #undef TARGET_LIBGCC_SHIFT_COUNT_MODE
10801 #define TARGET_LIBGCC_SHIFT_COUNT_MODE s390_libgcc_shift_count_mode
10803 #undef TARGET_LEGITIMATE_ADDRESS_P
10804 #define TARGET_LEGITIMATE_ADDRESS_P s390_legitimate_address_p
10806 #undef TARGET_LEGITIMATE_CONSTANT_P
10807 #define TARGET_LEGITIMATE_CONSTANT_P s390_legitimate_constant_p
10809 #undef TARGET_CAN_ELIMINATE
10810 #define TARGET_CAN_ELIMINATE s390_can_eliminate
10812 #undef TARGET_CONDITIONAL_REGISTER_USAGE
10813 #define TARGET_CONDITIONAL_REGISTER_USAGE s390_conditional_register_usage
10815 #undef TARGET_LOOP_UNROLL_ADJUST
10816 #define TARGET_LOOP_UNROLL_ADJUST s390_loop_unroll_adjust
10818 #undef TARGET_ASM_TRAMPOLINE_TEMPLATE
10819 #define TARGET_ASM_TRAMPOLINE_TEMPLATE s390_asm_trampoline_template
10820 #undef TARGET_TRAMPOLINE_INIT
10821 #define TARGET_TRAMPOLINE_INIT s390_trampoline_init
10823 #undef TARGET_UNWIND_WORD_MODE
10824 #define TARGET_UNWIND_WORD_MODE s390_unwind_word_mode
10826 struct gcc_target targetm
= TARGET_INITIALIZER
;
10828 #include "gt-s390.h"