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;
1654 /* Map for smallest class containing reg regno. */
1656 const enum reg_class regclass_map
[FIRST_PSEUDO_REGISTER
] =
1657 { GENERAL_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1658 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1659 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1660 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1661 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1662 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1663 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1664 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1665 ADDR_REGS
, CC_REGS
, ADDR_REGS
, ADDR_REGS
,
1666 ACCESS_REGS
, ACCESS_REGS
1669 /* Return attribute type of insn. */
1671 static enum attr_type
1672 s390_safe_attr_type (rtx insn
)
1674 if (recog_memoized (insn
) >= 0)
1675 return get_attr_type (insn
);
1680 /* Return true if DISP is a valid short displacement. */
1683 s390_short_displacement (rtx disp
)
1685 /* No displacement is OK. */
1689 /* Without the long displacement facility we don't need to
1690 distingiush between long and short displacement. */
1691 if (!TARGET_LONG_DISPLACEMENT
)
1694 /* Integer displacement in range. */
1695 if (GET_CODE (disp
) == CONST_INT
)
1696 return INTVAL (disp
) >= 0 && INTVAL (disp
) < 4096;
1698 /* GOT offset is not OK, the GOT can be large. */
1699 if (GET_CODE (disp
) == CONST
1700 && GET_CODE (XEXP (disp
, 0)) == UNSPEC
1701 && (XINT (XEXP (disp
, 0), 1) == UNSPEC_GOT
1702 || XINT (XEXP (disp
, 0), 1) == UNSPEC_GOTNTPOFF
))
1705 /* All other symbolic constants are literal pool references,
1706 which are OK as the literal pool must be small. */
1707 if (GET_CODE (disp
) == CONST
)
1713 /* Decompose a RTL expression ADDR for a memory address into
1714 its components, returned in OUT.
1716 Returns false if ADDR is not a valid memory address, true
1717 otherwise. If OUT is NULL, don't return the components,
1718 but check for validity only.
1720 Note: Only addresses in canonical form are recognized.
1721 LEGITIMIZE_ADDRESS should convert non-canonical forms to the
1722 canonical form so that they will be recognized. */
1725 s390_decompose_address (rtx addr
, struct s390_address
*out
)
1727 HOST_WIDE_INT offset
= 0;
1728 rtx base
= NULL_RTX
;
1729 rtx indx
= NULL_RTX
;
1730 rtx disp
= NULL_RTX
;
1732 bool pointer
= false;
1733 bool base_ptr
= false;
1734 bool indx_ptr
= false;
1735 bool literal_pool
= false;
1737 /* We may need to substitute the literal pool base register into the address
1738 below. However, at this point we do not know which register is going to
1739 be used as base, so we substitute the arg pointer register. This is going
1740 to be treated as holding a pointer below -- it shouldn't be used for any
1742 rtx fake_pool_base
= gen_rtx_REG (Pmode
, ARG_POINTER_REGNUM
);
1744 /* Decompose address into base + index + displacement. */
1746 if (GET_CODE (addr
) == REG
|| GET_CODE (addr
) == UNSPEC
)
1749 else if (GET_CODE (addr
) == PLUS
)
1751 rtx op0
= XEXP (addr
, 0);
1752 rtx op1
= XEXP (addr
, 1);
1753 enum rtx_code code0
= GET_CODE (op0
);
1754 enum rtx_code code1
= GET_CODE (op1
);
1756 if (code0
== REG
|| code0
== UNSPEC
)
1758 if (code1
== REG
|| code1
== UNSPEC
)
1760 indx
= op0
; /* index + base */
1766 base
= op0
; /* base + displacement */
1771 else if (code0
== PLUS
)
1773 indx
= XEXP (op0
, 0); /* index + base + disp */
1774 base
= XEXP (op0
, 1);
1785 disp
= addr
; /* displacement */
1787 /* Extract integer part of displacement. */
1791 if (GET_CODE (disp
) == CONST_INT
)
1793 offset
= INTVAL (disp
);
1796 else if (GET_CODE (disp
) == CONST
1797 && GET_CODE (XEXP (disp
, 0)) == PLUS
1798 && GET_CODE (XEXP (XEXP (disp
, 0), 1)) == CONST_INT
)
1800 offset
= INTVAL (XEXP (XEXP (disp
, 0), 1));
1801 disp
= XEXP (XEXP (disp
, 0), 0);
1805 /* Strip off CONST here to avoid special case tests later. */
1806 if (disp
&& GET_CODE (disp
) == CONST
)
1807 disp
= XEXP (disp
, 0);
1809 /* We can convert literal pool addresses to
1810 displacements by basing them off the base register. */
1811 if (disp
&& GET_CODE (disp
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (disp
))
1813 /* Either base or index must be free to hold the base register. */
1815 base
= fake_pool_base
, literal_pool
= true;
1817 indx
= fake_pool_base
, literal_pool
= true;
1821 /* Mark up the displacement. */
1822 disp
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, disp
),
1823 UNSPEC_LTREL_OFFSET
);
1826 /* Validate base register. */
1829 if (GET_CODE (base
) == UNSPEC
)
1830 switch (XINT (base
, 1))
1834 disp
= gen_rtx_UNSPEC (Pmode
,
1835 gen_rtvec (1, XVECEXP (base
, 0, 0)),
1836 UNSPEC_LTREL_OFFSET
);
1840 base
= XVECEXP (base
, 0, 1);
1843 case UNSPEC_LTREL_BASE
:
1844 if (XVECLEN (base
, 0) == 1)
1845 base
= fake_pool_base
, literal_pool
= true;
1847 base
= XVECEXP (base
, 0, 1);
1855 || (GET_MODE (base
) != SImode
1856 && GET_MODE (base
) != Pmode
))
1859 if (REGNO (base
) == STACK_POINTER_REGNUM
1860 || REGNO (base
) == FRAME_POINTER_REGNUM
1861 || ((reload_completed
|| reload_in_progress
)
1862 && frame_pointer_needed
1863 && REGNO (base
) == HARD_FRAME_POINTER_REGNUM
)
1864 || REGNO (base
) == ARG_POINTER_REGNUM
1866 && REGNO (base
) == PIC_OFFSET_TABLE_REGNUM
))
1867 pointer
= base_ptr
= true;
1869 if ((reload_completed
|| reload_in_progress
)
1870 && base
== cfun
->machine
->base_reg
)
1871 pointer
= base_ptr
= literal_pool
= true;
1874 /* Validate index register. */
1877 if (GET_CODE (indx
) == UNSPEC
)
1878 switch (XINT (indx
, 1))
1882 disp
= gen_rtx_UNSPEC (Pmode
,
1883 gen_rtvec (1, XVECEXP (indx
, 0, 0)),
1884 UNSPEC_LTREL_OFFSET
);
1888 indx
= XVECEXP (indx
, 0, 1);
1891 case UNSPEC_LTREL_BASE
:
1892 if (XVECLEN (indx
, 0) == 1)
1893 indx
= fake_pool_base
, literal_pool
= true;
1895 indx
= XVECEXP (indx
, 0, 1);
1903 || (GET_MODE (indx
) != SImode
1904 && GET_MODE (indx
) != Pmode
))
1907 if (REGNO (indx
) == STACK_POINTER_REGNUM
1908 || REGNO (indx
) == FRAME_POINTER_REGNUM
1909 || ((reload_completed
|| reload_in_progress
)
1910 && frame_pointer_needed
1911 && REGNO (indx
) == HARD_FRAME_POINTER_REGNUM
)
1912 || REGNO (indx
) == ARG_POINTER_REGNUM
1914 && REGNO (indx
) == PIC_OFFSET_TABLE_REGNUM
))
1915 pointer
= indx_ptr
= true;
1917 if ((reload_completed
|| reload_in_progress
)
1918 && indx
== cfun
->machine
->base_reg
)
1919 pointer
= indx_ptr
= literal_pool
= true;
1922 /* Prefer to use pointer as base, not index. */
1923 if (base
&& indx
&& !base_ptr
1924 && (indx_ptr
|| (!REG_POINTER (base
) && REG_POINTER (indx
))))
1931 /* Validate displacement. */
1934 /* If virtual registers are involved, the displacement will change later
1935 anyway as the virtual registers get eliminated. This could make a
1936 valid displacement invalid, but it is more likely to make an invalid
1937 displacement valid, because we sometimes access the register save area
1938 via negative offsets to one of those registers.
1939 Thus we don't check the displacement for validity here. If after
1940 elimination the displacement turns out to be invalid after all,
1941 this is fixed up by reload in any case. */
1942 if (base
!= arg_pointer_rtx
1943 && indx
!= arg_pointer_rtx
1944 && base
!= return_address_pointer_rtx
1945 && indx
!= return_address_pointer_rtx
1946 && base
!= frame_pointer_rtx
1947 && indx
!= frame_pointer_rtx
1948 && base
!= virtual_stack_vars_rtx
1949 && indx
!= virtual_stack_vars_rtx
)
1950 if (!DISP_IN_RANGE (offset
))
1955 /* All the special cases are pointers. */
1958 /* In the small-PIC case, the linker converts @GOT
1959 and @GOTNTPOFF offsets to possible displacements. */
1960 if (GET_CODE (disp
) == UNSPEC
1961 && (XINT (disp
, 1) == UNSPEC_GOT
1962 || XINT (disp
, 1) == UNSPEC_GOTNTPOFF
)
1968 /* Accept pool label offsets. */
1969 else if (GET_CODE (disp
) == UNSPEC
1970 && XINT (disp
, 1) == UNSPEC_POOL_OFFSET
)
1973 /* Accept literal pool references. */
1974 else if (GET_CODE (disp
) == UNSPEC
1975 && XINT (disp
, 1) == UNSPEC_LTREL_OFFSET
)
1977 /* In case CSE pulled a non literal pool reference out of
1978 the pool we have to reject the address. This is
1979 especially important when loading the GOT pointer on non
1980 zarch CPUs. In this case the literal pool contains an lt
1981 relative offset to the _GLOBAL_OFFSET_TABLE_ label which
1982 will most likely exceed the displacement. */
1983 if (GET_CODE (XVECEXP (disp
, 0, 0)) != SYMBOL_REF
1984 || !CONSTANT_POOL_ADDRESS_P (XVECEXP (disp
, 0, 0)))
1987 orig_disp
= gen_rtx_CONST (Pmode
, disp
);
1990 /* If we have an offset, make sure it does not
1991 exceed the size of the constant pool entry. */
1992 rtx sym
= XVECEXP (disp
, 0, 0);
1993 if (offset
>= GET_MODE_SIZE (get_pool_mode (sym
)))
1996 orig_disp
= plus_constant (Pmode
, orig_disp
, offset
);
2011 out
->disp
= orig_disp
;
2012 out
->pointer
= pointer
;
2013 out
->literal_pool
= literal_pool
;
2019 /* Decompose a RTL expression OP for a shift count into its components,
2020 and return the base register in BASE and the offset in OFFSET.
2022 Return true if OP is a valid shift count, false if not. */
2025 s390_decompose_shift_count (rtx op
, rtx
*base
, HOST_WIDE_INT
*offset
)
2027 HOST_WIDE_INT off
= 0;
2029 /* We can have an integer constant, an address register,
2030 or a sum of the two. */
2031 if (GET_CODE (op
) == CONST_INT
)
2036 if (op
&& GET_CODE (op
) == PLUS
&& GET_CODE (XEXP (op
, 1)) == CONST_INT
)
2038 off
= INTVAL (XEXP (op
, 1));
2041 while (op
&& GET_CODE (op
) == SUBREG
)
2042 op
= SUBREG_REG (op
);
2044 if (op
&& GET_CODE (op
) != REG
)
2056 /* Return true if CODE is a valid address without index. */
2059 s390_legitimate_address_without_index_p (rtx op
)
2061 struct s390_address addr
;
2063 if (!s390_decompose_address (XEXP (op
, 0), &addr
))
2072 /* Return true if ADDR is of kind symbol_ref or symbol_ref + const_int
2073 and return these parts in SYMREF and ADDEND. You can pass NULL in
2074 SYMREF and/or ADDEND if you are not interested in these values.
2075 Literal pool references are *not* considered symbol references. */
2078 s390_symref_operand_p (rtx addr
, rtx
*symref
, HOST_WIDE_INT
*addend
)
2080 HOST_WIDE_INT tmpaddend
= 0;
2082 if (GET_CODE (addr
) == CONST
)
2083 addr
= XEXP (addr
, 0);
2085 if (GET_CODE (addr
) == PLUS
)
2087 if (GET_CODE (XEXP (addr
, 0)) == SYMBOL_REF
2088 && !CONSTANT_POOL_ADDRESS_P (XEXP (addr
, 0))
2089 && CONST_INT_P (XEXP (addr
, 1)))
2091 tmpaddend
= INTVAL (XEXP (addr
, 1));
2092 addr
= XEXP (addr
, 0);
2098 if (GET_CODE (addr
) != SYMBOL_REF
|| CONSTANT_POOL_ADDRESS_P (addr
))
2104 *addend
= tmpaddend
;
2110 /* Return true if the address in OP is valid for constraint letter C
2111 if wrapped in a MEM rtx. Set LIT_POOL_OK to true if it literal
2112 pool MEMs should be accepted. Only the Q, R, S, T constraint
2113 letters are allowed for C. */
2116 s390_check_qrst_address (char c
, rtx op
, bool lit_pool_ok
)
2118 struct s390_address addr
;
2119 bool decomposed
= false;
2121 /* This check makes sure that no symbolic address (except literal
2122 pool references) are accepted by the R or T constraints. */
2123 if (s390_symref_operand_p (op
, NULL
, NULL
))
2126 /* Ensure literal pool references are only accepted if LIT_POOL_OK. */
2129 if (!s390_decompose_address (op
, &addr
))
2131 if (addr
.literal_pool
)
2138 case 'Q': /* no index short displacement */
2139 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2143 if (!s390_short_displacement (addr
.disp
))
2147 case 'R': /* with index short displacement */
2148 if (TARGET_LONG_DISPLACEMENT
)
2150 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2152 if (!s390_short_displacement (addr
.disp
))
2155 /* Any invalid address here will be fixed up by reload,
2156 so accept it for the most generic constraint. */
2159 case 'S': /* no index long displacement */
2160 if (!TARGET_LONG_DISPLACEMENT
)
2162 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2166 if (s390_short_displacement (addr
.disp
))
2170 case 'T': /* with index long displacement */
2171 if (!TARGET_LONG_DISPLACEMENT
)
2173 /* Any invalid address here will be fixed up by reload,
2174 so accept it for the most generic constraint. */
2175 if ((decomposed
|| s390_decompose_address (op
, &addr
))
2176 && s390_short_displacement (addr
.disp
))
2186 /* Evaluates constraint strings described by the regular expression
2187 ([A|B|Z](Q|R|S|T))|U|W|Y and returns 1 if OP is a valid operand for
2188 the constraint given in STR, or 0 else. */
2191 s390_mem_constraint (const char *str
, rtx op
)
2198 /* Check for offsettable variants of memory constraints. */
2199 if (!MEM_P (op
) || MEM_VOLATILE_P (op
))
2201 if ((reload_completed
|| reload_in_progress
)
2202 ? !offsettable_memref_p (op
) : !offsettable_nonstrict_memref_p (op
))
2204 return s390_check_qrst_address (str
[1], XEXP (op
, 0), true);
2206 /* Check for non-literal-pool variants of memory constraints. */
2209 return s390_check_qrst_address (str
[1], XEXP (op
, 0), false);
2214 if (GET_CODE (op
) != MEM
)
2216 return s390_check_qrst_address (c
, XEXP (op
, 0), true);
2218 return (s390_check_qrst_address ('Q', op
, true)
2219 || s390_check_qrst_address ('R', op
, true));
2221 return (s390_check_qrst_address ('S', op
, true)
2222 || s390_check_qrst_address ('T', op
, true));
2224 /* Simply check for the basic form of a shift count. Reload will
2225 take care of making sure we have a proper base register. */
2226 if (!s390_decompose_shift_count (op
, NULL
, NULL
))
2230 return s390_check_qrst_address (str
[1], op
, true);
2238 /* Evaluates constraint strings starting with letter O. Input
2239 parameter C is the second letter following the "O" in the constraint
2240 string. Returns 1 if VALUE meets the respective constraint and 0
2244 s390_O_constraint_str (const char c
, HOST_WIDE_INT value
)
2252 return trunc_int_for_mode (value
, SImode
) == value
;
2256 || s390_single_part (GEN_INT (value
), DImode
, SImode
, 0) == 1;
2259 return s390_single_part (GEN_INT (value
- 1), DImode
, SImode
, -1) == 1;
2267 /* Evaluates constraint strings starting with letter N. Parameter STR
2268 contains the letters following letter "N" in the constraint string.
2269 Returns true if VALUE matches the constraint. */
2272 s390_N_constraint_str (const char *str
, HOST_WIDE_INT value
)
2274 enum machine_mode mode
, part_mode
;
2276 int part
, part_goal
;
2282 part_goal
= str
[0] - '0';
2326 if (GET_MODE_SIZE (mode
) <= GET_MODE_SIZE (part_mode
))
2329 part
= s390_single_part (GEN_INT (value
), mode
, part_mode
, def
);
2332 if (part_goal
!= -1 && part_goal
!= part
)
2339 /* Returns true if the input parameter VALUE is a float zero. */
2342 s390_float_const_zero_p (rtx value
)
2344 return (GET_MODE_CLASS (GET_MODE (value
)) == MODE_FLOAT
2345 && value
== CONST0_RTX (GET_MODE (value
)));
2348 /* Implement TARGET_REGISTER_MOVE_COST. */
2351 s390_register_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED
,
2352 reg_class_t from
, reg_class_t to
)
2354 /* On s390, copy between fprs and gprs is expensive. */
2355 if ((reg_classes_intersect_p (from
, GENERAL_REGS
)
2356 && reg_classes_intersect_p (to
, FP_REGS
))
2357 || (reg_classes_intersect_p (from
, FP_REGS
)
2358 && reg_classes_intersect_p (to
, GENERAL_REGS
)))
2364 /* Implement TARGET_MEMORY_MOVE_COST. */
2367 s390_memory_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED
,
2368 reg_class_t rclass ATTRIBUTE_UNUSED
,
2369 bool in ATTRIBUTE_UNUSED
)
2374 /* Compute a (partial) cost for rtx X. Return true if the complete
2375 cost has been computed, and false if subexpressions should be
2376 scanned. In either case, *TOTAL contains the cost result.
2377 CODE contains GET_CODE (x), OUTER_CODE contains the code
2378 of the superexpression of x. */
2381 s390_rtx_costs (rtx x
, int code
, int outer_code
, int opno ATTRIBUTE_UNUSED
,
2382 int *total
, bool speed ATTRIBUTE_UNUSED
)
2405 *total
= COSTS_N_INSNS (1);
2410 *total
= COSTS_N_INSNS (1);
2414 switch (GET_MODE (x
))
2418 rtx left
= XEXP (x
, 0);
2419 rtx right
= XEXP (x
, 1);
2420 if (GET_CODE (right
) == CONST_INT
2421 && CONST_OK_FOR_K (INTVAL (right
)))
2422 *total
= s390_cost
->mhi
;
2423 else if (GET_CODE (left
) == SIGN_EXTEND
)
2424 *total
= s390_cost
->mh
;
2426 *total
= s390_cost
->ms
; /* msr, ms, msy */
2431 rtx left
= XEXP (x
, 0);
2432 rtx right
= XEXP (x
, 1);
2435 if (GET_CODE (right
) == CONST_INT
2436 && CONST_OK_FOR_K (INTVAL (right
)))
2437 *total
= s390_cost
->mghi
;
2438 else if (GET_CODE (left
) == SIGN_EXTEND
)
2439 *total
= s390_cost
->msgf
;
2441 *total
= s390_cost
->msg
; /* msgr, msg */
2443 else /* TARGET_31BIT */
2445 if (GET_CODE (left
) == SIGN_EXTEND
2446 && GET_CODE (right
) == SIGN_EXTEND
)
2447 /* mulsidi case: mr, m */
2448 *total
= s390_cost
->m
;
2449 else if (GET_CODE (left
) == ZERO_EXTEND
2450 && GET_CODE (right
) == ZERO_EXTEND
2451 && TARGET_CPU_ZARCH
)
2452 /* umulsidi case: ml, mlr */
2453 *total
= s390_cost
->ml
;
2455 /* Complex calculation is required. */
2456 *total
= COSTS_N_INSNS (40);
2462 *total
= s390_cost
->mult_df
;
2465 *total
= s390_cost
->mxbr
;
2473 switch (GET_MODE (x
))
2476 *total
= s390_cost
->madbr
;
2479 *total
= s390_cost
->maebr
;
2484 /* Negate in the third argument is free: FMSUB. */
2485 if (GET_CODE (XEXP (x
, 2)) == NEG
)
2487 *total
+= (rtx_cost (XEXP (x
, 0), FMA
, 0, speed
)
2488 + rtx_cost (XEXP (x
, 1), FMA
, 1, speed
)
2489 + rtx_cost (XEXP (XEXP (x
, 2), 0), FMA
, 2, speed
));
2496 if (GET_MODE (x
) == TImode
) /* 128 bit division */
2497 *total
= s390_cost
->dlgr
;
2498 else if (GET_MODE (x
) == DImode
)
2500 rtx right
= XEXP (x
, 1);
2501 if (GET_CODE (right
) == ZERO_EXTEND
) /* 64 by 32 bit division */
2502 *total
= s390_cost
->dlr
;
2503 else /* 64 by 64 bit division */
2504 *total
= s390_cost
->dlgr
;
2506 else if (GET_MODE (x
) == SImode
) /* 32 bit division */
2507 *total
= s390_cost
->dlr
;
2512 if (GET_MODE (x
) == DImode
)
2514 rtx right
= XEXP (x
, 1);
2515 if (GET_CODE (right
) == ZERO_EXTEND
) /* 64 by 32 bit division */
2517 *total
= s390_cost
->dsgfr
;
2519 *total
= s390_cost
->dr
;
2520 else /* 64 by 64 bit division */
2521 *total
= s390_cost
->dsgr
;
2523 else if (GET_MODE (x
) == SImode
) /* 32 bit division */
2524 *total
= s390_cost
->dlr
;
2525 else if (GET_MODE (x
) == SFmode
)
2527 *total
= s390_cost
->debr
;
2529 else if (GET_MODE (x
) == DFmode
)
2531 *total
= s390_cost
->ddbr
;
2533 else if (GET_MODE (x
) == TFmode
)
2535 *total
= s390_cost
->dxbr
;
2540 if (GET_MODE (x
) == SFmode
)
2541 *total
= s390_cost
->sqebr
;
2542 else if (GET_MODE (x
) == DFmode
)
2543 *total
= s390_cost
->sqdbr
;
2545 *total
= s390_cost
->sqxbr
;
2550 if (outer_code
== MULT
|| outer_code
== DIV
|| outer_code
== MOD
2551 || outer_code
== PLUS
|| outer_code
== MINUS
2552 || outer_code
== COMPARE
)
2557 *total
= COSTS_N_INSNS (1);
2558 if (GET_CODE (XEXP (x
, 0)) == AND
2559 && GET_CODE (XEXP (x
, 1)) == CONST_INT
2560 && GET_CODE (XEXP (XEXP (x
, 0), 1)) == CONST_INT
)
2562 rtx op0
= XEXP (XEXP (x
, 0), 0);
2563 rtx op1
= XEXP (XEXP (x
, 0), 1);
2564 rtx op2
= XEXP (x
, 1);
2566 if (memory_operand (op0
, GET_MODE (op0
))
2567 && s390_tm_ccmode (op1
, op2
, 0) != VOIDmode
)
2569 if (register_operand (op0
, GET_MODE (op0
))
2570 && s390_tm_ccmode (op1
, op2
, 1) != VOIDmode
)
2580 /* Return the cost of an address rtx ADDR. */
2583 s390_address_cost (rtx addr
, bool speed ATTRIBUTE_UNUSED
)
2585 struct s390_address ad
;
2586 if (!s390_decompose_address (addr
, &ad
))
2589 return ad
.indx
? COSTS_N_INSNS (1) + 1 : COSTS_N_INSNS (1);
2592 /* If OP is a SYMBOL_REF of a thread-local symbol, return its TLS mode,
2593 otherwise return 0. */
2596 tls_symbolic_operand (rtx op
)
2598 if (GET_CODE (op
) != SYMBOL_REF
)
2600 return SYMBOL_REF_TLS_MODEL (op
);
2603 /* Split DImode access register reference REG (on 64-bit) into its constituent
2604 low and high parts, and store them into LO and HI. Note that gen_lowpart/
2605 gen_highpart cannot be used as they assume all registers are word-sized,
2606 while our access registers have only half that size. */
2609 s390_split_access_reg (rtx reg
, rtx
*lo
, rtx
*hi
)
2611 gcc_assert (TARGET_64BIT
);
2612 gcc_assert (ACCESS_REG_P (reg
));
2613 gcc_assert (GET_MODE (reg
) == DImode
);
2614 gcc_assert (!(REGNO (reg
) & 1));
2616 *lo
= gen_rtx_REG (SImode
, REGNO (reg
) + 1);
2617 *hi
= gen_rtx_REG (SImode
, REGNO (reg
));
2620 /* Return true if OP contains a symbol reference */
2623 symbolic_reference_mentioned_p (rtx op
)
2628 if (GET_CODE (op
) == SYMBOL_REF
|| GET_CODE (op
) == LABEL_REF
)
2631 fmt
= GET_RTX_FORMAT (GET_CODE (op
));
2632 for (i
= GET_RTX_LENGTH (GET_CODE (op
)) - 1; i
>= 0; i
--)
2638 for (j
= XVECLEN (op
, i
) - 1; j
>= 0; j
--)
2639 if (symbolic_reference_mentioned_p (XVECEXP (op
, i
, j
)))
2643 else if (fmt
[i
] == 'e' && symbolic_reference_mentioned_p (XEXP (op
, i
)))
2650 /* Return true if OP contains a reference to a thread-local symbol. */
2653 tls_symbolic_reference_mentioned_p (rtx op
)
2658 if (GET_CODE (op
) == SYMBOL_REF
)
2659 return tls_symbolic_operand (op
);
2661 fmt
= GET_RTX_FORMAT (GET_CODE (op
));
2662 for (i
= GET_RTX_LENGTH (GET_CODE (op
)) - 1; i
>= 0; i
--)
2668 for (j
= XVECLEN (op
, i
) - 1; j
>= 0; j
--)
2669 if (tls_symbolic_reference_mentioned_p (XVECEXP (op
, i
, j
)))
2673 else if (fmt
[i
] == 'e' && tls_symbolic_reference_mentioned_p (XEXP (op
, i
)))
2681 /* Return true if OP is a legitimate general operand when
2682 generating PIC code. It is given that flag_pic is on
2683 and that OP satisfies CONSTANT_P or is a CONST_DOUBLE. */
2686 legitimate_pic_operand_p (rtx op
)
2688 /* Accept all non-symbolic constants. */
2689 if (!SYMBOLIC_CONST (op
))
2692 /* Reject everything else; must be handled
2693 via emit_symbolic_move. */
2697 /* Returns true if the constant value OP is a legitimate general operand.
2698 It is given that OP satisfies CONSTANT_P or is a CONST_DOUBLE. */
2701 s390_legitimate_constant_p (enum machine_mode mode
, rtx op
)
2703 /* Accept all non-symbolic constants. */
2704 if (!SYMBOLIC_CONST (op
))
2707 /* Accept immediate LARL operands. */
2708 if (TARGET_CPU_ZARCH
&& larl_operand (op
, mode
))
2711 /* Thread-local symbols are never legal constants. This is
2712 so that emit_call knows that computing such addresses
2713 might require a function call. */
2714 if (TLS_SYMBOLIC_CONST (op
))
2717 /* In the PIC case, symbolic constants must *not* be
2718 forced into the literal pool. We accept them here,
2719 so that they will be handled by emit_symbolic_move. */
2723 /* All remaining non-PIC symbolic constants are
2724 forced into the literal pool. */
2728 /* Determine if it's legal to put X into the constant pool. This
2729 is not possible if X contains the address of a symbol that is
2730 not constant (TLS) or not known at final link time (PIC). */
2733 s390_cannot_force_const_mem (enum machine_mode mode
, rtx x
)
2735 switch (GET_CODE (x
))
2739 /* Accept all non-symbolic constants. */
2743 /* Labels are OK iff we are non-PIC. */
2744 return flag_pic
!= 0;
2747 /* 'Naked' TLS symbol references are never OK,
2748 non-TLS symbols are OK iff we are non-PIC. */
2749 if (tls_symbolic_operand (x
))
2752 return flag_pic
!= 0;
2755 return s390_cannot_force_const_mem (mode
, XEXP (x
, 0));
2758 return s390_cannot_force_const_mem (mode
, XEXP (x
, 0))
2759 || s390_cannot_force_const_mem (mode
, XEXP (x
, 1));
2762 switch (XINT (x
, 1))
2764 /* Only lt-relative or GOT-relative UNSPECs are OK. */
2765 case UNSPEC_LTREL_OFFSET
:
2773 case UNSPEC_GOTNTPOFF
:
2774 case UNSPEC_INDNTPOFF
:
2777 /* If the literal pool shares the code section, be put
2778 execute template placeholders into the pool as well. */
2780 return TARGET_CPU_ZARCH
;
2792 /* Returns true if the constant value OP is a legitimate general
2793 operand during and after reload. The difference to
2794 legitimate_constant_p is that this function will not accept
2795 a constant that would need to be forced to the literal pool
2796 before it can be used as operand.
2797 This function accepts all constants which can be loaded directly
2801 legitimate_reload_constant_p (rtx op
)
2803 /* Accept la(y) operands. */
2804 if (GET_CODE (op
) == CONST_INT
2805 && DISP_IN_RANGE (INTVAL (op
)))
2808 /* Accept l(g)hi/l(g)fi operands. */
2809 if (GET_CODE (op
) == CONST_INT
2810 && (CONST_OK_FOR_K (INTVAL (op
)) || CONST_OK_FOR_Os (INTVAL (op
))))
2813 /* Accept lliXX operands. */
2815 && GET_CODE (op
) == CONST_INT
2816 && trunc_int_for_mode (INTVAL (op
), word_mode
) == INTVAL (op
)
2817 && s390_single_part (op
, word_mode
, HImode
, 0) >= 0)
2821 && GET_CODE (op
) == CONST_INT
2822 && trunc_int_for_mode (INTVAL (op
), word_mode
) == INTVAL (op
)
2823 && s390_single_part (op
, word_mode
, SImode
, 0) >= 0)
2826 /* Accept larl operands. */
2827 if (TARGET_CPU_ZARCH
2828 && larl_operand (op
, VOIDmode
))
2831 /* Accept floating-point zero operands that fit into a single GPR. */
2832 if (GET_CODE (op
) == CONST_DOUBLE
2833 && s390_float_const_zero_p (op
)
2834 && GET_MODE_SIZE (GET_MODE (op
)) <= UNITS_PER_WORD
)
2837 /* Accept double-word operands that can be split. */
2838 if (GET_CODE (op
) == CONST_INT
2839 && trunc_int_for_mode (INTVAL (op
), word_mode
) != INTVAL (op
))
2841 enum machine_mode dword_mode
= word_mode
== SImode
? DImode
: TImode
;
2842 rtx hi
= operand_subword (op
, 0, 0, dword_mode
);
2843 rtx lo
= operand_subword (op
, 1, 0, dword_mode
);
2844 return legitimate_reload_constant_p (hi
)
2845 && legitimate_reload_constant_p (lo
);
2848 /* Everything else cannot be handled without reload. */
2852 /* Returns true if the constant value OP is a legitimate fp operand
2853 during and after reload.
2854 This function accepts all constants which can be loaded directly
2858 legitimate_reload_fp_constant_p (rtx op
)
2860 /* Accept floating-point zero operands if the load zero instruction
2863 && GET_CODE (op
) == CONST_DOUBLE
2864 && s390_float_const_zero_p (op
))
2870 /* Given an rtx OP being reloaded into a reg required to be in class RCLASS,
2871 return the class of reg to actually use. */
2874 s390_preferred_reload_class (rtx op
, reg_class_t rclass
)
2876 switch (GET_CODE (op
))
2878 /* Constants we cannot reload into general registers
2879 must be forced into the literal pool. */
2882 if (reg_class_subset_p (GENERAL_REGS
, rclass
)
2883 && legitimate_reload_constant_p (op
))
2884 return GENERAL_REGS
;
2885 else if (reg_class_subset_p (ADDR_REGS
, rclass
)
2886 && legitimate_reload_constant_p (op
))
2888 else if (reg_class_subset_p (FP_REGS
, rclass
)
2889 && legitimate_reload_fp_constant_p (op
))
2893 /* If a symbolic constant or a PLUS is reloaded,
2894 it is most likely being used as an address, so
2895 prefer ADDR_REGS. If 'class' is not a superset
2896 of ADDR_REGS, e.g. FP_REGS, reject this reload. */
2900 if (!legitimate_reload_constant_p (op
))
2904 /* load address will be used. */
2905 if (reg_class_subset_p (ADDR_REGS
, rclass
))
2917 /* Return true if ADDR is SYMBOL_REF + addend with addend being a
2918 multiple of ALIGNMENT and the SYMBOL_REF being naturally
2922 s390_check_symref_alignment (rtx addr
, HOST_WIDE_INT alignment
)
2924 HOST_WIDE_INT addend
;
2927 if (!s390_symref_operand_p (addr
, &symref
, &addend
))
2930 return (!SYMBOL_REF_NOT_NATURALLY_ALIGNED_P (symref
)
2931 && !(addend
& (alignment
- 1)));
2934 /* ADDR is moved into REG using larl. If ADDR isn't a valid larl
2935 operand SCRATCH is used to reload the even part of the address and
2939 s390_reload_larl_operand (rtx reg
, rtx addr
, rtx scratch
)
2941 HOST_WIDE_INT addend
;
2944 if (!s390_symref_operand_p (addr
, &symref
, &addend
))
2948 /* Easy case. The addend is even so larl will do fine. */
2949 emit_move_insn (reg
, addr
);
2952 /* We can leave the scratch register untouched if the target
2953 register is a valid base register. */
2954 if (REGNO (reg
) < FIRST_PSEUDO_REGISTER
2955 && REGNO_REG_CLASS (REGNO (reg
)) == ADDR_REGS
)
2958 gcc_assert (REGNO (scratch
) < FIRST_PSEUDO_REGISTER
);
2959 gcc_assert (REGNO_REG_CLASS (REGNO (scratch
)) == ADDR_REGS
);
2962 emit_move_insn (scratch
,
2963 gen_rtx_CONST (Pmode
,
2964 gen_rtx_PLUS (Pmode
, symref
,
2965 GEN_INT (addend
- 1))));
2967 emit_move_insn (scratch
, symref
);
2969 /* Increment the address using la in order to avoid clobbering cc. */
2970 emit_move_insn (reg
, gen_rtx_PLUS (Pmode
, scratch
, const1_rtx
));
2974 /* Generate what is necessary to move between REG and MEM using
2975 SCRATCH. The direction is given by TOMEM. */
2978 s390_reload_symref_address (rtx reg
, rtx mem
, rtx scratch
, bool tomem
)
2980 /* Reload might have pulled a constant out of the literal pool.
2981 Force it back in. */
2982 if (CONST_INT_P (mem
) || GET_CODE (mem
) == CONST_DOUBLE
2983 || GET_CODE (mem
) == CONST
)
2984 mem
= force_const_mem (GET_MODE (reg
), mem
);
2986 gcc_assert (MEM_P (mem
));
2988 /* For a load from memory we can leave the scratch register
2989 untouched if the target register is a valid base register. */
2991 && REGNO (reg
) < FIRST_PSEUDO_REGISTER
2992 && REGNO_REG_CLASS (REGNO (reg
)) == ADDR_REGS
2993 && GET_MODE (reg
) == GET_MODE (scratch
))
2996 /* Load address into scratch register. Since we can't have a
2997 secondary reload for a secondary reload we have to cover the case
2998 where larl would need a secondary reload here as well. */
2999 s390_reload_larl_operand (scratch
, XEXP (mem
, 0), scratch
);
3001 /* Now we can use a standard load/store to do the move. */
3003 emit_move_insn (replace_equiv_address (mem
, scratch
), reg
);
3005 emit_move_insn (reg
, replace_equiv_address (mem
, scratch
));
3008 /* Inform reload about cases where moving X with a mode MODE to a register in
3009 RCLASS requires an extra scratch or immediate register. Return the class
3010 needed for the immediate register. */
3013 s390_secondary_reload (bool in_p
, rtx x
, reg_class_t rclass_i
,
3014 enum machine_mode mode
, secondary_reload_info
*sri
)
3016 enum reg_class rclass
= (enum reg_class
) rclass_i
;
3018 /* Intermediate register needed. */
3019 if (reg_classes_intersect_p (CC_REGS
, rclass
))
3020 return GENERAL_REGS
;
3024 HOST_WIDE_INT offset
;
3027 /* On z10 several optimizer steps may generate larl operands with
3030 && s390_symref_operand_p (x
, &symref
, &offset
)
3032 && !SYMBOL_REF_ALIGN1_P (symref
)
3033 && (offset
& 1) == 1)
3034 sri
->icode
= ((mode
== DImode
) ? CODE_FOR_reloaddi_larl_odd_addend_z10
3035 : CODE_FOR_reloadsi_larl_odd_addend_z10
);
3037 /* On z10 we need a scratch register when moving QI, TI or floating
3038 point mode values from or to a memory location with a SYMBOL_REF
3039 or if the symref addend of a SI or DI move is not aligned to the
3040 width of the access. */
3042 && s390_symref_operand_p (XEXP (x
, 0), NULL
, NULL
)
3043 && (mode
== QImode
|| mode
== TImode
|| FLOAT_MODE_P (mode
)
3044 || (!TARGET_ZARCH
&& mode
== DImode
)
3045 || ((mode
== HImode
|| mode
== SImode
|| mode
== DImode
)
3046 && (!s390_check_symref_alignment (XEXP (x
, 0),
3047 GET_MODE_SIZE (mode
))))))
3049 #define __SECONDARY_RELOAD_CASE(M,m) \
3052 sri->icode = in_p ? CODE_FOR_reload##m##di_toreg_z10 : \
3053 CODE_FOR_reload##m##di_tomem_z10; \
3055 sri->icode = in_p ? CODE_FOR_reload##m##si_toreg_z10 : \
3056 CODE_FOR_reload##m##si_tomem_z10; \
3059 switch (GET_MODE (x
))
3061 __SECONDARY_RELOAD_CASE (QI
, qi
);
3062 __SECONDARY_RELOAD_CASE (HI
, hi
);
3063 __SECONDARY_RELOAD_CASE (SI
, si
);
3064 __SECONDARY_RELOAD_CASE (DI
, di
);
3065 __SECONDARY_RELOAD_CASE (TI
, ti
);
3066 __SECONDARY_RELOAD_CASE (SF
, sf
);
3067 __SECONDARY_RELOAD_CASE (DF
, df
);
3068 __SECONDARY_RELOAD_CASE (TF
, tf
);
3069 __SECONDARY_RELOAD_CASE (SD
, sd
);
3070 __SECONDARY_RELOAD_CASE (DD
, dd
);
3071 __SECONDARY_RELOAD_CASE (TD
, td
);
3076 #undef __SECONDARY_RELOAD_CASE
3080 /* We need a scratch register when loading a PLUS expression which
3081 is not a legitimate operand of the LOAD ADDRESS instruction. */
3082 if (in_p
&& s390_plus_operand (x
, mode
))
3083 sri
->icode
= (TARGET_64BIT
?
3084 CODE_FOR_reloaddi_plus
: CODE_FOR_reloadsi_plus
);
3086 /* Performing a multiword move from or to memory we have to make sure the
3087 second chunk in memory is addressable without causing a displacement
3088 overflow. If that would be the case we calculate the address in
3089 a scratch register. */
3091 && GET_CODE (XEXP (x
, 0)) == PLUS
3092 && GET_CODE (XEXP (XEXP (x
, 0), 1)) == CONST_INT
3093 && !DISP_IN_RANGE (INTVAL (XEXP (XEXP (x
, 0), 1))
3094 + GET_MODE_SIZE (mode
) - 1))
3096 /* For GENERAL_REGS a displacement overflow is no problem if occurring
3097 in a s_operand address since we may fallback to lm/stm. So we only
3098 have to care about overflows in the b+i+d case. */
3099 if ((reg_classes_intersect_p (GENERAL_REGS
, rclass
)
3100 && s390_class_max_nregs (GENERAL_REGS
, mode
) > 1
3101 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == PLUS
)
3102 /* For FP_REGS no lm/stm is available so this check is triggered
3103 for displacement overflows in b+i+d and b+d like addresses. */
3104 || (reg_classes_intersect_p (FP_REGS
, rclass
)
3105 && s390_class_max_nregs (FP_REGS
, mode
) > 1))
3108 sri
->icode
= (TARGET_64BIT
?
3109 CODE_FOR_reloaddi_nonoffmem_in
:
3110 CODE_FOR_reloadsi_nonoffmem_in
);
3112 sri
->icode
= (TARGET_64BIT
?
3113 CODE_FOR_reloaddi_nonoffmem_out
:
3114 CODE_FOR_reloadsi_nonoffmem_out
);
3118 /* A scratch address register is needed when a symbolic constant is
3119 copied to r0 compiling with -fPIC. In other cases the target
3120 register might be used as temporary (see legitimize_pic_address). */
3121 if (in_p
&& SYMBOLIC_CONST (x
) && flag_pic
== 2 && rclass
!= ADDR_REGS
)
3122 sri
->icode
= (TARGET_64BIT
?
3123 CODE_FOR_reloaddi_PIC_addr
:
3124 CODE_FOR_reloadsi_PIC_addr
);
3126 /* Either scratch or no register needed. */
3130 /* Generate code to load SRC, which is PLUS that is not a
3131 legitimate operand for the LA instruction, into TARGET.
3132 SCRATCH may be used as scratch register. */
3135 s390_expand_plus_operand (rtx target
, rtx src
,
3139 struct s390_address ad
;
3141 /* src must be a PLUS; get its two operands. */
3142 gcc_assert (GET_CODE (src
) == PLUS
);
3143 gcc_assert (GET_MODE (src
) == Pmode
);
3145 /* Check if any of the two operands is already scheduled
3146 for replacement by reload. This can happen e.g. when
3147 float registers occur in an address. */
3148 sum1
= find_replacement (&XEXP (src
, 0));
3149 sum2
= find_replacement (&XEXP (src
, 1));
3150 src
= gen_rtx_PLUS (Pmode
, sum1
, sum2
);
3152 /* If the address is already strictly valid, there's nothing to do. */
3153 if (!s390_decompose_address (src
, &ad
)
3154 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
3155 || (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
))))
3157 /* Otherwise, one of the operands cannot be an address register;
3158 we reload its value into the scratch register. */
3159 if (true_regnum (sum1
) < 1 || true_regnum (sum1
) > 15)
3161 emit_move_insn (scratch
, sum1
);
3164 if (true_regnum (sum2
) < 1 || true_regnum (sum2
) > 15)
3166 emit_move_insn (scratch
, sum2
);
3170 /* According to the way these invalid addresses are generated
3171 in reload.c, it should never happen (at least on s390) that
3172 *neither* of the PLUS components, after find_replacements
3173 was applied, is an address register. */
3174 if (sum1
== scratch
&& sum2
== scratch
)
3180 src
= gen_rtx_PLUS (Pmode
, sum1
, sum2
);
3183 /* Emit the LOAD ADDRESS pattern. Note that reload of PLUS
3184 is only ever performed on addresses, so we can mark the
3185 sum as legitimate for LA in any case. */
3186 s390_load_address (target
, src
);
3190 /* Return true if ADDR is a valid memory address.
3191 STRICT specifies whether strict register checking applies. */
3194 s390_legitimate_address_p (enum machine_mode mode
, rtx addr
, bool strict
)
3196 struct s390_address ad
;
3199 && larl_operand (addr
, VOIDmode
)
3200 && (mode
== VOIDmode
3201 || s390_check_symref_alignment (addr
, GET_MODE_SIZE (mode
))))
3204 if (!s390_decompose_address (addr
, &ad
))
3209 if (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
3212 if (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
)))
3218 && !(REGNO (ad
.base
) >= FIRST_PSEUDO_REGISTER
3219 || REGNO_REG_CLASS (REGNO (ad
.base
)) == ADDR_REGS
))
3223 && !(REGNO (ad
.indx
) >= FIRST_PSEUDO_REGISTER
3224 || REGNO_REG_CLASS (REGNO (ad
.indx
)) == ADDR_REGS
))
3230 /* Return true if OP is a valid operand for the LA instruction.
3231 In 31-bit, we need to prove that the result is used as an
3232 address, as LA performs only a 31-bit addition. */
3235 legitimate_la_operand_p (rtx op
)
3237 struct s390_address addr
;
3238 if (!s390_decompose_address (op
, &addr
))
3241 return (TARGET_64BIT
|| addr
.pointer
);
3244 /* Return true if it is valid *and* preferable to use LA to
3245 compute the sum of OP1 and OP2. */
3248 preferred_la_operand_p (rtx op1
, rtx op2
)
3250 struct s390_address addr
;
3252 if (op2
!= const0_rtx
)
3253 op1
= gen_rtx_PLUS (Pmode
, op1
, op2
);
3255 if (!s390_decompose_address (op1
, &addr
))
3257 if (addr
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (addr
.base
)))
3259 if (addr
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (addr
.indx
)))
3262 /* Avoid LA instructions with index register on z196; it is
3263 preferable to use regular add instructions when possible. */
3264 if (addr
.indx
&& s390_tune
== PROCESSOR_2817_Z196
)
3267 if (!TARGET_64BIT
&& !addr
.pointer
)
3273 if ((addr
.base
&& REG_P (addr
.base
) && REG_POINTER (addr
.base
))
3274 || (addr
.indx
&& REG_P (addr
.indx
) && REG_POINTER (addr
.indx
)))
3280 /* Emit a forced load-address operation to load SRC into DST.
3281 This will use the LOAD ADDRESS instruction even in situations
3282 where legitimate_la_operand_p (SRC) returns false. */
3285 s390_load_address (rtx dst
, rtx src
)
3288 emit_move_insn (dst
, src
);
3290 emit_insn (gen_force_la_31 (dst
, src
));
3293 /* Return a legitimate reference for ORIG (an address) using the
3294 register REG. If REG is 0, a new pseudo is generated.
3296 There are two types of references that must be handled:
3298 1. Global data references must load the address from the GOT, via
3299 the PIC reg. An insn is emitted to do this load, and the reg is
3302 2. Static data references, constant pool addresses, and code labels
3303 compute the address as an offset from the GOT, whose base is in
3304 the PIC reg. Static data objects have SYMBOL_FLAG_LOCAL set to
3305 differentiate them from global data objects. The returned
3306 address is the PIC reg + an unspec constant.
3308 TARGET_LEGITIMIZE_ADDRESS_P rejects symbolic references unless the PIC
3309 reg also appears in the address. */
3312 legitimize_pic_address (rtx orig
, rtx reg
)
3318 gcc_assert (!TLS_SYMBOLIC_CONST (addr
));
3320 if (GET_CODE (addr
) == LABEL_REF
3321 || (GET_CODE (addr
) == SYMBOL_REF
&& SYMBOL_REF_LOCAL_P (addr
)))
3323 /* This is a local symbol. */
3324 if (TARGET_CPU_ZARCH
&& larl_operand (addr
, VOIDmode
))
3326 /* Access local symbols PC-relative via LARL.
3327 This is the same as in the non-PIC case, so it is
3328 handled automatically ... */
3332 /* Access local symbols relative to the GOT. */
3334 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3336 if (reload_in_progress
|| reload_completed
)
3337 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3339 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTOFF
);
3340 addr
= gen_rtx_CONST (Pmode
, addr
);
3341 addr
= force_const_mem (Pmode
, addr
);
3342 emit_move_insn (temp
, addr
);
3344 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3347 s390_load_address (reg
, new_rtx
);
3352 else if (GET_CODE (addr
) == SYMBOL_REF
)
3355 reg
= gen_reg_rtx (Pmode
);
3359 /* Assume GOT offset < 4k. This is handled the same way
3360 in both 31- and 64-bit code (@GOT). */
3362 if (reload_in_progress
|| reload_completed
)
3363 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3365 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOT
);
3366 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3367 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, new_rtx
);
3368 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3369 emit_move_insn (reg
, new_rtx
);
3372 else if (TARGET_CPU_ZARCH
)
3374 /* If the GOT offset might be >= 4k, we determine the position
3375 of the GOT entry via a PC-relative LARL (@GOTENT). */
3377 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3379 gcc_assert (REGNO (temp
) >= FIRST_PSEUDO_REGISTER
3380 || REGNO_REG_CLASS (REGNO (temp
)) == ADDR_REGS
);
3382 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTENT
);
3383 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3384 emit_move_insn (temp
, new_rtx
);
3386 new_rtx
= gen_const_mem (Pmode
, temp
);
3387 emit_move_insn (reg
, new_rtx
);
3392 /* If the GOT offset might be >= 4k, we have to load it
3393 from the literal pool (@GOT). */
3395 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3397 gcc_assert (REGNO (temp
) >= FIRST_PSEUDO_REGISTER
3398 || REGNO_REG_CLASS (REGNO (temp
)) == ADDR_REGS
);
3400 if (reload_in_progress
|| reload_completed
)
3401 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3403 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOT
);
3404 addr
= gen_rtx_CONST (Pmode
, addr
);
3405 addr
= force_const_mem (Pmode
, addr
);
3406 emit_move_insn (temp
, addr
);
3408 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3409 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3410 emit_move_insn (reg
, new_rtx
);
3416 if (GET_CODE (addr
) == CONST
)
3418 addr
= XEXP (addr
, 0);
3419 if (GET_CODE (addr
) == UNSPEC
)
3421 gcc_assert (XVECLEN (addr
, 0) == 1);
3422 switch (XINT (addr
, 1))
3424 /* If someone moved a GOT-relative UNSPEC
3425 out of the literal pool, force them back in. */
3428 new_rtx
= force_const_mem (Pmode
, orig
);
3431 /* @GOT is OK as is if small. */
3434 new_rtx
= force_const_mem (Pmode
, orig
);
3437 /* @GOTENT is OK as is. */
3441 /* @PLT is OK as is on 64-bit, must be converted to
3442 GOT-relative @PLTOFF on 31-bit. */
3444 if (!TARGET_CPU_ZARCH
)
3446 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3448 if (reload_in_progress
|| reload_completed
)
3449 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3451 addr
= XVECEXP (addr
, 0, 0);
3452 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
),
3454 addr
= gen_rtx_CONST (Pmode
, addr
);
3455 addr
= force_const_mem (Pmode
, addr
);
3456 emit_move_insn (temp
, addr
);
3458 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3461 s390_load_address (reg
, new_rtx
);
3467 /* Everything else cannot happen. */
3473 gcc_assert (GET_CODE (addr
) == PLUS
);
3475 if (GET_CODE (addr
) == PLUS
)
3477 rtx op0
= XEXP (addr
, 0), op1
= XEXP (addr
, 1);
3479 gcc_assert (!TLS_SYMBOLIC_CONST (op0
));
3480 gcc_assert (!TLS_SYMBOLIC_CONST (op1
));
3482 /* Check first to see if this is a constant offset
3483 from a local symbol reference. */
3484 if ((GET_CODE (op0
) == LABEL_REF
3485 || (GET_CODE (op0
) == SYMBOL_REF
&& SYMBOL_REF_LOCAL_P (op0
)))
3486 && GET_CODE (op1
) == CONST_INT
)
3488 if (TARGET_CPU_ZARCH
3489 && larl_operand (op0
, VOIDmode
)
3490 && INTVAL (op1
) < (HOST_WIDE_INT
)1 << 31
3491 && INTVAL (op1
) >= -((HOST_WIDE_INT
)1 << 31))
3493 if (INTVAL (op1
) & 1)
3495 /* LARL can't handle odd offsets, so emit a
3496 pair of LARL and LA. */
3497 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3499 if (!DISP_IN_RANGE (INTVAL (op1
)))
3501 HOST_WIDE_INT even
= INTVAL (op1
) - 1;
3502 op0
= gen_rtx_PLUS (Pmode
, op0
, GEN_INT (even
));
3503 op0
= gen_rtx_CONST (Pmode
, op0
);
3507 emit_move_insn (temp
, op0
);
3508 new_rtx
= gen_rtx_PLUS (Pmode
, temp
, op1
);
3512 s390_load_address (reg
, new_rtx
);
3518 /* If the offset is even, we can just use LARL.
3519 This will happen automatically. */
3524 /* Access local symbols relative to the GOT. */
3526 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3528 if (reload_in_progress
|| reload_completed
)
3529 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3531 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op0
),
3533 addr
= gen_rtx_PLUS (Pmode
, addr
, op1
);
3534 addr
= gen_rtx_CONST (Pmode
, addr
);
3535 addr
= force_const_mem (Pmode
, addr
);
3536 emit_move_insn (temp
, addr
);
3538 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3541 s390_load_address (reg
, new_rtx
);
3547 /* Now, check whether it is a GOT relative symbol plus offset
3548 that was pulled out of the literal pool. Force it back in. */
3550 else if (GET_CODE (op0
) == UNSPEC
3551 && GET_CODE (op1
) == CONST_INT
3552 && XINT (op0
, 1) == UNSPEC_GOTOFF
)
3554 gcc_assert (XVECLEN (op0
, 0) == 1);
3556 new_rtx
= force_const_mem (Pmode
, orig
);
3559 /* Otherwise, compute the sum. */
3562 base
= legitimize_pic_address (XEXP (addr
, 0), reg
);
3563 new_rtx
= legitimize_pic_address (XEXP (addr
, 1),
3564 base
== reg
? NULL_RTX
: reg
);
3565 if (GET_CODE (new_rtx
) == CONST_INT
)
3566 new_rtx
= plus_constant (Pmode
, base
, INTVAL (new_rtx
));
3569 if (GET_CODE (new_rtx
) == PLUS
&& CONSTANT_P (XEXP (new_rtx
, 1)))
3571 base
= gen_rtx_PLUS (Pmode
, base
, XEXP (new_rtx
, 0));
3572 new_rtx
= XEXP (new_rtx
, 1);
3574 new_rtx
= gen_rtx_PLUS (Pmode
, base
, new_rtx
);
3577 if (GET_CODE (new_rtx
) == CONST
)
3578 new_rtx
= XEXP (new_rtx
, 0);
3579 new_rtx
= force_operand (new_rtx
, 0);
3586 /* Load the thread pointer into a register. */
3589 s390_get_thread_pointer (void)
3591 rtx tp
= gen_reg_rtx (Pmode
);
3593 emit_move_insn (tp
, gen_rtx_REG (Pmode
, TP_REGNUM
));
3594 mark_reg_pointer (tp
, BITS_PER_WORD
);
3599 /* Emit a tls call insn. The call target is the SYMBOL_REF stored
3600 in s390_tls_symbol which always refers to __tls_get_offset.
3601 The returned offset is written to RESULT_REG and an USE rtx is
3602 generated for TLS_CALL. */
3604 static GTY(()) rtx s390_tls_symbol
;
3607 s390_emit_tls_call_insn (rtx result_reg
, rtx tls_call
)
3612 emit_insn (s390_load_got ());
3614 if (!s390_tls_symbol
)
3615 s390_tls_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "__tls_get_offset");
3617 insn
= s390_emit_call (s390_tls_symbol
, tls_call
, result_reg
,
3618 gen_rtx_REG (Pmode
, RETURN_REGNUM
));
3620 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), result_reg
);
3621 RTL_CONST_CALL_P (insn
) = 1;
3624 /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute
3625 this (thread-local) address. REG may be used as temporary. */
3628 legitimize_tls_address (rtx addr
, rtx reg
)
3630 rtx new_rtx
, tls_call
, temp
, base
, r2
, insn
;
3632 if (GET_CODE (addr
) == SYMBOL_REF
)
3633 switch (tls_symbolic_operand (addr
))
3635 case TLS_MODEL_GLOBAL_DYNAMIC
:
3637 r2
= gen_rtx_REG (Pmode
, 2);
3638 tls_call
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_TLSGD
);
3639 new_rtx
= gen_rtx_CONST (Pmode
, tls_call
);
3640 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3641 emit_move_insn (r2
, new_rtx
);
3642 s390_emit_tls_call_insn (r2
, tls_call
);
3643 insn
= get_insns ();
3646 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_NTPOFF
);
3647 temp
= gen_reg_rtx (Pmode
);
3648 emit_libcall_block (insn
, temp
, r2
, new_rtx
);
3650 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3653 s390_load_address (reg
, new_rtx
);
3658 case TLS_MODEL_LOCAL_DYNAMIC
:
3660 r2
= gen_rtx_REG (Pmode
, 2);
3661 tls_call
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, const0_rtx
), UNSPEC_TLSLDM
);
3662 new_rtx
= gen_rtx_CONST (Pmode
, tls_call
);
3663 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3664 emit_move_insn (r2
, new_rtx
);
3665 s390_emit_tls_call_insn (r2
, tls_call
);
3666 insn
= get_insns ();
3669 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, const0_rtx
), UNSPEC_TLSLDM_NTPOFF
);
3670 temp
= gen_reg_rtx (Pmode
);
3671 emit_libcall_block (insn
, temp
, r2
, new_rtx
);
3673 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3674 base
= gen_reg_rtx (Pmode
);
3675 s390_load_address (base
, new_rtx
);
3677 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_DTPOFF
);
3678 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3679 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3680 temp
= gen_reg_rtx (Pmode
);
3681 emit_move_insn (temp
, new_rtx
);
3683 new_rtx
= gen_rtx_PLUS (Pmode
, base
, temp
);
3686 s390_load_address (reg
, new_rtx
);
3691 case TLS_MODEL_INITIAL_EXEC
:
3694 /* Assume GOT offset < 4k. This is handled the same way
3695 in both 31- and 64-bit code. */
3697 if (reload_in_progress
|| reload_completed
)
3698 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3700 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTNTPOFF
);
3701 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3702 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, new_rtx
);
3703 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3704 temp
= gen_reg_rtx (Pmode
);
3705 emit_move_insn (temp
, new_rtx
);
3707 else if (TARGET_CPU_ZARCH
)
3709 /* If the GOT offset might be >= 4k, we determine the position
3710 of the GOT entry via a PC-relative LARL. */
3712 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_INDNTPOFF
);
3713 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3714 temp
= gen_reg_rtx (Pmode
);
3715 emit_move_insn (temp
, new_rtx
);
3717 new_rtx
= gen_const_mem (Pmode
, temp
);
3718 temp
= gen_reg_rtx (Pmode
);
3719 emit_move_insn (temp
, new_rtx
);
3723 /* If the GOT offset might be >= 4k, we have to load it
3724 from the literal pool. */
3726 if (reload_in_progress
|| reload_completed
)
3727 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3729 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTNTPOFF
);
3730 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3731 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3732 temp
= gen_reg_rtx (Pmode
);
3733 emit_move_insn (temp
, new_rtx
);
3735 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3736 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3738 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, new_rtx
, addr
), UNSPEC_TLS_LOAD
);
3739 temp
= gen_reg_rtx (Pmode
);
3740 emit_insn (gen_rtx_SET (Pmode
, temp
, new_rtx
));
3744 /* In position-dependent code, load the absolute address of
3745 the GOT entry from the literal pool. */
3747 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_INDNTPOFF
);
3748 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3749 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3750 temp
= gen_reg_rtx (Pmode
);
3751 emit_move_insn (temp
, new_rtx
);
3754 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3755 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, new_rtx
, addr
), UNSPEC_TLS_LOAD
);
3756 temp
= gen_reg_rtx (Pmode
);
3757 emit_insn (gen_rtx_SET (Pmode
, temp
, new_rtx
));
3760 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3763 s390_load_address (reg
, new_rtx
);
3768 case TLS_MODEL_LOCAL_EXEC
:
3769 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_NTPOFF
);
3770 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3771 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3772 temp
= gen_reg_rtx (Pmode
);
3773 emit_move_insn (temp
, new_rtx
);
3775 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3778 s390_load_address (reg
, new_rtx
);
3787 else if (GET_CODE (addr
) == CONST
&& GET_CODE (XEXP (addr
, 0)) == UNSPEC
)
3789 switch (XINT (XEXP (addr
, 0), 1))
3791 case UNSPEC_INDNTPOFF
:
3792 gcc_assert (TARGET_CPU_ZARCH
);
3801 else if (GET_CODE (addr
) == CONST
&& GET_CODE (XEXP (addr
, 0)) == PLUS
3802 && GET_CODE (XEXP (XEXP (addr
, 0), 1)) == CONST_INT
)
3804 new_rtx
= XEXP (XEXP (addr
, 0), 0);
3805 if (GET_CODE (new_rtx
) != SYMBOL_REF
)
3806 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3808 new_rtx
= legitimize_tls_address (new_rtx
, reg
);
3809 new_rtx
= plus_constant (Pmode
, new_rtx
,
3810 INTVAL (XEXP (XEXP (addr
, 0), 1)));
3811 new_rtx
= force_operand (new_rtx
, 0);
3815 gcc_unreachable (); /* for now ... */
3820 /* Emit insns making the address in operands[1] valid for a standard
3821 move to operands[0]. operands[1] is replaced by an address which
3822 should be used instead of the former RTX to emit the move
3826 emit_symbolic_move (rtx
*operands
)
3828 rtx temp
= !can_create_pseudo_p () ? operands
[0] : gen_reg_rtx (Pmode
);
3830 if (GET_CODE (operands
[0]) == MEM
)
3831 operands
[1] = force_reg (Pmode
, operands
[1]);
3832 else if (TLS_SYMBOLIC_CONST (operands
[1]))
3833 operands
[1] = legitimize_tls_address (operands
[1], temp
);
3835 operands
[1] = legitimize_pic_address (operands
[1], temp
);
3838 /* Try machine-dependent ways of modifying an illegitimate address X
3839 to be legitimate. If we find one, return the new, valid address.
3841 OLDX is the address as it was before break_out_memory_refs was called.
3842 In some cases it is useful to look at this to decide what needs to be done.
3844 MODE is the mode of the operand pointed to by X.
3846 When -fpic is used, special handling is needed for symbolic references.
3847 See comments by legitimize_pic_address for details. */
3850 s390_legitimize_address (rtx x
, rtx oldx ATTRIBUTE_UNUSED
,
3851 enum machine_mode mode ATTRIBUTE_UNUSED
)
3853 rtx constant_term
= const0_rtx
;
3855 if (TLS_SYMBOLIC_CONST (x
))
3857 x
= legitimize_tls_address (x
, 0);
3859 if (s390_legitimate_address_p (mode
, x
, FALSE
))
3862 else if (GET_CODE (x
) == PLUS
3863 && (TLS_SYMBOLIC_CONST (XEXP (x
, 0))
3864 || TLS_SYMBOLIC_CONST (XEXP (x
, 1))))
3870 if (SYMBOLIC_CONST (x
)
3871 || (GET_CODE (x
) == PLUS
3872 && (SYMBOLIC_CONST (XEXP (x
, 0))
3873 || SYMBOLIC_CONST (XEXP (x
, 1)))))
3874 x
= legitimize_pic_address (x
, 0);
3876 if (s390_legitimate_address_p (mode
, x
, FALSE
))
3880 x
= eliminate_constant_term (x
, &constant_term
);
3882 /* Optimize loading of large displacements by splitting them
3883 into the multiple of 4K and the rest; this allows the
3884 former to be CSE'd if possible.
3886 Don't do this if the displacement is added to a register
3887 pointing into the stack frame, as the offsets will
3888 change later anyway. */
3890 if (GET_CODE (constant_term
) == CONST_INT
3891 && !TARGET_LONG_DISPLACEMENT
3892 && !DISP_IN_RANGE (INTVAL (constant_term
))
3893 && !(REG_P (x
) && REGNO_PTR_FRAME_P (REGNO (x
))))
3895 HOST_WIDE_INT lower
= INTVAL (constant_term
) & 0xfff;
3896 HOST_WIDE_INT upper
= INTVAL (constant_term
) ^ lower
;
3898 rtx temp
= gen_reg_rtx (Pmode
);
3899 rtx val
= force_operand (GEN_INT (upper
), temp
);
3901 emit_move_insn (temp
, val
);
3903 x
= gen_rtx_PLUS (Pmode
, x
, temp
);
3904 constant_term
= GEN_INT (lower
);
3907 if (GET_CODE (x
) == PLUS
)
3909 if (GET_CODE (XEXP (x
, 0)) == REG
)
3911 rtx temp
= gen_reg_rtx (Pmode
);
3912 rtx val
= force_operand (XEXP (x
, 1), temp
);
3914 emit_move_insn (temp
, val
);
3916 x
= gen_rtx_PLUS (Pmode
, XEXP (x
, 0), temp
);
3919 else if (GET_CODE (XEXP (x
, 1)) == REG
)
3921 rtx temp
= gen_reg_rtx (Pmode
);
3922 rtx val
= force_operand (XEXP (x
, 0), temp
);
3924 emit_move_insn (temp
, val
);
3926 x
= gen_rtx_PLUS (Pmode
, temp
, XEXP (x
, 1));
3930 if (constant_term
!= const0_rtx
)
3931 x
= gen_rtx_PLUS (Pmode
, x
, constant_term
);
3936 /* Try a machine-dependent way of reloading an illegitimate address AD
3937 operand. If we find one, push the reload and return the new address.
3939 MODE is the mode of the enclosing MEM. OPNUM is the operand number
3940 and TYPE is the reload type of the current reload. */
3943 legitimize_reload_address (rtx ad
, enum machine_mode mode ATTRIBUTE_UNUSED
,
3944 int opnum
, int type
)
3946 if (!optimize
|| TARGET_LONG_DISPLACEMENT
)
3949 if (GET_CODE (ad
) == PLUS
)
3951 rtx tem
= simplify_binary_operation (PLUS
, Pmode
,
3952 XEXP (ad
, 0), XEXP (ad
, 1));
3957 if (GET_CODE (ad
) == PLUS
3958 && GET_CODE (XEXP (ad
, 0)) == REG
3959 && GET_CODE (XEXP (ad
, 1)) == CONST_INT
3960 && !DISP_IN_RANGE (INTVAL (XEXP (ad
, 1))))
3962 HOST_WIDE_INT lower
= INTVAL (XEXP (ad
, 1)) & 0xfff;
3963 HOST_WIDE_INT upper
= INTVAL (XEXP (ad
, 1)) ^ lower
;
3964 rtx cst
, tem
, new_rtx
;
3966 cst
= GEN_INT (upper
);
3967 if (!legitimate_reload_constant_p (cst
))
3968 cst
= force_const_mem (Pmode
, cst
);
3970 tem
= gen_rtx_PLUS (Pmode
, XEXP (ad
, 0), cst
);
3971 new_rtx
= gen_rtx_PLUS (Pmode
, tem
, GEN_INT (lower
));
3973 push_reload (XEXP (tem
, 1), 0, &XEXP (tem
, 1), 0,
3974 BASE_REG_CLASS
, Pmode
, VOIDmode
, 0, 0,
3975 opnum
, (enum reload_type
) type
);
3982 /* Emit code to move LEN bytes from DST to SRC. */
3985 s390_expand_movmem (rtx dst
, rtx src
, rtx len
)
3987 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) >= 0 && INTVAL (len
) <= 256)
3989 if (INTVAL (len
) > 0)
3990 emit_insn (gen_movmem_short (dst
, src
, GEN_INT (INTVAL (len
) - 1)));
3993 else if (TARGET_MVCLE
)
3995 emit_insn (gen_movmem_long (dst
, src
, convert_to_mode (Pmode
, len
, 1)));
4000 rtx dst_addr
, src_addr
, count
, blocks
, temp
;
4001 rtx loop_start_label
= gen_label_rtx ();
4002 rtx loop_end_label
= gen_label_rtx ();
4003 rtx end_label
= gen_label_rtx ();
4004 enum machine_mode mode
;
4006 mode
= GET_MODE (len
);
4007 if (mode
== VOIDmode
)
4010 dst_addr
= gen_reg_rtx (Pmode
);
4011 src_addr
= gen_reg_rtx (Pmode
);
4012 count
= gen_reg_rtx (mode
);
4013 blocks
= gen_reg_rtx (mode
);
4015 convert_move (count
, len
, 1);
4016 emit_cmp_and_jump_insns (count
, const0_rtx
,
4017 EQ
, NULL_RTX
, mode
, 1, end_label
);
4019 emit_move_insn (dst_addr
, force_operand (XEXP (dst
, 0), NULL_RTX
));
4020 emit_move_insn (src_addr
, force_operand (XEXP (src
, 0), NULL_RTX
));
4021 dst
= change_address (dst
, VOIDmode
, dst_addr
);
4022 src
= change_address (src
, VOIDmode
, src_addr
);
4024 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4027 emit_move_insn (count
, temp
);
4029 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4032 emit_move_insn (blocks
, temp
);
4034 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4035 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4037 emit_label (loop_start_label
);
4040 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 768))
4044 /* Issue a read prefetch for the +3 cache line. */
4045 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, src_addr
, GEN_INT (768)),
4046 const0_rtx
, const0_rtx
);
4047 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4048 emit_insn (prefetch
);
4050 /* Issue a write prefetch for the +3 cache line. */
4051 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (768)),
4052 const1_rtx
, const0_rtx
);
4053 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4054 emit_insn (prefetch
);
4057 emit_insn (gen_movmem_short (dst
, src
, GEN_INT (255)));
4058 s390_load_address (dst_addr
,
4059 gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (256)));
4060 s390_load_address (src_addr
,
4061 gen_rtx_PLUS (Pmode
, src_addr
, GEN_INT (256)));
4063 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4066 emit_move_insn (blocks
, temp
);
4068 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4069 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4071 emit_jump (loop_start_label
);
4072 emit_label (loop_end_label
);
4074 emit_insn (gen_movmem_short (dst
, src
,
4075 convert_to_mode (Pmode
, count
, 1)));
4076 emit_label (end_label
);
4080 /* Emit code to set LEN bytes at DST to VAL.
4081 Make use of clrmem if VAL is zero. */
4084 s390_expand_setmem (rtx dst
, rtx len
, rtx val
)
4086 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) == 0)
4089 gcc_assert (GET_CODE (val
) == CONST_INT
|| GET_MODE (val
) == QImode
);
4091 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) > 0 && INTVAL (len
) <= 257)
4093 if (val
== const0_rtx
&& INTVAL (len
) <= 256)
4094 emit_insn (gen_clrmem_short (dst
, GEN_INT (INTVAL (len
) - 1)));
4097 /* Initialize memory by storing the first byte. */
4098 emit_move_insn (adjust_address (dst
, QImode
, 0), val
);
4100 if (INTVAL (len
) > 1)
4102 /* Initiate 1 byte overlap move.
4103 The first byte of DST is propagated through DSTP1.
4104 Prepare a movmem for: DST+1 = DST (length = LEN - 1).
4105 DST is set to size 1 so the rest of the memory location
4106 does not count as source operand. */
4107 rtx dstp1
= adjust_address (dst
, VOIDmode
, 1);
4108 set_mem_size (dst
, 1);
4110 emit_insn (gen_movmem_short (dstp1
, dst
,
4111 GEN_INT (INTVAL (len
) - 2)));
4116 else if (TARGET_MVCLE
)
4118 val
= force_not_mem (convert_modes (Pmode
, QImode
, val
, 1));
4119 emit_insn (gen_setmem_long (dst
, convert_to_mode (Pmode
, len
, 1), val
));
4124 rtx dst_addr
, count
, blocks
, temp
, dstp1
= NULL_RTX
;
4125 rtx loop_start_label
= gen_label_rtx ();
4126 rtx loop_end_label
= gen_label_rtx ();
4127 rtx end_label
= gen_label_rtx ();
4128 enum machine_mode mode
;
4130 mode
= GET_MODE (len
);
4131 if (mode
== VOIDmode
)
4134 dst_addr
= gen_reg_rtx (Pmode
);
4135 count
= gen_reg_rtx (mode
);
4136 blocks
= gen_reg_rtx (mode
);
4138 convert_move (count
, len
, 1);
4139 emit_cmp_and_jump_insns (count
, const0_rtx
,
4140 EQ
, NULL_RTX
, mode
, 1, end_label
);
4142 emit_move_insn (dst_addr
, force_operand (XEXP (dst
, 0), NULL_RTX
));
4143 dst
= change_address (dst
, VOIDmode
, dst_addr
);
4145 if (val
== const0_rtx
)
4146 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4150 dstp1
= adjust_address (dst
, VOIDmode
, 1);
4151 set_mem_size (dst
, 1);
4153 /* Initialize memory by storing the first byte. */
4154 emit_move_insn (adjust_address (dst
, QImode
, 0), val
);
4156 /* If count is 1 we are done. */
4157 emit_cmp_and_jump_insns (count
, const1_rtx
,
4158 EQ
, NULL_RTX
, mode
, 1, end_label
);
4160 temp
= expand_binop (mode
, add_optab
, count
, GEN_INT (-2), count
, 1,
4164 emit_move_insn (count
, temp
);
4166 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4169 emit_move_insn (blocks
, temp
);
4171 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4172 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4174 emit_label (loop_start_label
);
4177 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 1024))
4179 /* Issue a write prefetch for the +4 cache line. */
4180 rtx prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, dst_addr
,
4182 const1_rtx
, const0_rtx
);
4183 emit_insn (prefetch
);
4184 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4187 if (val
== const0_rtx
)
4188 emit_insn (gen_clrmem_short (dst
, GEN_INT (255)));
4190 emit_insn (gen_movmem_short (dstp1
, dst
, GEN_INT (255)));
4191 s390_load_address (dst_addr
,
4192 gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (256)));
4194 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4197 emit_move_insn (blocks
, temp
);
4199 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4200 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4202 emit_jump (loop_start_label
);
4203 emit_label (loop_end_label
);
4205 if (val
== const0_rtx
)
4206 emit_insn (gen_clrmem_short (dst
, convert_to_mode (Pmode
, count
, 1)));
4208 emit_insn (gen_movmem_short (dstp1
, dst
, convert_to_mode (Pmode
, count
, 1)));
4209 emit_label (end_label
);
4213 /* Emit code to compare LEN bytes at OP0 with those at OP1,
4214 and return the result in TARGET. */
4217 s390_expand_cmpmem (rtx target
, rtx op0
, rtx op1
, rtx len
)
4219 rtx ccreg
= gen_rtx_REG (CCUmode
, CC_REGNUM
);
4222 /* As the result of CMPINT is inverted compared to what we need,
4223 we have to swap the operands. */
4224 tmp
= op0
; op0
= op1
; op1
= tmp
;
4226 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) >= 0 && INTVAL (len
) <= 256)
4228 if (INTVAL (len
) > 0)
4230 emit_insn (gen_cmpmem_short (op0
, op1
, GEN_INT (INTVAL (len
) - 1)));
4231 emit_insn (gen_cmpint (target
, ccreg
));
4234 emit_move_insn (target
, const0_rtx
);
4236 else if (TARGET_MVCLE
)
4238 emit_insn (gen_cmpmem_long (op0
, op1
, convert_to_mode (Pmode
, len
, 1)));
4239 emit_insn (gen_cmpint (target
, ccreg
));
4243 rtx addr0
, addr1
, count
, blocks
, temp
;
4244 rtx loop_start_label
= gen_label_rtx ();
4245 rtx loop_end_label
= gen_label_rtx ();
4246 rtx end_label
= gen_label_rtx ();
4247 enum machine_mode mode
;
4249 mode
= GET_MODE (len
);
4250 if (mode
== VOIDmode
)
4253 addr0
= gen_reg_rtx (Pmode
);
4254 addr1
= gen_reg_rtx (Pmode
);
4255 count
= gen_reg_rtx (mode
);
4256 blocks
= gen_reg_rtx (mode
);
4258 convert_move (count
, len
, 1);
4259 emit_cmp_and_jump_insns (count
, const0_rtx
,
4260 EQ
, NULL_RTX
, mode
, 1, end_label
);
4262 emit_move_insn (addr0
, force_operand (XEXP (op0
, 0), NULL_RTX
));
4263 emit_move_insn (addr1
, force_operand (XEXP (op1
, 0), NULL_RTX
));
4264 op0
= change_address (op0
, VOIDmode
, addr0
);
4265 op1
= change_address (op1
, VOIDmode
, addr1
);
4267 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4270 emit_move_insn (count
, temp
);
4272 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4275 emit_move_insn (blocks
, temp
);
4277 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4278 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4280 emit_label (loop_start_label
);
4283 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 512))
4287 /* Issue a read prefetch for the +2 cache line of operand 1. */
4288 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, addr0
, GEN_INT (512)),
4289 const0_rtx
, const0_rtx
);
4290 emit_insn (prefetch
);
4291 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4293 /* Issue a read prefetch for the +2 cache line of operand 2. */
4294 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, addr1
, GEN_INT (512)),
4295 const0_rtx
, const0_rtx
);
4296 emit_insn (prefetch
);
4297 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4300 emit_insn (gen_cmpmem_short (op0
, op1
, GEN_INT (255)));
4301 temp
= gen_rtx_NE (VOIDmode
, ccreg
, const0_rtx
);
4302 temp
= gen_rtx_IF_THEN_ELSE (VOIDmode
, temp
,
4303 gen_rtx_LABEL_REF (VOIDmode
, end_label
), pc_rtx
);
4304 temp
= gen_rtx_SET (VOIDmode
, pc_rtx
, temp
);
4305 emit_jump_insn (temp
);
4307 s390_load_address (addr0
,
4308 gen_rtx_PLUS (Pmode
, addr0
, GEN_INT (256)));
4309 s390_load_address (addr1
,
4310 gen_rtx_PLUS (Pmode
, addr1
, GEN_INT (256)));
4312 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4315 emit_move_insn (blocks
, temp
);
4317 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4318 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4320 emit_jump (loop_start_label
);
4321 emit_label (loop_end_label
);
4323 emit_insn (gen_cmpmem_short (op0
, op1
,
4324 convert_to_mode (Pmode
, count
, 1)));
4325 emit_label (end_label
);
4327 emit_insn (gen_cmpint (target
, ccreg
));
4332 /* Expand conditional increment or decrement using alc/slb instructions.
4333 Should generate code setting DST to either SRC or SRC + INCREMENT,
4334 depending on the result of the comparison CMP_OP0 CMP_CODE CMP_OP1.
4335 Returns true if successful, false otherwise.
4337 That makes it possible to implement some if-constructs without jumps e.g.:
4338 (borrow = CC0 | CC1 and carry = CC2 | CC3)
4339 unsigned int a, b, c;
4340 if (a < b) c++; -> CCU b > a -> CC2; c += carry;
4341 if (a < b) c--; -> CCL3 a - b -> borrow; c -= borrow;
4342 if (a <= b) c++; -> CCL3 b - a -> borrow; c += carry;
4343 if (a <= b) c--; -> CCU a <= b -> borrow; c -= borrow;
4345 Checks for EQ and NE with a nonzero value need an additional xor e.g.:
4346 if (a == b) c++; -> CCL3 a ^= b; 0 - a -> borrow; c += carry;
4347 if (a == b) c--; -> CCU a ^= b; a <= 0 -> CC0 | CC1; c -= borrow;
4348 if (a != b) c++; -> CCU a ^= b; a > 0 -> CC2; c += carry;
4349 if (a != b) c--; -> CCL3 a ^= b; 0 - a -> borrow; c -= borrow; */
4352 s390_expand_addcc (enum rtx_code cmp_code
, rtx cmp_op0
, rtx cmp_op1
,
4353 rtx dst
, rtx src
, rtx increment
)
4355 enum machine_mode cmp_mode
;
4356 enum machine_mode cc_mode
;
4362 if ((GET_MODE (cmp_op0
) == SImode
|| GET_MODE (cmp_op0
) == VOIDmode
)
4363 && (GET_MODE (cmp_op1
) == SImode
|| GET_MODE (cmp_op1
) == VOIDmode
))
4365 else if ((GET_MODE (cmp_op0
) == DImode
|| GET_MODE (cmp_op0
) == VOIDmode
)
4366 && (GET_MODE (cmp_op1
) == DImode
|| GET_MODE (cmp_op1
) == VOIDmode
))
4371 /* Try ADD LOGICAL WITH CARRY. */
4372 if (increment
== const1_rtx
)
4374 /* Determine CC mode to use. */
4375 if (cmp_code
== EQ
|| cmp_code
== NE
)
4377 if (cmp_op1
!= const0_rtx
)
4379 cmp_op0
= expand_simple_binop (cmp_mode
, XOR
, cmp_op0
, cmp_op1
,
4380 NULL_RTX
, 0, OPTAB_WIDEN
);
4381 cmp_op1
= const0_rtx
;
4384 cmp_code
= cmp_code
== EQ
? LEU
: GTU
;
4387 if (cmp_code
== LTU
|| cmp_code
== LEU
)
4392 cmp_code
= swap_condition (cmp_code
);
4409 /* Emit comparison instruction pattern. */
4410 if (!register_operand (cmp_op0
, cmp_mode
))
4411 cmp_op0
= force_reg (cmp_mode
, cmp_op0
);
4413 insn
= gen_rtx_SET (VOIDmode
, gen_rtx_REG (cc_mode
, CC_REGNUM
),
4414 gen_rtx_COMPARE (cc_mode
, cmp_op0
, cmp_op1
));
4415 /* We use insn_invalid_p here to add clobbers if required. */
4416 ret
= insn_invalid_p (emit_insn (insn
), false);
4419 /* Emit ALC instruction pattern. */
4420 op_res
= gen_rtx_fmt_ee (cmp_code
, GET_MODE (dst
),
4421 gen_rtx_REG (cc_mode
, CC_REGNUM
),
4424 if (src
!= const0_rtx
)
4426 if (!register_operand (src
, GET_MODE (dst
)))
4427 src
= force_reg (GET_MODE (dst
), src
);
4429 op_res
= gen_rtx_PLUS (GET_MODE (dst
), op_res
, src
);
4430 op_res
= gen_rtx_PLUS (GET_MODE (dst
), op_res
, const0_rtx
);
4433 p
= rtvec_alloc (2);
4435 gen_rtx_SET (VOIDmode
, dst
, op_res
);
4437 gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4438 emit_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
4443 /* Try SUBTRACT LOGICAL WITH BORROW. */
4444 if (increment
== constm1_rtx
)
4446 /* Determine CC mode to use. */
4447 if (cmp_code
== EQ
|| cmp_code
== NE
)
4449 if (cmp_op1
!= const0_rtx
)
4451 cmp_op0
= expand_simple_binop (cmp_mode
, XOR
, cmp_op0
, cmp_op1
,
4452 NULL_RTX
, 0, OPTAB_WIDEN
);
4453 cmp_op1
= const0_rtx
;
4456 cmp_code
= cmp_code
== EQ
? LEU
: GTU
;
4459 if (cmp_code
== GTU
|| cmp_code
== GEU
)
4464 cmp_code
= swap_condition (cmp_code
);
4481 /* Emit comparison instruction pattern. */
4482 if (!register_operand (cmp_op0
, cmp_mode
))
4483 cmp_op0
= force_reg (cmp_mode
, cmp_op0
);
4485 insn
= gen_rtx_SET (VOIDmode
, gen_rtx_REG (cc_mode
, CC_REGNUM
),
4486 gen_rtx_COMPARE (cc_mode
, cmp_op0
, cmp_op1
));
4487 /* We use insn_invalid_p here to add clobbers if required. */
4488 ret
= insn_invalid_p (emit_insn (insn
), false);
4491 /* Emit SLB instruction pattern. */
4492 if (!register_operand (src
, GET_MODE (dst
)))
4493 src
= force_reg (GET_MODE (dst
), src
);
4495 op_res
= gen_rtx_MINUS (GET_MODE (dst
),
4496 gen_rtx_MINUS (GET_MODE (dst
), src
, const0_rtx
),
4497 gen_rtx_fmt_ee (cmp_code
, GET_MODE (dst
),
4498 gen_rtx_REG (cc_mode
, CC_REGNUM
),
4500 p
= rtvec_alloc (2);
4502 gen_rtx_SET (VOIDmode
, dst
, op_res
);
4504 gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4505 emit_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
4513 /* Expand code for the insv template. Return true if successful. */
4516 s390_expand_insv (rtx dest
, rtx op1
, rtx op2
, rtx src
)
4518 int bitsize
= INTVAL (op1
);
4519 int bitpos
= INTVAL (op2
);
4521 /* On z10 we can use the risbg instruction to implement insv. */
4523 && ((GET_MODE (dest
) == DImode
&& GET_MODE (src
) == DImode
)
4524 || (GET_MODE (dest
) == SImode
&& GET_MODE (src
) == SImode
)))
4529 op
= gen_rtx_SET (GET_MODE(src
),
4530 gen_rtx_ZERO_EXTRACT (GET_MODE (dest
), dest
, op1
, op2
),
4532 clobber
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4533 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, op
, clobber
)));
4538 /* We need byte alignment. */
4539 if (bitsize
% BITS_PER_UNIT
)
4543 && memory_operand (dest
, VOIDmode
)
4544 && (register_operand (src
, word_mode
)
4545 || const_int_operand (src
, VOIDmode
)))
4547 /* Emit standard pattern if possible. */
4548 enum machine_mode mode
= smallest_mode_for_size (bitsize
, MODE_INT
);
4549 if (GET_MODE_BITSIZE (mode
) == bitsize
)
4550 emit_move_insn (adjust_address (dest
, mode
, 0), gen_lowpart (mode
, src
));
4552 /* (set (ze (mem)) (const_int)). */
4553 else if (const_int_operand (src
, VOIDmode
))
4555 int size
= bitsize
/ BITS_PER_UNIT
;
4556 rtx src_mem
= adjust_address (force_const_mem (word_mode
, src
), BLKmode
,
4557 GET_MODE_SIZE (word_mode
) - size
);
4559 dest
= adjust_address (dest
, BLKmode
, 0);
4560 set_mem_size (dest
, size
);
4561 s390_expand_movmem (dest
, src_mem
, GEN_INT (size
));
4564 /* (set (ze (mem)) (reg)). */
4565 else if (register_operand (src
, word_mode
))
4567 if (bitsize
<= GET_MODE_BITSIZE (SImode
))
4568 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
, op1
,
4572 /* Emit st,stcmh sequence. */
4573 int stcmh_width
= bitsize
- GET_MODE_BITSIZE (SImode
);
4574 int size
= stcmh_width
/ BITS_PER_UNIT
;
4576 emit_move_insn (adjust_address (dest
, SImode
, size
),
4577 gen_lowpart (SImode
, src
));
4578 set_mem_size (dest
, size
);
4579 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
, GEN_INT
4580 (stcmh_width
), const0_rtx
),
4581 gen_rtx_LSHIFTRT (word_mode
, src
, GEN_INT
4582 (GET_MODE_BITSIZE (SImode
))));
4591 /* (set (ze (reg)) (const_int)). */
4593 && register_operand (dest
, word_mode
)
4594 && (bitpos
% 16) == 0
4595 && (bitsize
% 16) == 0
4596 && const_int_operand (src
, VOIDmode
))
4598 HOST_WIDE_INT val
= INTVAL (src
);
4599 int regpos
= bitpos
+ bitsize
;
4601 while (regpos
> bitpos
)
4603 enum machine_mode putmode
;
4606 if (TARGET_EXTIMM
&& (regpos
% 32 == 0) && (regpos
>= bitpos
+ 32))
4611 putsize
= GET_MODE_BITSIZE (putmode
);
4613 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
,
4616 gen_int_mode (val
, putmode
));
4619 gcc_assert (regpos
== bitpos
);
4626 /* A subroutine of s390_expand_cs_hqi and s390_expand_atomic which returns a
4627 register that holds VAL of mode MODE shifted by COUNT bits. */
4630 s390_expand_mask_and_shift (rtx val
, enum machine_mode mode
, rtx count
)
4632 val
= expand_simple_binop (SImode
, AND
, val
, GEN_INT (GET_MODE_MASK (mode
)),
4633 NULL_RTX
, 1, OPTAB_DIRECT
);
4634 return expand_simple_binop (SImode
, ASHIFT
, val
, count
,
4635 NULL_RTX
, 1, OPTAB_DIRECT
);
4638 /* Structure to hold the initial parameters for a compare_and_swap operation
4639 in HImode and QImode. */
4641 struct alignment_context
4643 rtx memsi
; /* SI aligned memory location. */
4644 rtx shift
; /* Bit offset with regard to lsb. */
4645 rtx modemask
; /* Mask of the HQImode shifted by SHIFT bits. */
4646 rtx modemaski
; /* ~modemask */
4647 bool aligned
; /* True if memory is aligned, false else. */
4650 /* A subroutine of s390_expand_cs_hqi and s390_expand_atomic to initialize
4651 structure AC for transparent simplifying, if the memory alignment is known
4652 to be at least 32bit. MEM is the memory location for the actual operation
4653 and MODE its mode. */
4656 init_alignment_context (struct alignment_context
*ac
, rtx mem
,
4657 enum machine_mode mode
)
4659 ac
->shift
= GEN_INT (GET_MODE_SIZE (SImode
) - GET_MODE_SIZE (mode
));
4660 ac
->aligned
= (MEM_ALIGN (mem
) >= GET_MODE_BITSIZE (SImode
));
4663 ac
->memsi
= adjust_address (mem
, SImode
, 0); /* Memory is aligned. */
4666 /* Alignment is unknown. */
4667 rtx byteoffset
, addr
, align
;
4669 /* Force the address into a register. */
4670 addr
= force_reg (Pmode
, XEXP (mem
, 0));
4672 /* Align it to SImode. */
4673 align
= expand_simple_binop (Pmode
, AND
, addr
,
4674 GEN_INT (-GET_MODE_SIZE (SImode
)),
4675 NULL_RTX
, 1, OPTAB_DIRECT
);
4677 ac
->memsi
= gen_rtx_MEM (SImode
, align
);
4678 MEM_VOLATILE_P (ac
->memsi
) = MEM_VOLATILE_P (mem
);
4679 set_mem_alias_set (ac
->memsi
, ALIAS_SET_MEMORY_BARRIER
);
4680 set_mem_align (ac
->memsi
, GET_MODE_BITSIZE (SImode
));
4682 /* Calculate shiftcount. */
4683 byteoffset
= expand_simple_binop (Pmode
, AND
, addr
,
4684 GEN_INT (GET_MODE_SIZE (SImode
) - 1),
4685 NULL_RTX
, 1, OPTAB_DIRECT
);
4686 /* As we already have some offset, evaluate the remaining distance. */
4687 ac
->shift
= expand_simple_binop (SImode
, MINUS
, ac
->shift
, byteoffset
,
4688 NULL_RTX
, 1, OPTAB_DIRECT
);
4691 /* Shift is the byte count, but we need the bitcount. */
4692 ac
->shift
= expand_simple_binop (SImode
, MULT
, ac
->shift
, GEN_INT (BITS_PER_UNIT
),
4693 NULL_RTX
, 1, OPTAB_DIRECT
);
4694 /* Calculate masks. */
4695 ac
->modemask
= expand_simple_binop (SImode
, ASHIFT
,
4696 GEN_INT (GET_MODE_MASK (mode
)), ac
->shift
,
4697 NULL_RTX
, 1, OPTAB_DIRECT
);
4698 ac
->modemaski
= expand_simple_unop (SImode
, NOT
, ac
->modemask
, NULL_RTX
, 1);
4701 /* Expand an atomic compare and swap operation for HImode and QImode. MEM is
4702 the memory location, CMP the old value to compare MEM with and NEW_RTX the value
4703 to set if CMP == MEM.
4704 CMP is never in memory for compare_and_swap_cc because
4705 expand_bool_compare_and_swap puts it into a register for later compare. */
4708 s390_expand_cs_hqi (enum machine_mode mode
, rtx target
, rtx mem
, rtx cmp
, rtx new_rtx
)
4710 struct alignment_context ac
;
4711 rtx cmpv
, newv
, val
, resv
, cc
;
4712 rtx res
= gen_reg_rtx (SImode
);
4713 rtx csloop
= gen_label_rtx ();
4714 rtx csend
= gen_label_rtx ();
4716 gcc_assert (register_operand (target
, VOIDmode
));
4717 gcc_assert (MEM_P (mem
));
4719 init_alignment_context (&ac
, mem
, mode
);
4721 /* Shift the values to the correct bit positions. */
4722 if (!(ac
.aligned
&& MEM_P (cmp
)))
4723 cmp
= s390_expand_mask_and_shift (cmp
, mode
, ac
.shift
);
4724 if (!(ac
.aligned
&& MEM_P (new_rtx
)))
4725 new_rtx
= s390_expand_mask_and_shift (new_rtx
, mode
, ac
.shift
);
4727 /* Load full word. Subsequent loads are performed by CS. */
4728 val
= expand_simple_binop (SImode
, AND
, ac
.memsi
, ac
.modemaski
,
4729 NULL_RTX
, 1, OPTAB_DIRECT
);
4731 /* Start CS loop. */
4732 emit_label (csloop
);
4733 /* val = "<mem>00..0<mem>"
4734 * cmp = "00..0<cmp>00..0"
4735 * new = "00..0<new>00..0"
4738 /* Patch cmp and new with val at correct position. */
4739 if (ac
.aligned
&& MEM_P (cmp
))
4741 cmpv
= force_reg (SImode
, val
);
4742 store_bit_field (cmpv
, GET_MODE_BITSIZE (mode
), 0,
4746 cmpv
= force_reg (SImode
, expand_simple_binop (SImode
, IOR
, cmp
, val
,
4747 NULL_RTX
, 1, OPTAB_DIRECT
));
4748 if (ac
.aligned
&& MEM_P (new_rtx
))
4750 newv
= force_reg (SImode
, val
);
4751 store_bit_field (newv
, GET_MODE_BITSIZE (mode
), 0,
4752 0, 0, SImode
, new_rtx
);
4755 newv
= force_reg (SImode
, expand_simple_binop (SImode
, IOR
, new_rtx
, val
,
4756 NULL_RTX
, 1, OPTAB_DIRECT
));
4758 /* Jump to end if we're done (likely?). */
4759 s390_emit_jump (csend
, s390_emit_compare_and_swap (EQ
, res
, ac
.memsi
,
4762 /* Check for changes outside mode. */
4763 resv
= expand_simple_binop (SImode
, AND
, res
, ac
.modemaski
,
4764 NULL_RTX
, 1, OPTAB_DIRECT
);
4765 cc
= s390_emit_compare (NE
, resv
, val
);
4766 emit_move_insn (val
, resv
);
4767 /* Loop internal if so. */
4768 s390_emit_jump (csloop
, cc
);
4772 /* Return the correct part of the bitfield. */
4773 convert_move (target
, expand_simple_binop (SImode
, LSHIFTRT
, res
, ac
.shift
,
4774 NULL_RTX
, 1, OPTAB_DIRECT
), 1);
4777 /* Expand an atomic operation CODE of mode MODE. MEM is the memory location
4778 and VAL the value to play with. If AFTER is true then store the value
4779 MEM holds after the operation, if AFTER is false then store the value MEM
4780 holds before the operation. If TARGET is zero then discard that value, else
4781 store it to TARGET. */
4784 s390_expand_atomic (enum machine_mode mode
, enum rtx_code code
,
4785 rtx target
, rtx mem
, rtx val
, bool after
)
4787 struct alignment_context ac
;
4789 rtx new_rtx
= gen_reg_rtx (SImode
);
4790 rtx orig
= gen_reg_rtx (SImode
);
4791 rtx csloop
= gen_label_rtx ();
4793 gcc_assert (!target
|| register_operand (target
, VOIDmode
));
4794 gcc_assert (MEM_P (mem
));
4796 init_alignment_context (&ac
, mem
, mode
);
4798 /* Shift val to the correct bit positions.
4799 Preserve "icm", but prevent "ex icm". */
4800 if (!(ac
.aligned
&& code
== SET
&& MEM_P (val
)))
4801 val
= s390_expand_mask_and_shift (val
, mode
, ac
.shift
);
4803 /* Further preparation insns. */
4804 if (code
== PLUS
|| code
== MINUS
)
4805 emit_move_insn (orig
, val
);
4806 else if (code
== MULT
|| code
== AND
) /* val = "11..1<val>11..1" */
4807 val
= expand_simple_binop (SImode
, XOR
, val
, ac
.modemaski
,
4808 NULL_RTX
, 1, OPTAB_DIRECT
);
4810 /* Load full word. Subsequent loads are performed by CS. */
4811 cmp
= force_reg (SImode
, ac
.memsi
);
4813 /* Start CS loop. */
4814 emit_label (csloop
);
4815 emit_move_insn (new_rtx
, cmp
);
4817 /* Patch new with val at correct position. */
4822 val
= expand_simple_binop (SImode
, code
, new_rtx
, orig
,
4823 NULL_RTX
, 1, OPTAB_DIRECT
);
4824 val
= expand_simple_binop (SImode
, AND
, val
, ac
.modemask
,
4825 NULL_RTX
, 1, OPTAB_DIRECT
);
4828 if (ac
.aligned
&& MEM_P (val
))
4829 store_bit_field (new_rtx
, GET_MODE_BITSIZE (mode
), 0,
4833 new_rtx
= expand_simple_binop (SImode
, AND
, new_rtx
, ac
.modemaski
,
4834 NULL_RTX
, 1, OPTAB_DIRECT
);
4835 new_rtx
= expand_simple_binop (SImode
, IOR
, new_rtx
, val
,
4836 NULL_RTX
, 1, OPTAB_DIRECT
);
4842 new_rtx
= expand_simple_binop (SImode
, code
, new_rtx
, val
,
4843 NULL_RTX
, 1, OPTAB_DIRECT
);
4845 case MULT
: /* NAND */
4846 new_rtx
= expand_simple_binop (SImode
, AND
, new_rtx
, val
,
4847 NULL_RTX
, 1, OPTAB_DIRECT
);
4848 new_rtx
= expand_simple_binop (SImode
, XOR
, new_rtx
, ac
.modemask
,
4849 NULL_RTX
, 1, OPTAB_DIRECT
);
4855 s390_emit_jump (csloop
, s390_emit_compare_and_swap (NE
, cmp
,
4856 ac
.memsi
, cmp
, new_rtx
));
4858 /* Return the correct part of the bitfield. */
4860 convert_move (target
, expand_simple_binop (SImode
, LSHIFTRT
,
4861 after
? new_rtx
: cmp
, ac
.shift
,
4862 NULL_RTX
, 1, OPTAB_DIRECT
), 1);
4865 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
4866 We need to emit DTP-relative relocations. */
4868 static void s390_output_dwarf_dtprel (FILE *, int, rtx
) ATTRIBUTE_UNUSED
;
4871 s390_output_dwarf_dtprel (FILE *file
, int size
, rtx x
)
4876 fputs ("\t.long\t", file
);
4879 fputs ("\t.quad\t", file
);
4884 output_addr_const (file
, x
);
4885 fputs ("@DTPOFF", file
);
4888 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
4889 /* Implement TARGET_MANGLE_TYPE. */
4892 s390_mangle_type (const_tree type
)
4894 if (TYPE_MAIN_VARIANT (type
) == long_double_type_node
4895 && TARGET_LONG_DOUBLE_128
)
4898 /* For all other types, use normal C++ mangling. */
4903 /* In the name of slightly smaller debug output, and to cater to
4904 general assembler lossage, recognize various UNSPEC sequences
4905 and turn them back into a direct symbol reference. */
4908 s390_delegitimize_address (rtx orig_x
)
4912 orig_x
= delegitimize_mem_from_attrs (orig_x
);
4915 /* Extract the symbol ref from:
4916 (plus:SI (reg:SI 12 %r12)
4917 (const:SI (unspec:SI [(symbol_ref/f:SI ("*.LC0"))]
4918 UNSPEC_GOTOFF/PLTOFF)))
4920 (plus:SI (reg:SI 12 %r12)
4921 (const:SI (plus:SI (unspec:SI [(symbol_ref:SI ("L"))]
4922 UNSPEC_GOTOFF/PLTOFF)
4923 (const_int 4 [0x4])))) */
4924 if (GET_CODE (x
) == PLUS
4925 && REG_P (XEXP (x
, 0))
4926 && REGNO (XEXP (x
, 0)) == PIC_OFFSET_TABLE_REGNUM
4927 && GET_CODE (XEXP (x
, 1)) == CONST
)
4929 HOST_WIDE_INT offset
= 0;
4931 /* The const operand. */
4932 y
= XEXP (XEXP (x
, 1), 0);
4934 if (GET_CODE (y
) == PLUS
4935 && GET_CODE (XEXP (y
, 1)) == CONST_INT
)
4937 offset
= INTVAL (XEXP (y
, 1));
4941 if (GET_CODE (y
) == UNSPEC
4942 && (XINT (y
, 1) == UNSPEC_GOTOFF
4943 || XINT (y
, 1) == UNSPEC_PLTOFF
))
4944 return plus_constant (Pmode
, XVECEXP (y
, 0, 0), offset
);
4947 if (GET_CODE (x
) != MEM
)
4951 if (GET_CODE (x
) == PLUS
4952 && GET_CODE (XEXP (x
, 1)) == CONST
4953 && GET_CODE (XEXP (x
, 0)) == REG
4954 && REGNO (XEXP (x
, 0)) == PIC_OFFSET_TABLE_REGNUM
)
4956 y
= XEXP (XEXP (x
, 1), 0);
4957 if (GET_CODE (y
) == UNSPEC
4958 && XINT (y
, 1) == UNSPEC_GOT
)
4959 y
= XVECEXP (y
, 0, 0);
4963 else if (GET_CODE (x
) == CONST
)
4965 /* Extract the symbol ref from:
4966 (mem:QI (const:DI (unspec:DI [(symbol_ref:DI ("foo"))]
4967 UNSPEC_PLT/GOTENT))) */
4970 if (GET_CODE (y
) == UNSPEC
4971 && (XINT (y
, 1) == UNSPEC_GOTENT
4972 || XINT (y
, 1) == UNSPEC_PLT
))
4973 y
= XVECEXP (y
, 0, 0);
4980 if (GET_MODE (orig_x
) != Pmode
)
4982 if (GET_MODE (orig_x
) == BLKmode
)
4984 y
= lowpart_subreg (GET_MODE (orig_x
), y
, Pmode
);
4991 /* Output operand OP to stdio stream FILE.
4992 OP is an address (register + offset) which is not used to address data;
4993 instead the rightmost bits are interpreted as the value. */
4996 print_shift_count_operand (FILE *file
, rtx op
)
4998 HOST_WIDE_INT offset
;
5001 /* Extract base register and offset. */
5002 if (!s390_decompose_shift_count (op
, &base
, &offset
))
5008 gcc_assert (GET_CODE (base
) == REG
);
5009 gcc_assert (REGNO (base
) < FIRST_PSEUDO_REGISTER
);
5010 gcc_assert (REGNO_REG_CLASS (REGNO (base
)) == ADDR_REGS
);
5013 /* Offsets are constricted to twelve bits. */
5014 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, offset
& ((1 << 12) - 1));
5016 fprintf (file
, "(%s)", reg_names
[REGNO (base
)]);
5019 /* See 'get_some_local_dynamic_name'. */
5022 get_some_local_dynamic_name_1 (rtx
*px
, void *data ATTRIBUTE_UNUSED
)
5026 if (GET_CODE (x
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (x
))
5028 x
= get_pool_constant (x
);
5029 return for_each_rtx (&x
, get_some_local_dynamic_name_1
, 0);
5032 if (GET_CODE (x
) == SYMBOL_REF
5033 && tls_symbolic_operand (x
) == TLS_MODEL_LOCAL_DYNAMIC
)
5035 cfun
->machine
->some_ld_name
= XSTR (x
, 0);
5042 /* Locate some local-dynamic symbol still in use by this function
5043 so that we can print its name in local-dynamic base patterns. */
5046 get_some_local_dynamic_name (void)
5050 if (cfun
->machine
->some_ld_name
)
5051 return cfun
->machine
->some_ld_name
;
5053 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
5055 && for_each_rtx (&PATTERN (insn
), get_some_local_dynamic_name_1
, 0))
5056 return cfun
->machine
->some_ld_name
;
5061 /* Output machine-dependent UNSPECs occurring in address constant X
5062 in assembler syntax to stdio stream FILE. Returns true if the
5063 constant X could be recognized, false otherwise. */
5066 s390_output_addr_const_extra (FILE *file
, rtx x
)
5068 if (GET_CODE (x
) == UNSPEC
&& XVECLEN (x
, 0) == 1)
5069 switch (XINT (x
, 1))
5072 output_addr_const (file
, XVECEXP (x
, 0, 0));
5073 fprintf (file
, "@GOTENT");
5076 output_addr_const (file
, XVECEXP (x
, 0, 0));
5077 fprintf (file
, "@GOT");
5080 output_addr_const (file
, XVECEXP (x
, 0, 0));
5081 fprintf (file
, "@GOTOFF");
5084 output_addr_const (file
, XVECEXP (x
, 0, 0));
5085 fprintf (file
, "@PLT");
5088 output_addr_const (file
, XVECEXP (x
, 0, 0));
5089 fprintf (file
, "@PLTOFF");
5092 output_addr_const (file
, XVECEXP (x
, 0, 0));
5093 fprintf (file
, "@TLSGD");
5096 assemble_name (file
, get_some_local_dynamic_name ());
5097 fprintf (file
, "@TLSLDM");
5100 output_addr_const (file
, XVECEXP (x
, 0, 0));
5101 fprintf (file
, "@DTPOFF");
5104 output_addr_const (file
, XVECEXP (x
, 0, 0));
5105 fprintf (file
, "@NTPOFF");
5107 case UNSPEC_GOTNTPOFF
:
5108 output_addr_const (file
, XVECEXP (x
, 0, 0));
5109 fprintf (file
, "@GOTNTPOFF");
5111 case UNSPEC_INDNTPOFF
:
5112 output_addr_const (file
, XVECEXP (x
, 0, 0));
5113 fprintf (file
, "@INDNTPOFF");
5117 if (GET_CODE (x
) == UNSPEC
&& XVECLEN (x
, 0) == 2)
5118 switch (XINT (x
, 1))
5120 case UNSPEC_POOL_OFFSET
:
5121 x
= gen_rtx_MINUS (GET_MODE (x
), XVECEXP (x
, 0, 0), XVECEXP (x
, 0, 1));
5122 output_addr_const (file
, x
);
5128 /* Output address operand ADDR in assembler syntax to
5129 stdio stream FILE. */
5132 print_operand_address (FILE *file
, rtx addr
)
5134 struct s390_address ad
;
5136 if (s390_symref_operand_p (addr
, NULL
, NULL
))
5140 output_operand_lossage ("symbolic memory references are "
5141 "only supported on z10 or later");
5144 output_addr_const (file
, addr
);
5148 if (!s390_decompose_address (addr
, &ad
)
5149 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5150 || (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
))))
5151 output_operand_lossage ("cannot decompose address");
5154 output_addr_const (file
, ad
.disp
);
5156 fprintf (file
, "0");
5158 if (ad
.base
&& ad
.indx
)
5159 fprintf (file
, "(%s,%s)", reg_names
[REGNO (ad
.indx
)],
5160 reg_names
[REGNO (ad
.base
)]);
5162 fprintf (file
, "(%s)", reg_names
[REGNO (ad
.base
)]);
5165 /* Output operand X in assembler syntax to stdio stream FILE.
5166 CODE specified the format flag. The following format flags
5169 'C': print opcode suffix for branch condition.
5170 'D': print opcode suffix for inverse branch condition.
5171 'E': print opcode suffix for branch on index instruction.
5172 'J': print tls_load/tls_gdcall/tls_ldcall suffix
5173 'G': print the size of the operand in bytes.
5174 'O': print only the displacement of a memory reference.
5175 'R': print only the base register of a memory reference.
5176 'S': print S-type memory reference (base+displacement).
5177 'N': print the second word of a DImode operand.
5178 'M': print the second word of a TImode operand.
5179 'Y': print shift count operand.
5181 'b': print integer X as if it's an unsigned byte.
5182 'c': print integer X as if it's an signed byte.
5183 'x': print integer X as if it's an unsigned halfword.
5184 'h': print integer X as if it's a signed halfword.
5185 'i': print the first nonzero HImode part of X.
5186 'j': print the first HImode part unequal to -1 of X.
5187 'k': print the first nonzero SImode part of X.
5188 'm': print the first SImode part unequal to -1 of X.
5189 'o': print integer X as if it's an unsigned 32bit word. */
5192 print_operand (FILE *file
, rtx x
, int code
)
5197 fprintf (file
, s390_branch_condition_mnemonic (x
, FALSE
));
5201 fprintf (file
, s390_branch_condition_mnemonic (x
, TRUE
));
5205 if (GET_CODE (x
) == LE
)
5206 fprintf (file
, "l");
5207 else if (GET_CODE (x
) == GT
)
5208 fprintf (file
, "h");
5210 output_operand_lossage ("invalid comparison operator "
5211 "for 'E' output modifier");
5215 if (GET_CODE (x
) == SYMBOL_REF
)
5217 fprintf (file
, "%s", ":tls_load:");
5218 output_addr_const (file
, x
);
5220 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLSGD
)
5222 fprintf (file
, "%s", ":tls_gdcall:");
5223 output_addr_const (file
, XVECEXP (x
, 0, 0));
5225 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLSLDM
)
5227 fprintf (file
, "%s", ":tls_ldcall:");
5228 assemble_name (file
, get_some_local_dynamic_name ());
5231 output_operand_lossage ("invalid reference for 'J' output modifier");
5235 fprintf (file
, "%u", GET_MODE_SIZE (GET_MODE (x
)));
5240 struct s390_address ad
;
5245 output_operand_lossage ("memory reference expected for "
5246 "'O' output modifier");
5250 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5253 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5256 output_operand_lossage ("invalid address for 'O' output modifier");
5261 output_addr_const (file
, ad
.disp
);
5263 fprintf (file
, "0");
5269 struct s390_address ad
;
5274 output_operand_lossage ("memory reference expected for "
5275 "'R' output modifier");
5279 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5282 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5285 output_operand_lossage ("invalid address for 'R' output modifier");
5290 fprintf (file
, "%s", reg_names
[REGNO (ad
.base
)]);
5292 fprintf (file
, "0");
5298 struct s390_address ad
;
5303 output_operand_lossage ("memory reference expected for "
5304 "'S' output modifier");
5307 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5310 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5313 output_operand_lossage ("invalid address for 'S' output modifier");
5318 output_addr_const (file
, ad
.disp
);
5320 fprintf (file
, "0");
5323 fprintf (file
, "(%s)", reg_names
[REGNO (ad
.base
)]);
5328 if (GET_CODE (x
) == REG
)
5329 x
= gen_rtx_REG (GET_MODE (x
), REGNO (x
) + 1);
5330 else if (GET_CODE (x
) == MEM
)
5331 x
= change_address (x
, VOIDmode
,
5332 plus_constant (Pmode
, XEXP (x
, 0), 4));
5334 output_operand_lossage ("register or memory expression expected "
5335 "for 'N' output modifier");
5339 if (GET_CODE (x
) == REG
)
5340 x
= gen_rtx_REG (GET_MODE (x
), REGNO (x
) + 1);
5341 else if (GET_CODE (x
) == MEM
)
5342 x
= change_address (x
, VOIDmode
,
5343 plus_constant (Pmode
, XEXP (x
, 0), 8));
5345 output_operand_lossage ("register or memory expression expected "
5346 "for 'M' output modifier");
5350 print_shift_count_operand (file
, x
);
5354 switch (GET_CODE (x
))
5357 fprintf (file
, "%s", reg_names
[REGNO (x
)]);
5361 output_address (XEXP (x
, 0));
5368 output_addr_const (file
, x
);
5373 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xff);
5374 else if (code
== 'c')
5375 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ((INTVAL (x
) & 0xff) ^ 0x80) - 0x80);
5376 else if (code
== 'x')
5377 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xffff);
5378 else if (code
== 'h')
5379 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ((INTVAL (x
) & 0xffff) ^ 0x8000) - 0x8000);
5380 else if (code
== 'i')
5381 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5382 s390_extract_part (x
, HImode
, 0));
5383 else if (code
== 'j')
5384 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5385 s390_extract_part (x
, HImode
, -1));
5386 else if (code
== 'k')
5387 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5388 s390_extract_part (x
, SImode
, 0));
5389 else if (code
== 'm')
5390 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5391 s390_extract_part (x
, SImode
, -1));
5392 else if (code
== 'o')
5393 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xffffffff);
5395 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
));
5399 gcc_assert (GET_MODE (x
) == VOIDmode
);
5401 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, CONST_DOUBLE_LOW (x
) & 0xff);
5402 else if (code
== 'x')
5403 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, CONST_DOUBLE_LOW (x
) & 0xffff);
5404 else if (code
== 'h')
5405 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5406 ((CONST_DOUBLE_LOW (x
) & 0xffff) ^ 0x8000) - 0x8000);
5410 output_operand_lossage ("invalid constant - try using "
5411 "an output modifier");
5413 output_operand_lossage ("invalid constant for output modifier '%c'",
5420 output_operand_lossage ("invalid expression - try using "
5421 "an output modifier");
5423 output_operand_lossage ("invalid expression for output "
5424 "modifier '%c'", code
);
5429 /* Target hook for assembling integer objects. We need to define it
5430 here to work a round a bug in some versions of GAS, which couldn't
5431 handle values smaller than INT_MIN when printed in decimal. */
5434 s390_assemble_integer (rtx x
, unsigned int size
, int aligned_p
)
5436 if (size
== 8 && aligned_p
5437 && GET_CODE (x
) == CONST_INT
&& INTVAL (x
) < INT_MIN
)
5439 fprintf (asm_out_file
, "\t.quad\t" HOST_WIDE_INT_PRINT_HEX
"\n",
5443 return default_assemble_integer (x
, size
, aligned_p
);
5446 /* Returns true if register REGNO is used for forming
5447 a memory address in expression X. */
5450 reg_used_in_mem_p (int regno
, rtx x
)
5452 enum rtx_code code
= GET_CODE (x
);
5458 if (refers_to_regno_p (regno
, regno
+1,
5462 else if (code
== SET
5463 && GET_CODE (SET_DEST (x
)) == PC
)
5465 if (refers_to_regno_p (regno
, regno
+1,
5470 fmt
= GET_RTX_FORMAT (code
);
5471 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
5474 && reg_used_in_mem_p (regno
, XEXP (x
, i
)))
5477 else if (fmt
[i
] == 'E')
5478 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5479 if (reg_used_in_mem_p (regno
, XVECEXP (x
, i
, j
)))
5485 /* Returns true if expression DEP_RTX sets an address register
5486 used by instruction INSN to address memory. */
5489 addr_generation_dependency_p (rtx dep_rtx
, rtx insn
)
5493 if (GET_CODE (dep_rtx
) == INSN
)
5494 dep_rtx
= PATTERN (dep_rtx
);
5496 if (GET_CODE (dep_rtx
) == SET
)
5498 target
= SET_DEST (dep_rtx
);
5499 if (GET_CODE (target
) == STRICT_LOW_PART
)
5500 target
= XEXP (target
, 0);
5501 while (GET_CODE (target
) == SUBREG
)
5502 target
= SUBREG_REG (target
);
5504 if (GET_CODE (target
) == REG
)
5506 int regno
= REGNO (target
);
5508 if (s390_safe_attr_type (insn
) == TYPE_LA
)
5510 pat
= PATTERN (insn
);
5511 if (GET_CODE (pat
) == PARALLEL
)
5513 gcc_assert (XVECLEN (pat
, 0) == 2);
5514 pat
= XVECEXP (pat
, 0, 0);
5516 gcc_assert (GET_CODE (pat
) == SET
);
5517 return refers_to_regno_p (regno
, regno
+1, SET_SRC (pat
), 0);
5519 else if (get_attr_atype (insn
) == ATYPE_AGEN
)
5520 return reg_used_in_mem_p (regno
, PATTERN (insn
));
5526 /* Return 1, if dep_insn sets register used in insn in the agen unit. */
5529 s390_agen_dep_p (rtx dep_insn
, rtx insn
)
5531 rtx dep_rtx
= PATTERN (dep_insn
);
5534 if (GET_CODE (dep_rtx
) == SET
5535 && addr_generation_dependency_p (dep_rtx
, insn
))
5537 else if (GET_CODE (dep_rtx
) == PARALLEL
)
5539 for (i
= 0; i
< XVECLEN (dep_rtx
, 0); i
++)
5541 if (addr_generation_dependency_p (XVECEXP (dep_rtx
, 0, i
), insn
))
5549 /* A C statement (sans semicolon) to update the integer scheduling priority
5550 INSN_PRIORITY (INSN). Increase the priority to execute the INSN earlier,
5551 reduce the priority to execute INSN later. Do not define this macro if
5552 you do not need to adjust the scheduling priorities of insns.
5554 A STD instruction should be scheduled earlier,
5555 in order to use the bypass. */
5557 s390_adjust_priority (rtx insn ATTRIBUTE_UNUSED
, int priority
)
5559 if (! INSN_P (insn
))
5562 if (s390_tune
!= PROCESSOR_2084_Z990
5563 && s390_tune
!= PROCESSOR_2094_Z9_109
5564 && s390_tune
!= PROCESSOR_2097_Z10
5565 && s390_tune
!= PROCESSOR_2817_Z196
)
5568 switch (s390_safe_attr_type (insn
))
5572 priority
= priority
<< 3;
5576 priority
= priority
<< 1;
5585 /* The number of instructions that can be issued per cycle. */
5588 s390_issue_rate (void)
5592 case PROCESSOR_2084_Z990
:
5593 case PROCESSOR_2094_Z9_109
:
5594 case PROCESSOR_2817_Z196
:
5596 case PROCESSOR_2097_Z10
:
5604 s390_first_cycle_multipass_dfa_lookahead (void)
5609 /* Annotate every literal pool reference in X by an UNSPEC_LTREF expression.
5610 Fix up MEMs as required. */
5613 annotate_constant_pool_refs (rtx
*x
)
5618 gcc_assert (GET_CODE (*x
) != SYMBOL_REF
5619 || !CONSTANT_POOL_ADDRESS_P (*x
));
5621 /* Literal pool references can only occur inside a MEM ... */
5622 if (GET_CODE (*x
) == MEM
)
5624 rtx memref
= XEXP (*x
, 0);
5626 if (GET_CODE (memref
) == SYMBOL_REF
5627 && CONSTANT_POOL_ADDRESS_P (memref
))
5629 rtx base
= cfun
->machine
->base_reg
;
5630 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, memref
, base
),
5633 *x
= replace_equiv_address (*x
, addr
);
5637 if (GET_CODE (memref
) == CONST
5638 && GET_CODE (XEXP (memref
, 0)) == PLUS
5639 && GET_CODE (XEXP (XEXP (memref
, 0), 1)) == CONST_INT
5640 && GET_CODE (XEXP (XEXP (memref
, 0), 0)) == SYMBOL_REF
5641 && CONSTANT_POOL_ADDRESS_P (XEXP (XEXP (memref
, 0), 0)))
5643 HOST_WIDE_INT off
= INTVAL (XEXP (XEXP (memref
, 0), 1));
5644 rtx sym
= XEXP (XEXP (memref
, 0), 0);
5645 rtx base
= cfun
->machine
->base_reg
;
5646 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, sym
, base
),
5649 *x
= replace_equiv_address (*x
, plus_constant (Pmode
, addr
, off
));
5654 /* ... or a load-address type pattern. */
5655 if (GET_CODE (*x
) == SET
)
5657 rtx addrref
= SET_SRC (*x
);
5659 if (GET_CODE (addrref
) == SYMBOL_REF
5660 && CONSTANT_POOL_ADDRESS_P (addrref
))
5662 rtx base
= cfun
->machine
->base_reg
;
5663 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, addrref
, base
),
5666 SET_SRC (*x
) = addr
;
5670 if (GET_CODE (addrref
) == CONST
5671 && GET_CODE (XEXP (addrref
, 0)) == PLUS
5672 && GET_CODE (XEXP (XEXP (addrref
, 0), 1)) == CONST_INT
5673 && GET_CODE (XEXP (XEXP (addrref
, 0), 0)) == SYMBOL_REF
5674 && CONSTANT_POOL_ADDRESS_P (XEXP (XEXP (addrref
, 0), 0)))
5676 HOST_WIDE_INT off
= INTVAL (XEXP (XEXP (addrref
, 0), 1));
5677 rtx sym
= XEXP (XEXP (addrref
, 0), 0);
5678 rtx base
= cfun
->machine
->base_reg
;
5679 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, sym
, base
),
5682 SET_SRC (*x
) = plus_constant (Pmode
, addr
, off
);
5687 /* Annotate LTREL_BASE as well. */
5688 if (GET_CODE (*x
) == UNSPEC
5689 && XINT (*x
, 1) == UNSPEC_LTREL_BASE
)
5691 rtx base
= cfun
->machine
->base_reg
;
5692 *x
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, XVECEXP (*x
, 0, 0), base
),
5697 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5698 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5702 annotate_constant_pool_refs (&XEXP (*x
, i
));
5704 else if (fmt
[i
] == 'E')
5706 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5707 annotate_constant_pool_refs (&XVECEXP (*x
, i
, j
));
5712 /* Split all branches that exceed the maximum distance.
5713 Returns true if this created a new literal pool entry. */
5716 s390_split_branches (void)
5718 rtx temp_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
5719 int new_literal
= 0, ret
;
5720 rtx insn
, pat
, tmp
, target
;
5723 /* We need correct insn addresses. */
5725 shorten_branches (get_insns ());
5727 /* Find all branches that exceed 64KB, and split them. */
5729 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
5731 if (GET_CODE (insn
) != JUMP_INSN
)
5734 pat
= PATTERN (insn
);
5735 if (GET_CODE (pat
) == PARALLEL
&& XVECLEN (pat
, 0) > 2)
5736 pat
= XVECEXP (pat
, 0, 0);
5737 if (GET_CODE (pat
) != SET
|| SET_DEST (pat
) != pc_rtx
)
5740 if (GET_CODE (SET_SRC (pat
)) == LABEL_REF
)
5742 label
= &SET_SRC (pat
);
5744 else if (GET_CODE (SET_SRC (pat
)) == IF_THEN_ELSE
)
5746 if (GET_CODE (XEXP (SET_SRC (pat
), 1)) == LABEL_REF
)
5747 label
= &XEXP (SET_SRC (pat
), 1);
5748 else if (GET_CODE (XEXP (SET_SRC (pat
), 2)) == LABEL_REF
)
5749 label
= &XEXP (SET_SRC (pat
), 2);
5756 if (get_attr_length (insn
) <= 4)
5759 /* We are going to use the return register as scratch register,
5760 make sure it will be saved/restored by the prologue/epilogue. */
5761 cfun_frame_layout
.save_return_addr_p
= 1;
5766 tmp
= force_const_mem (Pmode
, *label
);
5767 tmp
= emit_insn_before (gen_rtx_SET (Pmode
, temp_reg
, tmp
), insn
);
5768 INSN_ADDRESSES_NEW (tmp
, -1);
5769 annotate_constant_pool_refs (&PATTERN (tmp
));
5776 target
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, *label
),
5777 UNSPEC_LTREL_OFFSET
);
5778 target
= gen_rtx_CONST (Pmode
, target
);
5779 target
= force_const_mem (Pmode
, target
);
5780 tmp
= emit_insn_before (gen_rtx_SET (Pmode
, temp_reg
, target
), insn
);
5781 INSN_ADDRESSES_NEW (tmp
, -1);
5782 annotate_constant_pool_refs (&PATTERN (tmp
));
5784 target
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, XEXP (target
, 0),
5785 cfun
->machine
->base_reg
),
5787 target
= gen_rtx_PLUS (Pmode
, temp_reg
, target
);
5790 ret
= validate_change (insn
, label
, target
, 0);
5798 /* Find an annotated literal pool symbol referenced in RTX X,
5799 and store it at REF. Will abort if X contains references to
5800 more than one such pool symbol; multiple references to the same
5801 symbol are allowed, however.
5803 The rtx pointed to by REF must be initialized to NULL_RTX
5804 by the caller before calling this routine. */
5807 find_constant_pool_ref (rtx x
, rtx
*ref
)
5812 /* Ignore LTREL_BASE references. */
5813 if (GET_CODE (x
) == UNSPEC
5814 && XINT (x
, 1) == UNSPEC_LTREL_BASE
)
5816 /* Likewise POOL_ENTRY insns. */
5817 if (GET_CODE (x
) == UNSPEC_VOLATILE
5818 && XINT (x
, 1) == UNSPECV_POOL_ENTRY
)
5821 gcc_assert (GET_CODE (x
) != SYMBOL_REF
5822 || !CONSTANT_POOL_ADDRESS_P (x
));
5824 if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_LTREF
)
5826 rtx sym
= XVECEXP (x
, 0, 0);
5827 gcc_assert (GET_CODE (sym
) == SYMBOL_REF
5828 && CONSTANT_POOL_ADDRESS_P (sym
));
5830 if (*ref
== NULL_RTX
)
5833 gcc_assert (*ref
== sym
);
5838 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
5839 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
5843 find_constant_pool_ref (XEXP (x
, i
), ref
);
5845 else if (fmt
[i
] == 'E')
5847 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5848 find_constant_pool_ref (XVECEXP (x
, i
, j
), ref
);
5853 /* Replace every reference to the annotated literal pool
5854 symbol REF in X by its base plus OFFSET. */
5857 replace_constant_pool_ref (rtx
*x
, rtx ref
, rtx offset
)
5862 gcc_assert (*x
!= ref
);
5864 if (GET_CODE (*x
) == UNSPEC
5865 && XINT (*x
, 1) == UNSPEC_LTREF
5866 && XVECEXP (*x
, 0, 0) == ref
)
5868 *x
= gen_rtx_PLUS (Pmode
, XVECEXP (*x
, 0, 1), offset
);
5872 if (GET_CODE (*x
) == PLUS
5873 && GET_CODE (XEXP (*x
, 1)) == CONST_INT
5874 && GET_CODE (XEXP (*x
, 0)) == UNSPEC
5875 && XINT (XEXP (*x
, 0), 1) == UNSPEC_LTREF
5876 && XVECEXP (XEXP (*x
, 0), 0, 0) == ref
)
5878 rtx addr
= gen_rtx_PLUS (Pmode
, XVECEXP (XEXP (*x
, 0), 0, 1), offset
);
5879 *x
= plus_constant (Pmode
, addr
, INTVAL (XEXP (*x
, 1)));
5883 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5884 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5888 replace_constant_pool_ref (&XEXP (*x
, i
), ref
, offset
);
5890 else if (fmt
[i
] == 'E')
5892 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5893 replace_constant_pool_ref (&XVECEXP (*x
, i
, j
), ref
, offset
);
5898 /* Check whether X contains an UNSPEC_LTREL_BASE.
5899 Return its constant pool symbol if found, NULL_RTX otherwise. */
5902 find_ltrel_base (rtx x
)
5907 if (GET_CODE (x
) == UNSPEC
5908 && XINT (x
, 1) == UNSPEC_LTREL_BASE
)
5909 return XVECEXP (x
, 0, 0);
5911 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
5912 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
5916 rtx fnd
= find_ltrel_base (XEXP (x
, i
));
5920 else if (fmt
[i
] == 'E')
5922 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5924 rtx fnd
= find_ltrel_base (XVECEXP (x
, i
, j
));
5934 /* Replace any occurrence of UNSPEC_LTREL_BASE in X with its base. */
5937 replace_ltrel_base (rtx
*x
)
5942 if (GET_CODE (*x
) == UNSPEC
5943 && XINT (*x
, 1) == UNSPEC_LTREL_BASE
)
5945 *x
= XVECEXP (*x
, 0, 1);
5949 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5950 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5954 replace_ltrel_base (&XEXP (*x
, i
));
5956 else if (fmt
[i
] == 'E')
5958 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5959 replace_ltrel_base (&XVECEXP (*x
, i
, j
));
5965 /* We keep a list of constants which we have to add to internal
5966 constant tables in the middle of large functions. */
5968 #define NR_C_MODES 11
5969 enum machine_mode constant_modes
[NR_C_MODES
] =
5971 TFmode
, TImode
, TDmode
,
5972 DFmode
, DImode
, DDmode
,
5973 SFmode
, SImode
, SDmode
,
5980 struct constant
*next
;
5985 struct constant_pool
5987 struct constant_pool
*next
;
5991 rtx emit_pool_after
;
5993 struct constant
*constants
[NR_C_MODES
];
5994 struct constant
*execute
;
5999 /* Allocate new constant_pool structure. */
6001 static struct constant_pool
*
6002 s390_alloc_pool (void)
6004 struct constant_pool
*pool
;
6007 pool
= (struct constant_pool
*) xmalloc (sizeof *pool
);
6009 for (i
= 0; i
< NR_C_MODES
; i
++)
6010 pool
->constants
[i
] = NULL
;
6012 pool
->execute
= NULL
;
6013 pool
->label
= gen_label_rtx ();
6014 pool
->first_insn
= NULL_RTX
;
6015 pool
->pool_insn
= NULL_RTX
;
6016 pool
->insns
= BITMAP_ALLOC (NULL
);
6018 pool
->emit_pool_after
= NULL_RTX
;
6023 /* Create new constant pool covering instructions starting at INSN
6024 and chain it to the end of POOL_LIST. */
6026 static struct constant_pool
*
6027 s390_start_pool (struct constant_pool
**pool_list
, rtx insn
)
6029 struct constant_pool
*pool
, **prev
;
6031 pool
= s390_alloc_pool ();
6032 pool
->first_insn
= insn
;
6034 for (prev
= pool_list
; *prev
; prev
= &(*prev
)->next
)
6041 /* End range of instructions covered by POOL at INSN and emit
6042 placeholder insn representing the pool. */
6045 s390_end_pool (struct constant_pool
*pool
, rtx insn
)
6047 rtx pool_size
= GEN_INT (pool
->size
+ 8 /* alignment slop */);
6050 insn
= get_last_insn ();
6052 pool
->pool_insn
= emit_insn_after (gen_pool (pool_size
), insn
);
6053 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6056 /* Add INSN to the list of insns covered by POOL. */
6059 s390_add_pool_insn (struct constant_pool
*pool
, rtx insn
)
6061 bitmap_set_bit (pool
->insns
, INSN_UID (insn
));
6064 /* Return pool out of POOL_LIST that covers INSN. */
6066 static struct constant_pool
*
6067 s390_find_pool (struct constant_pool
*pool_list
, rtx insn
)
6069 struct constant_pool
*pool
;
6071 for (pool
= pool_list
; pool
; pool
= pool
->next
)
6072 if (bitmap_bit_p (pool
->insns
, INSN_UID (insn
)))
6078 /* Add constant VAL of mode MODE to the constant pool POOL. */
6081 s390_add_constant (struct constant_pool
*pool
, rtx val
, enum machine_mode mode
)
6086 for (i
= 0; i
< NR_C_MODES
; i
++)
6087 if (constant_modes
[i
] == mode
)
6089 gcc_assert (i
!= NR_C_MODES
);
6091 for (c
= pool
->constants
[i
]; c
!= NULL
; c
= c
->next
)
6092 if (rtx_equal_p (val
, c
->value
))
6097 c
= (struct constant
*) xmalloc (sizeof *c
);
6099 c
->label
= gen_label_rtx ();
6100 c
->next
= pool
->constants
[i
];
6101 pool
->constants
[i
] = c
;
6102 pool
->size
+= GET_MODE_SIZE (mode
);
6106 /* Return an rtx that represents the offset of X from the start of
6110 s390_pool_offset (struct constant_pool
*pool
, rtx x
)
6114 label
= gen_rtx_LABEL_REF (GET_MODE (x
), pool
->label
);
6115 x
= gen_rtx_UNSPEC (GET_MODE (x
), gen_rtvec (2, x
, label
),
6116 UNSPEC_POOL_OFFSET
);
6117 return gen_rtx_CONST (GET_MODE (x
), x
);
6120 /* Find constant VAL of mode MODE in the constant pool POOL.
6121 Return an RTX describing the distance from the start of
6122 the pool to the location of the new constant. */
6125 s390_find_constant (struct constant_pool
*pool
, rtx val
,
6126 enum machine_mode mode
)
6131 for (i
= 0; i
< NR_C_MODES
; i
++)
6132 if (constant_modes
[i
] == mode
)
6134 gcc_assert (i
!= NR_C_MODES
);
6136 for (c
= pool
->constants
[i
]; c
!= NULL
; c
= c
->next
)
6137 if (rtx_equal_p (val
, c
->value
))
6142 return s390_pool_offset (pool
, gen_rtx_LABEL_REF (Pmode
, c
->label
));
6145 /* Check whether INSN is an execute. Return the label_ref to its
6146 execute target template if so, NULL_RTX otherwise. */
6149 s390_execute_label (rtx insn
)
6151 if (GET_CODE (insn
) == INSN
6152 && GET_CODE (PATTERN (insn
)) == PARALLEL
6153 && GET_CODE (XVECEXP (PATTERN (insn
), 0, 0)) == UNSPEC
6154 && XINT (XVECEXP (PATTERN (insn
), 0, 0), 1) == UNSPEC_EXECUTE
)
6155 return XVECEXP (XVECEXP (PATTERN (insn
), 0, 0), 0, 2);
6160 /* Add execute target for INSN to the constant pool POOL. */
6163 s390_add_execute (struct constant_pool
*pool
, rtx insn
)
6167 for (c
= pool
->execute
; c
!= NULL
; c
= c
->next
)
6168 if (INSN_UID (insn
) == INSN_UID (c
->value
))
6173 c
= (struct constant
*) xmalloc (sizeof *c
);
6175 c
->label
= gen_label_rtx ();
6176 c
->next
= pool
->execute
;
6182 /* Find execute target for INSN in the constant pool POOL.
6183 Return an RTX describing the distance from the start of
6184 the pool to the location of the execute target. */
6187 s390_find_execute (struct constant_pool
*pool
, rtx insn
)
6191 for (c
= pool
->execute
; c
!= NULL
; c
= c
->next
)
6192 if (INSN_UID (insn
) == INSN_UID (c
->value
))
6197 return s390_pool_offset (pool
, gen_rtx_LABEL_REF (Pmode
, c
->label
));
6200 /* For an execute INSN, extract the execute target template. */
6203 s390_execute_target (rtx insn
)
6205 rtx pattern
= PATTERN (insn
);
6206 gcc_assert (s390_execute_label (insn
));
6208 if (XVECLEN (pattern
, 0) == 2)
6210 pattern
= copy_rtx (XVECEXP (pattern
, 0, 1));
6214 rtvec vec
= rtvec_alloc (XVECLEN (pattern
, 0) - 1);
6217 for (i
= 0; i
< XVECLEN (pattern
, 0) - 1; i
++)
6218 RTVEC_ELT (vec
, i
) = copy_rtx (XVECEXP (pattern
, 0, i
+ 1));
6220 pattern
= gen_rtx_PARALLEL (VOIDmode
, vec
);
6226 /* Indicate that INSN cannot be duplicated. This is the case for
6227 execute insns that carry a unique label. */
6230 s390_cannot_copy_insn_p (rtx insn
)
6232 rtx label
= s390_execute_label (insn
);
6233 return label
&& label
!= const0_rtx
;
6236 /* Dump out the constants in POOL. If REMOTE_LABEL is true,
6237 do not emit the pool base label. */
6240 s390_dump_pool (struct constant_pool
*pool
, bool remote_label
)
6243 rtx insn
= pool
->pool_insn
;
6246 /* Switch to rodata section. */
6247 if (TARGET_CPU_ZARCH
)
6249 insn
= emit_insn_after (gen_pool_section_start (), insn
);
6250 INSN_ADDRESSES_NEW (insn
, -1);
6253 /* Ensure minimum pool alignment. */
6254 if (TARGET_CPU_ZARCH
)
6255 insn
= emit_insn_after (gen_pool_align (GEN_INT (8)), insn
);
6257 insn
= emit_insn_after (gen_pool_align (GEN_INT (4)), insn
);
6258 INSN_ADDRESSES_NEW (insn
, -1);
6260 /* Emit pool base label. */
6263 insn
= emit_label_after (pool
->label
, insn
);
6264 INSN_ADDRESSES_NEW (insn
, -1);
6267 /* Dump constants in descending alignment requirement order,
6268 ensuring proper alignment for every constant. */
6269 for (i
= 0; i
< NR_C_MODES
; i
++)
6270 for (c
= pool
->constants
[i
]; c
; c
= c
->next
)
6272 /* Convert UNSPEC_LTREL_OFFSET unspecs to pool-relative references. */
6273 rtx value
= copy_rtx (c
->value
);
6274 if (GET_CODE (value
) == CONST
6275 && GET_CODE (XEXP (value
, 0)) == UNSPEC
6276 && XINT (XEXP (value
, 0), 1) == UNSPEC_LTREL_OFFSET
6277 && XVECLEN (XEXP (value
, 0), 0) == 1)
6278 value
= s390_pool_offset (pool
, XVECEXP (XEXP (value
, 0), 0, 0));
6280 insn
= emit_label_after (c
->label
, insn
);
6281 INSN_ADDRESSES_NEW (insn
, -1);
6283 value
= gen_rtx_UNSPEC_VOLATILE (constant_modes
[i
],
6284 gen_rtvec (1, value
),
6285 UNSPECV_POOL_ENTRY
);
6286 insn
= emit_insn_after (value
, insn
);
6287 INSN_ADDRESSES_NEW (insn
, -1);
6290 /* Ensure minimum alignment for instructions. */
6291 insn
= emit_insn_after (gen_pool_align (GEN_INT (2)), insn
);
6292 INSN_ADDRESSES_NEW (insn
, -1);
6294 /* Output in-pool execute template insns. */
6295 for (c
= pool
->execute
; c
; c
= c
->next
)
6297 insn
= emit_label_after (c
->label
, insn
);
6298 INSN_ADDRESSES_NEW (insn
, -1);
6300 insn
= emit_insn_after (s390_execute_target (c
->value
), insn
);
6301 INSN_ADDRESSES_NEW (insn
, -1);
6304 /* Switch back to previous section. */
6305 if (TARGET_CPU_ZARCH
)
6307 insn
= emit_insn_after (gen_pool_section_end (), insn
);
6308 INSN_ADDRESSES_NEW (insn
, -1);
6311 insn
= emit_barrier_after (insn
);
6312 INSN_ADDRESSES_NEW (insn
, -1);
6314 /* Remove placeholder insn. */
6315 remove_insn (pool
->pool_insn
);
6318 /* Free all memory used by POOL. */
6321 s390_free_pool (struct constant_pool
*pool
)
6323 struct constant
*c
, *next
;
6326 for (i
= 0; i
< NR_C_MODES
; i
++)
6327 for (c
= pool
->constants
[i
]; c
; c
= next
)
6333 for (c
= pool
->execute
; c
; c
= next
)
6339 BITMAP_FREE (pool
->insns
);
6344 /* Collect main literal pool. Return NULL on overflow. */
6346 static struct constant_pool
*
6347 s390_mainpool_start (void)
6349 struct constant_pool
*pool
;
6352 pool
= s390_alloc_pool ();
6354 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6356 if (GET_CODE (insn
) == INSN
6357 && GET_CODE (PATTERN (insn
)) == SET
6358 && GET_CODE (SET_SRC (PATTERN (insn
))) == UNSPEC_VOLATILE
6359 && XINT (SET_SRC (PATTERN (insn
)), 1) == UNSPECV_MAIN_POOL
)
6361 gcc_assert (!pool
->pool_insn
);
6362 pool
->pool_insn
= insn
;
6365 if (!TARGET_CPU_ZARCH
&& s390_execute_label (insn
))
6367 s390_add_execute (pool
, insn
);
6369 else if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6371 rtx pool_ref
= NULL_RTX
;
6372 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6375 rtx constant
= get_pool_constant (pool_ref
);
6376 enum machine_mode mode
= get_pool_mode (pool_ref
);
6377 s390_add_constant (pool
, constant
, mode
);
6381 /* If hot/cold partitioning is enabled we have to make sure that
6382 the literal pool is emitted in the same section where the
6383 initialization of the literal pool base pointer takes place.
6384 emit_pool_after is only used in the non-overflow case on non
6385 Z cpus where we can emit the literal pool at the end of the
6386 function body within the text section. */
6388 && NOTE_KIND (insn
) == NOTE_INSN_SWITCH_TEXT_SECTIONS
6389 && !pool
->emit_pool_after
)
6390 pool
->emit_pool_after
= PREV_INSN (insn
);
6393 gcc_assert (pool
->pool_insn
|| pool
->size
== 0);
6395 if (pool
->size
>= 4096)
6397 /* We're going to chunkify the pool, so remove the main
6398 pool placeholder insn. */
6399 remove_insn (pool
->pool_insn
);
6401 s390_free_pool (pool
);
6405 /* If the functions ends with the section where the literal pool
6406 should be emitted set the marker to its end. */
6407 if (pool
&& !pool
->emit_pool_after
)
6408 pool
->emit_pool_after
= get_last_insn ();
6413 /* POOL holds the main literal pool as collected by s390_mainpool_start.
6414 Modify the current function to output the pool constants as well as
6415 the pool register setup instruction. */
6418 s390_mainpool_finish (struct constant_pool
*pool
)
6420 rtx base_reg
= cfun
->machine
->base_reg
;
6423 /* If the pool is empty, we're done. */
6424 if (pool
->size
== 0)
6426 /* We don't actually need a base register after all. */
6427 cfun
->machine
->base_reg
= NULL_RTX
;
6429 if (pool
->pool_insn
)
6430 remove_insn (pool
->pool_insn
);
6431 s390_free_pool (pool
);
6435 /* We need correct insn addresses. */
6436 shorten_branches (get_insns ());
6438 /* On zSeries, we use a LARL to load the pool register. The pool is
6439 located in the .rodata section, so we emit it after the function. */
6440 if (TARGET_CPU_ZARCH
)
6442 insn
= gen_main_base_64 (base_reg
, pool
->label
);
6443 insn
= emit_insn_after (insn
, pool
->pool_insn
);
6444 INSN_ADDRESSES_NEW (insn
, -1);
6445 remove_insn (pool
->pool_insn
);
6447 insn
= get_last_insn ();
6448 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6449 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6451 s390_dump_pool (pool
, 0);
6454 /* On S/390, if the total size of the function's code plus literal pool
6455 does not exceed 4096 bytes, we use BASR to set up a function base
6456 pointer, and emit the literal pool at the end of the function. */
6457 else if (INSN_ADDRESSES (INSN_UID (pool
->emit_pool_after
))
6458 + pool
->size
+ 8 /* alignment slop */ < 4096)
6460 insn
= gen_main_base_31_small (base_reg
, pool
->label
);
6461 insn
= emit_insn_after (insn
, pool
->pool_insn
);
6462 INSN_ADDRESSES_NEW (insn
, -1);
6463 remove_insn (pool
->pool_insn
);
6465 insn
= emit_label_after (pool
->label
, insn
);
6466 INSN_ADDRESSES_NEW (insn
, -1);
6468 /* emit_pool_after will be set by s390_mainpool_start to the
6469 last insn of the section where the literal pool should be
6471 insn
= pool
->emit_pool_after
;
6473 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6474 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6476 s390_dump_pool (pool
, 1);
6479 /* Otherwise, we emit an inline literal pool and use BASR to branch
6480 over it, setting up the pool register at the same time. */
6483 rtx pool_end
= gen_label_rtx ();
6485 insn
= gen_main_base_31_large (base_reg
, pool
->label
, pool_end
);
6486 insn
= emit_jump_insn_after (insn
, pool
->pool_insn
);
6487 JUMP_LABEL (insn
) = pool_end
;
6488 INSN_ADDRESSES_NEW (insn
, -1);
6489 remove_insn (pool
->pool_insn
);
6491 insn
= emit_label_after (pool
->label
, insn
);
6492 INSN_ADDRESSES_NEW (insn
, -1);
6494 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6495 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6497 insn
= emit_label_after (pool_end
, pool
->pool_insn
);
6498 INSN_ADDRESSES_NEW (insn
, -1);
6500 s390_dump_pool (pool
, 1);
6504 /* Replace all literal pool references. */
6506 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6509 replace_ltrel_base (&PATTERN (insn
));
6511 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6513 rtx addr
, pool_ref
= NULL_RTX
;
6514 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6517 if (s390_execute_label (insn
))
6518 addr
= s390_find_execute (pool
, insn
);
6520 addr
= s390_find_constant (pool
, get_pool_constant (pool_ref
),
6521 get_pool_mode (pool_ref
));
6523 replace_constant_pool_ref (&PATTERN (insn
), pool_ref
, addr
);
6524 INSN_CODE (insn
) = -1;
6530 /* Free the pool. */
6531 s390_free_pool (pool
);
6534 /* POOL holds the main literal pool as collected by s390_mainpool_start.
6535 We have decided we cannot use this pool, so revert all changes
6536 to the current function that were done by s390_mainpool_start. */
6538 s390_mainpool_cancel (struct constant_pool
*pool
)
6540 /* We didn't actually change the instruction stream, so simply
6541 free the pool memory. */
6542 s390_free_pool (pool
);
6546 /* Chunkify the literal pool. */
6548 #define S390_POOL_CHUNK_MIN 0xc00
6549 #define S390_POOL_CHUNK_MAX 0xe00
6551 static struct constant_pool
*
6552 s390_chunkify_start (void)
6554 struct constant_pool
*curr_pool
= NULL
, *pool_list
= NULL
;
6557 rtx pending_ltrel
= NULL_RTX
;
6560 rtx (*gen_reload_base
) (rtx
, rtx
) =
6561 TARGET_CPU_ZARCH
? gen_reload_base_64
: gen_reload_base_31
;
6564 /* We need correct insn addresses. */
6566 shorten_branches (get_insns ());
6568 /* Scan all insns and move literals to pool chunks. */
6570 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6572 bool section_switch_p
= false;
6574 /* Check for pending LTREL_BASE. */
6577 rtx ltrel_base
= find_ltrel_base (PATTERN (insn
));
6580 gcc_assert (ltrel_base
== pending_ltrel
);
6581 pending_ltrel
= NULL_RTX
;
6585 if (!TARGET_CPU_ZARCH
&& s390_execute_label (insn
))
6588 curr_pool
= s390_start_pool (&pool_list
, insn
);
6590 s390_add_execute (curr_pool
, insn
);
6591 s390_add_pool_insn (curr_pool
, insn
);
6593 else if (GET_CODE (insn
) == INSN
|| CALL_P (insn
))
6595 rtx pool_ref
= NULL_RTX
;
6596 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6599 rtx constant
= get_pool_constant (pool_ref
);
6600 enum machine_mode mode
= get_pool_mode (pool_ref
);
6603 curr_pool
= s390_start_pool (&pool_list
, insn
);
6605 s390_add_constant (curr_pool
, constant
, mode
);
6606 s390_add_pool_insn (curr_pool
, insn
);
6608 /* Don't split the pool chunk between a LTREL_OFFSET load
6609 and the corresponding LTREL_BASE. */
6610 if (GET_CODE (constant
) == CONST
6611 && GET_CODE (XEXP (constant
, 0)) == UNSPEC
6612 && XINT (XEXP (constant
, 0), 1) == UNSPEC_LTREL_OFFSET
)
6614 gcc_assert (!pending_ltrel
);
6615 pending_ltrel
= pool_ref
;
6620 if (GET_CODE (insn
) == JUMP_INSN
|| GET_CODE (insn
) == CODE_LABEL
)
6623 s390_add_pool_insn (curr_pool
, insn
);
6624 /* An LTREL_BASE must follow within the same basic block. */
6625 gcc_assert (!pending_ltrel
);
6629 switch (NOTE_KIND (insn
))
6631 case NOTE_INSN_SWITCH_TEXT_SECTIONS
:
6632 section_switch_p
= true;
6634 case NOTE_INSN_VAR_LOCATION
:
6635 case NOTE_INSN_CALL_ARG_LOCATION
:
6642 || INSN_ADDRESSES_SIZE () <= (size_t) INSN_UID (insn
)
6643 || INSN_ADDRESSES (INSN_UID (insn
)) == -1)
6646 if (TARGET_CPU_ZARCH
)
6648 if (curr_pool
->size
< S390_POOL_CHUNK_MAX
)
6651 s390_end_pool (curr_pool
, NULL_RTX
);
6656 int chunk_size
= INSN_ADDRESSES (INSN_UID (insn
))
6657 - INSN_ADDRESSES (INSN_UID (curr_pool
->first_insn
))
6660 /* We will later have to insert base register reload insns.
6661 Those will have an effect on code size, which we need to
6662 consider here. This calculation makes rather pessimistic
6663 worst-case assumptions. */
6664 if (GET_CODE (insn
) == CODE_LABEL
)
6667 if (chunk_size
< S390_POOL_CHUNK_MIN
6668 && curr_pool
->size
< S390_POOL_CHUNK_MIN
6669 && !section_switch_p
)
6672 /* Pool chunks can only be inserted after BARRIERs ... */
6673 if (GET_CODE (insn
) == BARRIER
)
6675 s390_end_pool (curr_pool
, insn
);
6680 /* ... so if we don't find one in time, create one. */
6681 else if (chunk_size
> S390_POOL_CHUNK_MAX
6682 || curr_pool
->size
> S390_POOL_CHUNK_MAX
6683 || section_switch_p
)
6685 rtx label
, jump
, barrier
, next
, prev
;
6687 if (!section_switch_p
)
6689 /* We can insert the barrier only after a 'real' insn. */
6690 if (GET_CODE (insn
) != INSN
&& GET_CODE (insn
) != CALL_INSN
)
6692 if (get_attr_length (insn
) == 0)
6694 /* Don't separate LTREL_BASE from the corresponding
6695 LTREL_OFFSET load. */
6702 next
= NEXT_INSN (insn
);
6706 && (NOTE_KIND (next
) == NOTE_INSN_VAR_LOCATION
6707 || NOTE_KIND (next
) == NOTE_INSN_CALL_ARG_LOCATION
));
6711 gcc_assert (!pending_ltrel
);
6713 /* The old pool has to end before the section switch
6714 note in order to make it part of the current
6716 insn
= PREV_INSN (insn
);
6719 label
= gen_label_rtx ();
6721 if (prev
&& NOTE_P (prev
))
6722 prev
= prev_nonnote_insn (prev
);
6724 jump
= emit_jump_insn_after_setloc (gen_jump (label
), insn
,
6725 INSN_LOCATOR (prev
));
6727 jump
= emit_jump_insn_after_noloc (gen_jump (label
), insn
);
6728 barrier
= emit_barrier_after (jump
);
6729 insn
= emit_label_after (label
, barrier
);
6730 JUMP_LABEL (jump
) = label
;
6731 LABEL_NUSES (label
) = 1;
6733 INSN_ADDRESSES_NEW (jump
, -1);
6734 INSN_ADDRESSES_NEW (barrier
, -1);
6735 INSN_ADDRESSES_NEW (insn
, -1);
6737 s390_end_pool (curr_pool
, barrier
);
6745 s390_end_pool (curr_pool
, NULL_RTX
);
6746 gcc_assert (!pending_ltrel
);
6748 /* Find all labels that are branched into
6749 from an insn belonging to a different chunk. */
6751 far_labels
= BITMAP_ALLOC (NULL
);
6753 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6755 /* Labels marked with LABEL_PRESERVE_P can be target
6756 of non-local jumps, so we have to mark them.
6757 The same holds for named labels.
6759 Don't do that, however, if it is the label before
6762 if (GET_CODE (insn
) == CODE_LABEL
6763 && (LABEL_PRESERVE_P (insn
) || LABEL_NAME (insn
)))
6765 rtx vec_insn
= next_real_insn (insn
);
6766 rtx vec_pat
= vec_insn
&& GET_CODE (vec_insn
) == JUMP_INSN
?
6767 PATTERN (vec_insn
) : NULL_RTX
;
6769 || !(GET_CODE (vec_pat
) == ADDR_VEC
6770 || GET_CODE (vec_pat
) == ADDR_DIFF_VEC
))
6771 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (insn
));
6774 /* If we have a direct jump (conditional or unconditional)
6775 or a casesi jump, check all potential targets. */
6776 else if (GET_CODE (insn
) == JUMP_INSN
)
6778 rtx pat
= PATTERN (insn
);
6779 if (GET_CODE (pat
) == PARALLEL
&& XVECLEN (pat
, 0) > 2)
6780 pat
= XVECEXP (pat
, 0, 0);
6782 if (GET_CODE (pat
) == SET
)
6784 rtx label
= JUMP_LABEL (insn
);
6787 if (s390_find_pool (pool_list
, label
)
6788 != s390_find_pool (pool_list
, insn
))
6789 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (label
));
6792 else if (GET_CODE (pat
) == PARALLEL
6793 && XVECLEN (pat
, 0) == 2
6794 && GET_CODE (XVECEXP (pat
, 0, 0)) == SET
6795 && GET_CODE (XVECEXP (pat
, 0, 1)) == USE
6796 && GET_CODE (XEXP (XVECEXP (pat
, 0, 1), 0)) == LABEL_REF
)
6798 /* Find the jump table used by this casesi jump. */
6799 rtx vec_label
= XEXP (XEXP (XVECEXP (pat
, 0, 1), 0), 0);
6800 rtx vec_insn
= next_real_insn (vec_label
);
6801 rtx vec_pat
= vec_insn
&& GET_CODE (vec_insn
) == JUMP_INSN
?
6802 PATTERN (vec_insn
) : NULL_RTX
;
6804 && (GET_CODE (vec_pat
) == ADDR_VEC
6805 || GET_CODE (vec_pat
) == ADDR_DIFF_VEC
))
6807 int i
, diff_p
= GET_CODE (vec_pat
) == ADDR_DIFF_VEC
;
6809 for (i
= 0; i
< XVECLEN (vec_pat
, diff_p
); i
++)
6811 rtx label
= XEXP (XVECEXP (vec_pat
, diff_p
, i
), 0);
6813 if (s390_find_pool (pool_list
, label
)
6814 != s390_find_pool (pool_list
, insn
))
6815 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (label
));
6822 /* Insert base register reload insns before every pool. */
6824 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6826 rtx new_insn
= gen_reload_base (cfun
->machine
->base_reg
,
6828 rtx insn
= curr_pool
->first_insn
;
6829 INSN_ADDRESSES_NEW (emit_insn_before (new_insn
, insn
), -1);
6832 /* Insert base register reload insns at every far label. */
6834 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6835 if (GET_CODE (insn
) == CODE_LABEL
6836 && bitmap_bit_p (far_labels
, CODE_LABEL_NUMBER (insn
)))
6838 struct constant_pool
*pool
= s390_find_pool (pool_list
, insn
);
6841 rtx new_insn
= gen_reload_base (cfun
->machine
->base_reg
,
6843 INSN_ADDRESSES_NEW (emit_insn_after (new_insn
, insn
), -1);
6848 BITMAP_FREE (far_labels
);
6851 /* Recompute insn addresses. */
6853 init_insn_lengths ();
6854 shorten_branches (get_insns ());
6859 /* POOL_LIST is a chunk list as prepared by s390_chunkify_start.
6860 After we have decided to use this list, finish implementing
6861 all changes to the current function as required. */
6864 s390_chunkify_finish (struct constant_pool
*pool_list
)
6866 struct constant_pool
*curr_pool
= NULL
;
6870 /* Replace all literal pool references. */
6872 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6875 replace_ltrel_base (&PATTERN (insn
));
6877 curr_pool
= s390_find_pool (pool_list
, insn
);
6881 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6883 rtx addr
, pool_ref
= NULL_RTX
;
6884 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6887 if (s390_execute_label (insn
))
6888 addr
= s390_find_execute (curr_pool
, insn
);
6890 addr
= s390_find_constant (curr_pool
,
6891 get_pool_constant (pool_ref
),
6892 get_pool_mode (pool_ref
));
6894 replace_constant_pool_ref (&PATTERN (insn
), pool_ref
, addr
);
6895 INSN_CODE (insn
) = -1;
6900 /* Dump out all literal pools. */
6902 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6903 s390_dump_pool (curr_pool
, 0);
6905 /* Free pool list. */
6909 struct constant_pool
*next
= pool_list
->next
;
6910 s390_free_pool (pool_list
);
6915 /* POOL_LIST is a chunk list as prepared by s390_chunkify_start.
6916 We have decided we cannot use this list, so revert all changes
6917 to the current function that were done by s390_chunkify_start. */
6920 s390_chunkify_cancel (struct constant_pool
*pool_list
)
6922 struct constant_pool
*curr_pool
= NULL
;
6925 /* Remove all pool placeholder insns. */
6927 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6929 /* Did we insert an extra barrier? Remove it. */
6930 rtx barrier
= PREV_INSN (curr_pool
->pool_insn
);
6931 rtx jump
= barrier
? PREV_INSN (barrier
) : NULL_RTX
;
6932 rtx label
= NEXT_INSN (curr_pool
->pool_insn
);
6934 if (jump
&& GET_CODE (jump
) == JUMP_INSN
6935 && barrier
&& GET_CODE (barrier
) == BARRIER
6936 && label
&& GET_CODE (label
) == CODE_LABEL
6937 && GET_CODE (PATTERN (jump
)) == SET
6938 && SET_DEST (PATTERN (jump
)) == pc_rtx
6939 && GET_CODE (SET_SRC (PATTERN (jump
))) == LABEL_REF
6940 && XEXP (SET_SRC (PATTERN (jump
)), 0) == label
)
6943 remove_insn (barrier
);
6944 remove_insn (label
);
6947 remove_insn (curr_pool
->pool_insn
);
6950 /* Remove all base register reload insns. */
6952 for (insn
= get_insns (); insn
; )
6954 rtx next_insn
= NEXT_INSN (insn
);
6956 if (GET_CODE (insn
) == INSN
6957 && GET_CODE (PATTERN (insn
)) == SET
6958 && GET_CODE (SET_SRC (PATTERN (insn
))) == UNSPEC
6959 && XINT (SET_SRC (PATTERN (insn
)), 1) == UNSPEC_RELOAD_BASE
)
6965 /* Free pool list. */
6969 struct constant_pool
*next
= pool_list
->next
;
6970 s390_free_pool (pool_list
);
6975 /* Output the constant pool entry EXP in mode MODE with alignment ALIGN. */
6978 s390_output_pool_entry (rtx exp
, enum machine_mode mode
, unsigned int align
)
6982 switch (GET_MODE_CLASS (mode
))
6985 case MODE_DECIMAL_FLOAT
:
6986 gcc_assert (GET_CODE (exp
) == CONST_DOUBLE
);
6988 REAL_VALUE_FROM_CONST_DOUBLE (r
, exp
);
6989 assemble_real (r
, mode
, align
);
6993 assemble_integer (exp
, GET_MODE_SIZE (mode
), align
, 1);
6994 mark_symbol_refs_as_used (exp
);
7003 /* Return an RTL expression representing the value of the return address
7004 for the frame COUNT steps up from the current frame. FRAME is the
7005 frame pointer of that frame. */
7008 s390_return_addr_rtx (int count
, rtx frame ATTRIBUTE_UNUSED
)
7013 /* Without backchain, we fail for all but the current frame. */
7015 if (!TARGET_BACKCHAIN
&& count
> 0)
7018 /* For the current frame, we need to make sure the initial
7019 value of RETURN_REGNUM is actually saved. */
7023 /* On non-z architectures branch splitting could overwrite r14. */
7024 if (TARGET_CPU_ZARCH
)
7025 return get_hard_reg_initial_val (Pmode
, RETURN_REGNUM
);
7028 cfun_frame_layout
.save_return_addr_p
= true;
7029 return gen_rtx_MEM (Pmode
, return_address_pointer_rtx
);
7033 if (TARGET_PACKED_STACK
)
7034 offset
= -2 * UNITS_PER_LONG
;
7036 offset
= RETURN_REGNUM
* UNITS_PER_LONG
;
7038 addr
= plus_constant (Pmode
, frame
, offset
);
7039 addr
= memory_address (Pmode
, addr
);
7040 return gen_rtx_MEM (Pmode
, addr
);
7043 /* Return an RTL expression representing the back chain stored in
7044 the current stack frame. */
7047 s390_back_chain_rtx (void)
7051 gcc_assert (TARGET_BACKCHAIN
);
7053 if (TARGET_PACKED_STACK
)
7054 chain
= plus_constant (Pmode
, stack_pointer_rtx
,
7055 STACK_POINTER_OFFSET
- UNITS_PER_LONG
);
7057 chain
= stack_pointer_rtx
;
7059 chain
= gen_rtx_MEM (Pmode
, chain
);
7063 /* Find first call clobbered register unused in a function.
7064 This could be used as base register in a leaf function
7065 or for holding the return address before epilogue. */
7068 find_unused_clobbered_reg (void)
7071 for (i
= 0; i
< 6; i
++)
7072 if (!df_regs_ever_live_p (i
))
7078 /* Helper function for s390_regs_ever_clobbered. Sets the fields in DATA for all
7079 clobbered hard regs in SETREG. */
7082 s390_reg_clobbered_rtx (rtx setreg
, const_rtx set_insn ATTRIBUTE_UNUSED
, void *data
)
7084 int *regs_ever_clobbered
= (int *)data
;
7085 unsigned int i
, regno
;
7086 enum machine_mode mode
= GET_MODE (setreg
);
7088 if (GET_CODE (setreg
) == SUBREG
)
7090 rtx inner
= SUBREG_REG (setreg
);
7091 if (!GENERAL_REG_P (inner
))
7093 regno
= subreg_regno (setreg
);
7095 else if (GENERAL_REG_P (setreg
))
7096 regno
= REGNO (setreg
);
7101 i
< regno
+ HARD_REGNO_NREGS (regno
, mode
);
7103 regs_ever_clobbered
[i
] = 1;
7106 /* Walks through all basic blocks of the current function looking
7107 for clobbered hard regs using s390_reg_clobbered_rtx. The fields
7108 of the passed integer array REGS_EVER_CLOBBERED are set to one for
7109 each of those regs. */
7112 s390_regs_ever_clobbered (int *regs_ever_clobbered
)
7118 memset (regs_ever_clobbered
, 0, 16 * sizeof (int));
7120 /* For non-leaf functions we have to consider all call clobbered regs to be
7122 if (!current_function_is_leaf
)
7124 for (i
= 0; i
< 16; i
++)
7125 regs_ever_clobbered
[i
] = call_really_used_regs
[i
];
7128 /* Make the "magic" eh_return registers live if necessary. For regs_ever_live
7129 this work is done by liveness analysis (mark_regs_live_at_end).
7130 Special care is needed for functions containing landing pads. Landing pads
7131 may use the eh registers, but the code which sets these registers is not
7132 contained in that function. Hence s390_regs_ever_clobbered is not able to
7133 deal with this automatically. */
7134 if (crtl
->calls_eh_return
|| cfun
->machine
->has_landing_pad_p
)
7135 for (i
= 0; EH_RETURN_DATA_REGNO (i
) != INVALID_REGNUM
; i
++)
7136 if (crtl
->calls_eh_return
7137 || (cfun
->machine
->has_landing_pad_p
7138 && df_regs_ever_live_p (EH_RETURN_DATA_REGNO (i
))))
7139 regs_ever_clobbered
[EH_RETURN_DATA_REGNO (i
)] = 1;
7141 /* For nonlocal gotos all call-saved registers have to be saved.
7142 This flag is also set for the unwinding code in libgcc.
7143 See expand_builtin_unwind_init. For regs_ever_live this is done by
7145 if (cfun
->has_nonlocal_label
)
7146 for (i
= 0; i
< 16; i
++)
7147 if (!call_really_used_regs
[i
])
7148 regs_ever_clobbered
[i
] = 1;
7150 FOR_EACH_BB (cur_bb
)
7152 FOR_BB_INSNS (cur_bb
, cur_insn
)
7154 if (INSN_P (cur_insn
))
7155 note_stores (PATTERN (cur_insn
),
7156 s390_reg_clobbered_rtx
,
7157 regs_ever_clobbered
);
7162 /* Determine the frame area which actually has to be accessed
7163 in the function epilogue. The values are stored at the
7164 given pointers AREA_BOTTOM (address of the lowest used stack
7165 address) and AREA_TOP (address of the first item which does
7166 not belong to the stack frame). */
7169 s390_frame_area (int *area_bottom
, int *area_top
)
7177 if (cfun_frame_layout
.first_restore_gpr
!= -1)
7179 b
= (cfun_frame_layout
.gprs_offset
7180 + cfun_frame_layout
.first_restore_gpr
* UNITS_PER_LONG
);
7181 t
= b
+ (cfun_frame_layout
.last_restore_gpr
7182 - cfun_frame_layout
.first_restore_gpr
+ 1) * UNITS_PER_LONG
;
7185 if (TARGET_64BIT
&& cfun_save_high_fprs_p
)
7187 b
= MIN (b
, cfun_frame_layout
.f8_offset
);
7188 t
= MAX (t
, (cfun_frame_layout
.f8_offset
7189 + cfun_frame_layout
.high_fprs
* 8));
7193 for (i
= 2; i
< 4; i
++)
7194 if (cfun_fpr_bit_p (i
))
7196 b
= MIN (b
, cfun_frame_layout
.f4_offset
+ (i
- 2) * 8);
7197 t
= MAX (t
, cfun_frame_layout
.f4_offset
+ (i
- 1) * 8);
7204 /* Fill cfun->machine with info about register usage of current function.
7205 Return in CLOBBERED_REGS which GPRs are currently considered set. */
7208 s390_register_info (int clobbered_regs
[])
7212 /* fprs 8 - 15 are call saved for 64 Bit ABI. */
7213 cfun_frame_layout
.fpr_bitmap
= 0;
7214 cfun_frame_layout
.high_fprs
= 0;
7216 for (i
= 24; i
< 32; i
++)
7217 if (df_regs_ever_live_p (i
) && !global_regs
[i
])
7219 cfun_set_fpr_bit (i
- 16);
7220 cfun_frame_layout
.high_fprs
++;
7223 /* Find first and last gpr to be saved. We trust regs_ever_live
7224 data, except that we don't save and restore global registers.
7226 Also, all registers with special meaning to the compiler need
7227 to be handled extra. */
7229 s390_regs_ever_clobbered (clobbered_regs
);
7231 for (i
= 0; i
< 16; i
++)
7232 clobbered_regs
[i
] = clobbered_regs
[i
] && !global_regs
[i
] && !fixed_regs
[i
];
7234 if (frame_pointer_needed
)
7235 clobbered_regs
[HARD_FRAME_POINTER_REGNUM
] = 1;
7238 clobbered_regs
[PIC_OFFSET_TABLE_REGNUM
]
7239 |= df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM
);
7241 clobbered_regs
[BASE_REGNUM
]
7242 |= (cfun
->machine
->base_reg
7243 && REGNO (cfun
->machine
->base_reg
) == BASE_REGNUM
);
7245 clobbered_regs
[RETURN_REGNUM
]
7246 |= (!current_function_is_leaf
7247 || TARGET_TPF_PROFILING
7248 || cfun
->machine
->split_branches_pending_p
7249 || cfun_frame_layout
.save_return_addr_p
7250 || crtl
->calls_eh_return
7253 clobbered_regs
[STACK_POINTER_REGNUM
]
7254 |= (!current_function_is_leaf
7255 || TARGET_TPF_PROFILING
7256 || cfun_save_high_fprs_p
7257 || get_frame_size () > 0
7258 || cfun
->calls_alloca
7261 for (i
= 6; i
< 16; i
++)
7262 if (df_regs_ever_live_p (i
) || clobbered_regs
[i
])
7264 for (j
= 15; j
> i
; j
--)
7265 if (df_regs_ever_live_p (j
) || clobbered_regs
[j
])
7270 /* Nothing to save/restore. */
7271 cfun_frame_layout
.first_save_gpr_slot
= -1;
7272 cfun_frame_layout
.last_save_gpr_slot
= -1;
7273 cfun_frame_layout
.first_save_gpr
= -1;
7274 cfun_frame_layout
.first_restore_gpr
= -1;
7275 cfun_frame_layout
.last_save_gpr
= -1;
7276 cfun_frame_layout
.last_restore_gpr
= -1;
7280 /* Save slots for gprs from i to j. */
7281 cfun_frame_layout
.first_save_gpr_slot
= i
;
7282 cfun_frame_layout
.last_save_gpr_slot
= j
;
7284 for (i
= cfun_frame_layout
.first_save_gpr_slot
;
7285 i
< cfun_frame_layout
.last_save_gpr_slot
+ 1;
7287 if (clobbered_regs
[i
])
7290 for (j
= cfun_frame_layout
.last_save_gpr_slot
; j
> i
; j
--)
7291 if (clobbered_regs
[j
])
7294 if (i
== cfun_frame_layout
.last_save_gpr_slot
+ 1)
7296 /* Nothing to save/restore. */
7297 cfun_frame_layout
.first_save_gpr
= -1;
7298 cfun_frame_layout
.first_restore_gpr
= -1;
7299 cfun_frame_layout
.last_save_gpr
= -1;
7300 cfun_frame_layout
.last_restore_gpr
= -1;
7304 /* Save / Restore from gpr i to j. */
7305 cfun_frame_layout
.first_save_gpr
= i
;
7306 cfun_frame_layout
.first_restore_gpr
= i
;
7307 cfun_frame_layout
.last_save_gpr
= j
;
7308 cfun_frame_layout
.last_restore_gpr
= j
;
7314 /* Varargs functions need to save gprs 2 to 6. */
7315 if (cfun
->va_list_gpr_size
7316 && crtl
->args
.info
.gprs
< GP_ARG_NUM_REG
)
7318 int min_gpr
= crtl
->args
.info
.gprs
;
7319 int max_gpr
= min_gpr
+ cfun
->va_list_gpr_size
;
7320 if (max_gpr
> GP_ARG_NUM_REG
)
7321 max_gpr
= GP_ARG_NUM_REG
;
7323 if (cfun_frame_layout
.first_save_gpr
== -1
7324 || cfun_frame_layout
.first_save_gpr
> 2 + min_gpr
)
7326 cfun_frame_layout
.first_save_gpr
= 2 + min_gpr
;
7327 cfun_frame_layout
.first_save_gpr_slot
= 2 + min_gpr
;
7330 if (cfun_frame_layout
.last_save_gpr
== -1
7331 || cfun_frame_layout
.last_save_gpr
< 2 + max_gpr
- 1)
7333 cfun_frame_layout
.last_save_gpr
= 2 + max_gpr
- 1;
7334 cfun_frame_layout
.last_save_gpr_slot
= 2 + max_gpr
- 1;
7338 /* Mark f0, f2 for 31 bit and f0-f4 for 64 bit to be saved. */
7339 if (TARGET_HARD_FLOAT
&& cfun
->va_list_fpr_size
7340 && crtl
->args
.info
.fprs
< FP_ARG_NUM_REG
)
7342 int min_fpr
= crtl
->args
.info
.fprs
;
7343 int max_fpr
= min_fpr
+ cfun
->va_list_fpr_size
;
7344 if (max_fpr
> FP_ARG_NUM_REG
)
7345 max_fpr
= FP_ARG_NUM_REG
;
7347 /* ??? This is currently required to ensure proper location
7348 of the fpr save slots within the va_list save area. */
7349 if (TARGET_PACKED_STACK
)
7352 for (i
= min_fpr
; i
< max_fpr
; i
++)
7353 cfun_set_fpr_bit (i
);
7358 for (i
= 2; i
< 4; i
++)
7359 if (df_regs_ever_live_p (i
+ 16) && !global_regs
[i
+ 16])
7360 cfun_set_fpr_bit (i
);
7363 /* Fill cfun->machine with info about frame of current function. */
7366 s390_frame_info (void)
7370 cfun_frame_layout
.frame_size
= get_frame_size ();
7371 if (!TARGET_64BIT
&& cfun_frame_layout
.frame_size
> 0x7fff0000)
7372 fatal_error ("total size of local variables exceeds architecture limit");
7374 if (!TARGET_PACKED_STACK
)
7376 cfun_frame_layout
.backchain_offset
= 0;
7377 cfun_frame_layout
.f0_offset
= 16 * UNITS_PER_LONG
;
7378 cfun_frame_layout
.f4_offset
= cfun_frame_layout
.f0_offset
+ 2 * 8;
7379 cfun_frame_layout
.f8_offset
= -cfun_frame_layout
.high_fprs
* 8;
7380 cfun_frame_layout
.gprs_offset
= (cfun_frame_layout
.first_save_gpr_slot
7383 else if (TARGET_BACKCHAIN
) /* kernel stack layout */
7385 cfun_frame_layout
.backchain_offset
= (STACK_POINTER_OFFSET
7387 cfun_frame_layout
.gprs_offset
7388 = (cfun_frame_layout
.backchain_offset
7389 - (STACK_POINTER_REGNUM
- cfun_frame_layout
.first_save_gpr_slot
+ 1)
7394 cfun_frame_layout
.f4_offset
7395 = (cfun_frame_layout
.gprs_offset
7396 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7398 cfun_frame_layout
.f0_offset
7399 = (cfun_frame_layout
.f4_offset
7400 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7404 /* On 31 bit we have to care about alignment of the
7405 floating point regs to provide fastest access. */
7406 cfun_frame_layout
.f0_offset
7407 = ((cfun_frame_layout
.gprs_offset
7408 & ~(STACK_BOUNDARY
/ BITS_PER_UNIT
- 1))
7409 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7411 cfun_frame_layout
.f4_offset
7412 = (cfun_frame_layout
.f0_offset
7413 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7416 else /* no backchain */
7418 cfun_frame_layout
.f4_offset
7419 = (STACK_POINTER_OFFSET
7420 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7422 cfun_frame_layout
.f0_offset
7423 = (cfun_frame_layout
.f4_offset
7424 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7426 cfun_frame_layout
.gprs_offset
7427 = cfun_frame_layout
.f0_offset
- cfun_gprs_save_area_size
;
7430 if (current_function_is_leaf
7431 && !TARGET_TPF_PROFILING
7432 && cfun_frame_layout
.frame_size
== 0
7433 && !cfun_save_high_fprs_p
7434 && !cfun
->calls_alloca
7438 if (!TARGET_PACKED_STACK
)
7439 cfun_frame_layout
.frame_size
+= (STACK_POINTER_OFFSET
7440 + crtl
->outgoing_args_size
7441 + cfun_frame_layout
.high_fprs
* 8);
7444 if (TARGET_BACKCHAIN
)
7445 cfun_frame_layout
.frame_size
+= UNITS_PER_LONG
;
7447 /* No alignment trouble here because f8-f15 are only saved under
7449 cfun_frame_layout
.f8_offset
= (MIN (MIN (cfun_frame_layout
.f0_offset
,
7450 cfun_frame_layout
.f4_offset
),
7451 cfun_frame_layout
.gprs_offset
)
7452 - cfun_frame_layout
.high_fprs
* 8);
7454 cfun_frame_layout
.frame_size
+= cfun_frame_layout
.high_fprs
* 8;
7456 for (i
= 0; i
< 8; i
++)
7457 if (cfun_fpr_bit_p (i
))
7458 cfun_frame_layout
.frame_size
+= 8;
7460 cfun_frame_layout
.frame_size
+= cfun_gprs_save_area_size
;
7462 /* If under 31 bit an odd number of gprs has to be saved we have to adjust
7463 the frame size to sustain 8 byte alignment of stack frames. */
7464 cfun_frame_layout
.frame_size
= ((cfun_frame_layout
.frame_size
+
7465 STACK_BOUNDARY
/ BITS_PER_UNIT
- 1)
7466 & ~(STACK_BOUNDARY
/ BITS_PER_UNIT
- 1));
7468 cfun_frame_layout
.frame_size
+= crtl
->outgoing_args_size
;
7472 /* Generate frame layout. Fills in register and frame data for the current
7473 function in cfun->machine. This routine can be called multiple times;
7474 it will re-do the complete frame layout every time. */
7477 s390_init_frame_layout (void)
7479 HOST_WIDE_INT frame_size
;
7481 int clobbered_regs
[16];
7483 /* On S/390 machines, we may need to perform branch splitting, which
7484 will require both base and return address register. We have no
7485 choice but to assume we're going to need them until right at the
7486 end of the machine dependent reorg phase. */
7487 if (!TARGET_CPU_ZARCH
)
7488 cfun
->machine
->split_branches_pending_p
= true;
7492 frame_size
= cfun_frame_layout
.frame_size
;
7494 /* Try to predict whether we'll need the base register. */
7495 base_used
= cfun
->machine
->split_branches_pending_p
7496 || crtl
->uses_const_pool
7497 || (!DISP_IN_RANGE (frame_size
)
7498 && !CONST_OK_FOR_K (frame_size
));
7500 /* Decide which register to use as literal pool base. In small
7501 leaf functions, try to use an unused call-clobbered register
7502 as base register to avoid save/restore overhead. */
7504 cfun
->machine
->base_reg
= NULL_RTX
;
7505 else if (current_function_is_leaf
&& !df_regs_ever_live_p (5))
7506 cfun
->machine
->base_reg
= gen_rtx_REG (Pmode
, 5);
7508 cfun
->machine
->base_reg
= gen_rtx_REG (Pmode
, BASE_REGNUM
);
7510 s390_register_info (clobbered_regs
);
7513 while (frame_size
!= cfun_frame_layout
.frame_size
);
7516 /* Update frame layout. Recompute actual register save data based on
7517 current info and update regs_ever_live for the special registers.
7518 May be called multiple times, but may never cause *more* registers
7519 to be saved than s390_init_frame_layout allocated room for. */
7522 s390_update_frame_layout (void)
7524 int clobbered_regs
[16];
7526 s390_register_info (clobbered_regs
);
7528 df_set_regs_ever_live (BASE_REGNUM
,
7529 clobbered_regs
[BASE_REGNUM
] ? true : false);
7530 df_set_regs_ever_live (RETURN_REGNUM
,
7531 clobbered_regs
[RETURN_REGNUM
] ? true : false);
7532 df_set_regs_ever_live (STACK_POINTER_REGNUM
,
7533 clobbered_regs
[STACK_POINTER_REGNUM
] ? true : false);
7535 if (cfun
->machine
->base_reg
)
7536 df_set_regs_ever_live (REGNO (cfun
->machine
->base_reg
), true);
7539 /* Return true if it is legal to put a value with MODE into REGNO. */
7542 s390_hard_regno_mode_ok (unsigned int regno
, enum machine_mode mode
)
7544 switch (REGNO_REG_CLASS (regno
))
7547 if (REGNO_PAIR_OK (regno
, mode
))
7549 if (mode
== SImode
|| mode
== DImode
)
7552 if (FLOAT_MODE_P (mode
) && GET_MODE_CLASS (mode
) != MODE_VECTOR_FLOAT
)
7557 if (FRAME_REGNO_P (regno
) && mode
== Pmode
)
7562 if (REGNO_PAIR_OK (regno
, mode
))
7565 || (mode
!= TFmode
&& mode
!= TCmode
&& mode
!= TDmode
))
7570 if (GET_MODE_CLASS (mode
) == MODE_CC
)
7574 if (REGNO_PAIR_OK (regno
, mode
))
7576 if (mode
== SImode
|| mode
== Pmode
)
7587 /* Return nonzero if register OLD_REG can be renamed to register NEW_REG. */
7590 s390_hard_regno_rename_ok (unsigned int old_reg
, unsigned int new_reg
)
7592 /* Once we've decided upon a register to use as base register, it must
7593 no longer be used for any other purpose. */
7594 if (cfun
->machine
->base_reg
)
7595 if (REGNO (cfun
->machine
->base_reg
) == old_reg
7596 || REGNO (cfun
->machine
->base_reg
) == new_reg
)
7602 /* Maximum number of registers to represent a value of mode MODE
7603 in a register of class RCLASS. */
7606 s390_class_max_nregs (enum reg_class rclass
, enum machine_mode mode
)
7611 if (GET_MODE_CLASS (mode
) == MODE_COMPLEX_FLOAT
)
7612 return 2 * ((GET_MODE_SIZE (mode
) / 2 + 8 - 1) / 8);
7614 return (GET_MODE_SIZE (mode
) + 8 - 1) / 8;
7616 return (GET_MODE_SIZE (mode
) + 4 - 1) / 4;
7620 return (GET_MODE_SIZE (mode
) + UNITS_PER_WORD
- 1) / UNITS_PER_WORD
;
7623 /* Return true if register FROM can be eliminated via register TO. */
7626 s390_can_eliminate (const int from
, const int to
)
7628 /* On zSeries machines, we have not marked the base register as fixed.
7629 Instead, we have an elimination rule BASE_REGNUM -> BASE_REGNUM.
7630 If a function requires the base register, we say here that this
7631 elimination cannot be performed. This will cause reload to free
7632 up the base register (as if it were fixed). On the other hand,
7633 if the current function does *not* require the base register, we
7634 say here the elimination succeeds, which in turn allows reload
7635 to allocate the base register for any other purpose. */
7636 if (from
== BASE_REGNUM
&& to
== BASE_REGNUM
)
7638 if (TARGET_CPU_ZARCH
)
7640 s390_init_frame_layout ();
7641 return cfun
->machine
->base_reg
== NULL_RTX
;
7647 /* Everything else must point into the stack frame. */
7648 gcc_assert (to
== STACK_POINTER_REGNUM
7649 || to
== HARD_FRAME_POINTER_REGNUM
);
7651 gcc_assert (from
== FRAME_POINTER_REGNUM
7652 || from
== ARG_POINTER_REGNUM
7653 || from
== RETURN_ADDRESS_POINTER_REGNUM
);
7655 /* Make sure we actually saved the return address. */
7656 if (from
== RETURN_ADDRESS_POINTER_REGNUM
)
7657 if (!crtl
->calls_eh_return
7659 && !cfun_frame_layout
.save_return_addr_p
)
7665 /* Return offset between register FROM and TO initially after prolog. */
7668 s390_initial_elimination_offset (int from
, int to
)
7670 HOST_WIDE_INT offset
;
7673 /* ??? Why are we called for non-eliminable pairs? */
7674 if (!s390_can_eliminate (from
, to
))
7679 case FRAME_POINTER_REGNUM
:
7680 offset
= (get_frame_size()
7681 + STACK_POINTER_OFFSET
7682 + crtl
->outgoing_args_size
);
7685 case ARG_POINTER_REGNUM
:
7686 s390_init_frame_layout ();
7687 offset
= cfun_frame_layout
.frame_size
+ STACK_POINTER_OFFSET
;
7690 case RETURN_ADDRESS_POINTER_REGNUM
:
7691 s390_init_frame_layout ();
7692 index
= RETURN_REGNUM
- cfun_frame_layout
.first_save_gpr_slot
;
7693 gcc_assert (index
>= 0);
7694 offset
= cfun_frame_layout
.frame_size
+ cfun_frame_layout
.gprs_offset
;
7695 offset
+= index
* UNITS_PER_LONG
;
7709 /* Emit insn to save fpr REGNUM at offset OFFSET relative
7710 to register BASE. Return generated insn. */
7713 save_fpr (rtx base
, int offset
, int regnum
)
7716 addr
= gen_rtx_MEM (DFmode
, plus_constant (Pmode
, base
, offset
));
7718 if (regnum
>= 16 && regnum
<= (16 + FP_ARG_NUM_REG
))
7719 set_mem_alias_set (addr
, get_varargs_alias_set ());
7721 set_mem_alias_set (addr
, get_frame_alias_set ());
7723 return emit_move_insn (addr
, gen_rtx_REG (DFmode
, regnum
));
7726 /* Emit insn to restore fpr REGNUM from offset OFFSET relative
7727 to register BASE. Return generated insn. */
7730 restore_fpr (rtx base
, int offset
, int regnum
)
7733 addr
= gen_rtx_MEM (DFmode
, plus_constant (Pmode
, base
, offset
));
7734 set_mem_alias_set (addr
, get_frame_alias_set ());
7736 return emit_move_insn (gen_rtx_REG (DFmode
, regnum
), addr
);
7739 /* Return true if REGNO is a global register, but not one
7740 of the special ones that need to be saved/restored in anyway. */
7743 global_not_special_regno_p (int regno
)
7745 return (global_regs
[regno
]
7746 /* These registers are special and need to be
7747 restored in any case. */
7748 && !(regno
== STACK_POINTER_REGNUM
7749 || regno
== RETURN_REGNUM
7750 || regno
== BASE_REGNUM
7751 || (flag_pic
&& regno
== (int)PIC_OFFSET_TABLE_REGNUM
)));
7754 /* Generate insn to save registers FIRST to LAST into
7755 the register save area located at offset OFFSET
7756 relative to register BASE. */
7759 save_gprs (rtx base
, int offset
, int first
, int last
)
7761 rtx addr
, insn
, note
;
7764 addr
= plus_constant (Pmode
, base
, offset
);
7765 addr
= gen_rtx_MEM (Pmode
, addr
);
7767 set_mem_alias_set (addr
, get_frame_alias_set ());
7769 /* Special-case single register. */
7773 insn
= gen_movdi (addr
, gen_rtx_REG (Pmode
, first
));
7775 insn
= gen_movsi (addr
, gen_rtx_REG (Pmode
, first
));
7777 if (!global_not_special_regno_p (first
))
7778 RTX_FRAME_RELATED_P (insn
) = 1;
7783 insn
= gen_store_multiple (addr
,
7784 gen_rtx_REG (Pmode
, first
),
7785 GEN_INT (last
- first
+ 1));
7787 if (first
<= 6 && cfun
->stdarg
)
7788 for (i
= 0; i
< XVECLEN (PATTERN (insn
), 0); i
++)
7790 rtx mem
= XEXP (XVECEXP (PATTERN (insn
), 0, i
), 0);
7793 set_mem_alias_set (mem
, get_varargs_alias_set ());
7796 /* We need to set the FRAME_RELATED flag on all SETs
7797 inside the store-multiple pattern.
7799 However, we must not emit DWARF records for registers 2..5
7800 if they are stored for use by variable arguments ...
7802 ??? Unfortunately, it is not enough to simply not the
7803 FRAME_RELATED flags for those SETs, because the first SET
7804 of the PARALLEL is always treated as if it had the flag
7805 set, even if it does not. Therefore we emit a new pattern
7806 without those registers as REG_FRAME_RELATED_EXPR note. */
7808 if (first
>= 6 && !global_not_special_regno_p (first
))
7810 rtx pat
= PATTERN (insn
);
7812 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
7813 if (GET_CODE (XVECEXP (pat
, 0, i
)) == SET
7814 && !global_not_special_regno_p (REGNO (SET_SRC (XVECEXP (pat
,
7816 RTX_FRAME_RELATED_P (XVECEXP (pat
, 0, i
)) = 1;
7818 RTX_FRAME_RELATED_P (insn
) = 1;
7824 for (start
= first
>= 6 ? first
: 6; start
<= last
; start
++)
7825 if (!global_not_special_regno_p (start
))
7831 addr
= plus_constant (Pmode
, base
,
7832 offset
+ (start
- first
) * UNITS_PER_LONG
);
7833 note
= gen_store_multiple (gen_rtx_MEM (Pmode
, addr
),
7834 gen_rtx_REG (Pmode
, start
),
7835 GEN_INT (last
- start
+ 1));
7836 note
= PATTERN (note
);
7838 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
, note
);
7840 for (i
= 0; i
< XVECLEN (note
, 0); i
++)
7841 if (GET_CODE (XVECEXP (note
, 0, i
)) == SET
7842 && !global_not_special_regno_p (REGNO (SET_SRC (XVECEXP (note
,
7844 RTX_FRAME_RELATED_P (XVECEXP (note
, 0, i
)) = 1;
7846 RTX_FRAME_RELATED_P (insn
) = 1;
7852 /* Generate insn to restore registers FIRST to LAST from
7853 the register save area located at offset OFFSET
7854 relative to register BASE. */
7857 restore_gprs (rtx base
, int offset
, int first
, int last
)
7861 addr
= plus_constant (Pmode
, base
, offset
);
7862 addr
= gen_rtx_MEM (Pmode
, addr
);
7863 set_mem_alias_set (addr
, get_frame_alias_set ());
7865 /* Special-case single register. */
7869 insn
= gen_movdi (gen_rtx_REG (Pmode
, first
), addr
);
7871 insn
= gen_movsi (gen_rtx_REG (Pmode
, first
), addr
);
7876 insn
= gen_load_multiple (gen_rtx_REG (Pmode
, first
),
7878 GEN_INT (last
- first
+ 1));
7882 /* Return insn sequence to load the GOT register. */
7884 static GTY(()) rtx got_symbol
;
7886 s390_load_got (void)
7890 /* We cannot use pic_offset_table_rtx here since we use this
7891 function also for non-pic if __tls_get_offset is called and in
7892 that case PIC_OFFSET_TABLE_REGNUM as well as pic_offset_table_rtx
7894 rtx got_rtx
= gen_rtx_REG (Pmode
, 12);
7898 got_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "_GLOBAL_OFFSET_TABLE_");
7899 SYMBOL_REF_FLAGS (got_symbol
) = SYMBOL_FLAG_LOCAL
;
7904 if (TARGET_CPU_ZARCH
)
7906 emit_move_insn (got_rtx
, got_symbol
);
7912 offset
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, got_symbol
),
7913 UNSPEC_LTREL_OFFSET
);
7914 offset
= gen_rtx_CONST (Pmode
, offset
);
7915 offset
= force_const_mem (Pmode
, offset
);
7917 emit_move_insn (got_rtx
, offset
);
7919 offset
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, XEXP (offset
, 0)),
7921 offset
= gen_rtx_PLUS (Pmode
, got_rtx
, offset
);
7923 emit_move_insn (got_rtx
, offset
);
7926 insns
= get_insns ();
7931 /* This ties together stack memory (MEM with an alias set of frame_alias_set)
7932 and the change to the stack pointer. */
7935 s390_emit_stack_tie (void)
7937 rtx mem
= gen_frame_mem (BLKmode
,
7938 gen_rtx_REG (Pmode
, STACK_POINTER_REGNUM
));
7940 emit_insn (gen_stack_tie (mem
));
7943 /* Expand the prologue into a bunch of separate insns. */
7946 s390_emit_prologue (void)
7954 /* Complete frame layout. */
7956 s390_update_frame_layout ();
7958 /* Annotate all constant pool references to let the scheduler know
7959 they implicitly use the base register. */
7961 push_topmost_sequence ();
7963 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
7966 annotate_constant_pool_refs (&PATTERN (insn
));
7967 df_insn_rescan (insn
);
7970 pop_topmost_sequence ();
7972 /* Choose best register to use for temp use within prologue.
7973 See below for why TPF must use the register 1. */
7975 if (!has_hard_reg_initial_val (Pmode
, RETURN_REGNUM
)
7976 && !current_function_is_leaf
7977 && !TARGET_TPF_PROFILING
)
7978 temp_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
7980 temp_reg
= gen_rtx_REG (Pmode
, 1);
7982 /* Save call saved gprs. */
7983 if (cfun_frame_layout
.first_save_gpr
!= -1)
7985 insn
= save_gprs (stack_pointer_rtx
,
7986 cfun_frame_layout
.gprs_offset
+
7987 UNITS_PER_LONG
* (cfun_frame_layout
.first_save_gpr
7988 - cfun_frame_layout
.first_save_gpr_slot
),
7989 cfun_frame_layout
.first_save_gpr
,
7990 cfun_frame_layout
.last_save_gpr
);
7994 /* Dummy insn to mark literal pool slot. */
7996 if (cfun
->machine
->base_reg
)
7997 emit_insn (gen_main_pool (cfun
->machine
->base_reg
));
7999 offset
= cfun_frame_layout
.f0_offset
;
8001 /* Save f0 and f2. */
8002 for (i
= 0; i
< 2; i
++)
8004 if (cfun_fpr_bit_p (i
))
8006 save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8009 else if (!TARGET_PACKED_STACK
)
8013 /* Save f4 and f6. */
8014 offset
= cfun_frame_layout
.f4_offset
;
8015 for (i
= 2; i
< 4; i
++)
8017 if (cfun_fpr_bit_p (i
))
8019 insn
= save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8022 /* If f4 and f6 are call clobbered they are saved due to stdargs and
8023 therefore are not frame related. */
8024 if (!call_really_used_regs
[i
+ 16])
8025 RTX_FRAME_RELATED_P (insn
) = 1;
8027 else if (!TARGET_PACKED_STACK
)
8031 if (TARGET_PACKED_STACK
8032 && cfun_save_high_fprs_p
8033 && cfun_frame_layout
.f8_offset
+ cfun_frame_layout
.high_fprs
* 8 > 0)
8035 offset
= (cfun_frame_layout
.f8_offset
8036 + (cfun_frame_layout
.high_fprs
- 1) * 8);
8038 for (i
= 15; i
> 7 && offset
>= 0; i
--)
8039 if (cfun_fpr_bit_p (i
))
8041 insn
= save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8043 RTX_FRAME_RELATED_P (insn
) = 1;
8046 if (offset
>= cfun_frame_layout
.f8_offset
)
8050 if (!TARGET_PACKED_STACK
)
8051 next_fpr
= cfun_save_high_fprs_p
? 31 : 0;
8053 if (flag_stack_usage_info
)
8054 current_function_static_stack_size
= cfun_frame_layout
.frame_size
;
8056 /* Decrement stack pointer. */
8058 if (cfun_frame_layout
.frame_size
> 0)
8060 rtx frame_off
= GEN_INT (-cfun_frame_layout
.frame_size
);
8063 if (s390_stack_size
)
8065 HOST_WIDE_INT stack_guard
;
8067 if (s390_stack_guard
)
8068 stack_guard
= s390_stack_guard
;
8071 /* If no value for stack guard is provided the smallest power of 2
8072 larger than the current frame size is chosen. */
8074 while (stack_guard
< cfun_frame_layout
.frame_size
)
8078 if (cfun_frame_layout
.frame_size
>= s390_stack_size
)
8080 warning (0, "frame size of function %qs is %wd"
8081 " bytes exceeding user provided stack limit of "
8083 "An unconditional trap is added.",
8084 current_function_name(), cfun_frame_layout
.frame_size
,
8086 emit_insn (gen_trap ());
8090 /* stack_guard has to be smaller than s390_stack_size.
8091 Otherwise we would emit an AND with zero which would
8092 not match the test under mask pattern. */
8093 if (stack_guard
>= s390_stack_size
)
8095 warning (0, "frame size of function %qs is %wd"
8096 " bytes which is more than half the stack size. "
8097 "The dynamic check would not be reliable. "
8098 "No check emitted for this function.",
8099 current_function_name(),
8100 cfun_frame_layout
.frame_size
);
8104 HOST_WIDE_INT stack_check_mask
= ((s390_stack_size
- 1)
8105 & ~(stack_guard
- 1));
8107 rtx t
= gen_rtx_AND (Pmode
, stack_pointer_rtx
,
8108 GEN_INT (stack_check_mask
));
8110 emit_insn (gen_ctrapdi4 (gen_rtx_EQ (VOIDmode
,
8112 t
, const0_rtx
, const0_rtx
));
8114 emit_insn (gen_ctrapsi4 (gen_rtx_EQ (VOIDmode
,
8116 t
, const0_rtx
, const0_rtx
));
8121 if (s390_warn_framesize
> 0
8122 && cfun_frame_layout
.frame_size
>= s390_warn_framesize
)
8123 warning (0, "frame size of %qs is %wd bytes",
8124 current_function_name (), cfun_frame_layout
.frame_size
);
8126 if (s390_warn_dynamicstack_p
&& cfun
->calls_alloca
)
8127 warning (0, "%qs uses dynamic stack allocation", current_function_name ());
8129 /* Save incoming stack pointer into temp reg. */
8130 if (TARGET_BACKCHAIN
|| next_fpr
)
8131 insn
= emit_insn (gen_move_insn (temp_reg
, stack_pointer_rtx
));
8133 /* Subtract frame size from stack pointer. */
8135 if (DISP_IN_RANGE (INTVAL (frame_off
)))
8137 insn
= gen_rtx_SET (VOIDmode
, stack_pointer_rtx
,
8138 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
8140 insn
= emit_insn (insn
);
8144 if (!CONST_OK_FOR_K (INTVAL (frame_off
)))
8145 frame_off
= force_const_mem (Pmode
, frame_off
);
8147 insn
= emit_insn (gen_add2_insn (stack_pointer_rtx
, frame_off
));
8148 annotate_constant_pool_refs (&PATTERN (insn
));
8151 RTX_FRAME_RELATED_P (insn
) = 1;
8152 real_frame_off
= GEN_INT (-cfun_frame_layout
.frame_size
);
8153 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
8154 gen_rtx_SET (VOIDmode
, stack_pointer_rtx
,
8155 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
8158 /* Set backchain. */
8160 if (TARGET_BACKCHAIN
)
8162 if (cfun_frame_layout
.backchain_offset
)
8163 addr
= gen_rtx_MEM (Pmode
,
8164 plus_constant (Pmode
, stack_pointer_rtx
,
8165 cfun_frame_layout
.backchain_offset
));
8167 addr
= gen_rtx_MEM (Pmode
, stack_pointer_rtx
);
8168 set_mem_alias_set (addr
, get_frame_alias_set ());
8169 insn
= emit_insn (gen_move_insn (addr
, temp_reg
));
8172 /* If we support non-call exceptions (e.g. for Java),
8173 we need to make sure the backchain pointer is set up
8174 before any possibly trapping memory access. */
8175 if (TARGET_BACKCHAIN
&& cfun
->can_throw_non_call_exceptions
)
8177 addr
= gen_rtx_MEM (BLKmode
, gen_rtx_SCRATCH (VOIDmode
));
8178 emit_clobber (addr
);
8182 /* Save fprs 8 - 15 (64 bit ABI). */
8184 if (cfun_save_high_fprs_p
&& next_fpr
)
8186 /* If the stack might be accessed through a different register
8187 we have to make sure that the stack pointer decrement is not
8188 moved below the use of the stack slots. */
8189 s390_emit_stack_tie ();
8191 insn
= emit_insn (gen_add2_insn (temp_reg
,
8192 GEN_INT (cfun_frame_layout
.f8_offset
)));
8196 for (i
= 24; i
<= next_fpr
; i
++)
8197 if (cfun_fpr_bit_p (i
- 16))
8199 rtx addr
= plus_constant (Pmode
, stack_pointer_rtx
,
8200 cfun_frame_layout
.frame_size
8201 + cfun_frame_layout
.f8_offset
8204 insn
= save_fpr (temp_reg
, offset
, i
);
8206 RTX_FRAME_RELATED_P (insn
) = 1;
8207 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
8208 gen_rtx_SET (VOIDmode
,
8209 gen_rtx_MEM (DFmode
, addr
),
8210 gen_rtx_REG (DFmode
, i
)));
8214 /* Set frame pointer, if needed. */
8216 if (frame_pointer_needed
)
8218 insn
= emit_move_insn (hard_frame_pointer_rtx
, stack_pointer_rtx
);
8219 RTX_FRAME_RELATED_P (insn
) = 1;
8222 /* Set up got pointer, if needed. */
8224 if (flag_pic
&& df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM
))
8226 rtx insns
= s390_load_got ();
8228 for (insn
= insns
; insn
; insn
= NEXT_INSN (insn
))
8229 annotate_constant_pool_refs (&PATTERN (insn
));
8234 if (TARGET_TPF_PROFILING
)
8236 /* Generate a BAS instruction to serve as a function
8237 entry intercept to facilitate the use of tracing
8238 algorithms located at the branch target. */
8239 emit_insn (gen_prologue_tpf ());
8241 /* Emit a blockage here so that all code
8242 lies between the profiling mechanisms. */
8243 emit_insn (gen_blockage ());
8247 /* Expand the epilogue into a bunch of separate insns. */
8250 s390_emit_epilogue (bool sibcall
)
8252 rtx frame_pointer
, return_reg
, cfa_restores
= NULL_RTX
;
8253 int area_bottom
, area_top
, offset
= 0;
8258 if (TARGET_TPF_PROFILING
)
8261 /* Generate a BAS instruction to serve as a function
8262 entry intercept to facilitate the use of tracing
8263 algorithms located at the branch target. */
8265 /* Emit a blockage here so that all code
8266 lies between the profiling mechanisms. */
8267 emit_insn (gen_blockage ());
8269 emit_insn (gen_epilogue_tpf ());
8272 /* Check whether to use frame or stack pointer for restore. */
8274 frame_pointer
= (frame_pointer_needed
8275 ? hard_frame_pointer_rtx
: stack_pointer_rtx
);
8277 s390_frame_area (&area_bottom
, &area_top
);
8279 /* Check whether we can access the register save area.
8280 If not, increment the frame pointer as required. */
8282 if (area_top
<= area_bottom
)
8284 /* Nothing to restore. */
8286 else if (DISP_IN_RANGE (cfun_frame_layout
.frame_size
+ area_bottom
)
8287 && DISP_IN_RANGE (cfun_frame_layout
.frame_size
+ area_top
- 1))
8289 /* Area is in range. */
8290 offset
= cfun_frame_layout
.frame_size
;
8294 rtx insn
, frame_off
, cfa
;
8296 offset
= area_bottom
< 0 ? -area_bottom
: 0;
8297 frame_off
= GEN_INT (cfun_frame_layout
.frame_size
- offset
);
8299 cfa
= gen_rtx_SET (VOIDmode
, frame_pointer
,
8300 gen_rtx_PLUS (Pmode
, frame_pointer
, frame_off
));
8301 if (DISP_IN_RANGE (INTVAL (frame_off
)))
8303 insn
= gen_rtx_SET (VOIDmode
, frame_pointer
,
8304 gen_rtx_PLUS (Pmode
, frame_pointer
, frame_off
));
8305 insn
= emit_insn (insn
);
8309 if (!CONST_OK_FOR_K (INTVAL (frame_off
)))
8310 frame_off
= force_const_mem (Pmode
, frame_off
);
8312 insn
= emit_insn (gen_add2_insn (frame_pointer
, frame_off
));
8313 annotate_constant_pool_refs (&PATTERN (insn
));
8315 add_reg_note (insn
, REG_CFA_ADJUST_CFA
, cfa
);
8316 RTX_FRAME_RELATED_P (insn
) = 1;
8319 /* Restore call saved fprs. */
8323 if (cfun_save_high_fprs_p
)
8325 next_offset
= cfun_frame_layout
.f8_offset
;
8326 for (i
= 24; i
< 32; i
++)
8328 if (cfun_fpr_bit_p (i
- 16))
8330 restore_fpr (frame_pointer
,
8331 offset
+ next_offset
, i
);
8333 = alloc_reg_note (REG_CFA_RESTORE
,
8334 gen_rtx_REG (DFmode
, i
), cfa_restores
);
8343 next_offset
= cfun_frame_layout
.f4_offset
;
8344 for (i
= 18; i
< 20; i
++)
8346 if (cfun_fpr_bit_p (i
- 16))
8348 restore_fpr (frame_pointer
,
8349 offset
+ next_offset
, i
);
8351 = alloc_reg_note (REG_CFA_RESTORE
,
8352 gen_rtx_REG (DFmode
, i
), cfa_restores
);
8355 else if (!TARGET_PACKED_STACK
)
8361 /* Return register. */
8363 return_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
8365 /* Restore call saved gprs. */
8367 if (cfun_frame_layout
.first_restore_gpr
!= -1)
8372 /* Check for global register and save them
8373 to stack location from where they get restored. */
8375 for (i
= cfun_frame_layout
.first_restore_gpr
;
8376 i
<= cfun_frame_layout
.last_restore_gpr
;
8379 if (global_not_special_regno_p (i
))
8381 addr
= plus_constant (Pmode
, frame_pointer
,
8382 offset
+ cfun_frame_layout
.gprs_offset
8383 + (i
- cfun_frame_layout
.first_save_gpr_slot
)
8385 addr
= gen_rtx_MEM (Pmode
, addr
);
8386 set_mem_alias_set (addr
, get_frame_alias_set ());
8387 emit_move_insn (addr
, gen_rtx_REG (Pmode
, i
));
8391 = alloc_reg_note (REG_CFA_RESTORE
,
8392 gen_rtx_REG (Pmode
, i
), cfa_restores
);
8397 /* Fetch return address from stack before load multiple,
8398 this will do good for scheduling. */
8400 if (cfun_frame_layout
.save_return_addr_p
8401 || (cfun_frame_layout
.first_restore_gpr
< BASE_REGNUM
8402 && cfun_frame_layout
.last_restore_gpr
> RETURN_REGNUM
))
8404 int return_regnum
= find_unused_clobbered_reg();
8407 return_reg
= gen_rtx_REG (Pmode
, return_regnum
);
8409 addr
= plus_constant (Pmode
, frame_pointer
,
8410 offset
+ cfun_frame_layout
.gprs_offset
8412 - cfun_frame_layout
.first_save_gpr_slot
)
8414 addr
= gen_rtx_MEM (Pmode
, addr
);
8415 set_mem_alias_set (addr
, get_frame_alias_set ());
8416 emit_move_insn (return_reg
, addr
);
8420 insn
= restore_gprs (frame_pointer
,
8421 offset
+ cfun_frame_layout
.gprs_offset
8422 + (cfun_frame_layout
.first_restore_gpr
8423 - cfun_frame_layout
.first_save_gpr_slot
)
8425 cfun_frame_layout
.first_restore_gpr
,
8426 cfun_frame_layout
.last_restore_gpr
);
8427 insn
= emit_insn (insn
);
8428 REG_NOTES (insn
) = cfa_restores
;
8429 add_reg_note (insn
, REG_CFA_DEF_CFA
,
8430 plus_constant (Pmode
, stack_pointer_rtx
,
8431 STACK_POINTER_OFFSET
));
8432 RTX_FRAME_RELATED_P (insn
) = 1;
8438 /* Return to caller. */
8440 p
= rtvec_alloc (2);
8442 RTVEC_ELT (p
, 0) = ret_rtx
;
8443 RTVEC_ELT (p
, 1) = gen_rtx_USE (VOIDmode
, return_reg
);
8444 emit_jump_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
8449 /* Return the size in bytes of a function argument of
8450 type TYPE and/or mode MODE. At least one of TYPE or
8451 MODE must be specified. */
8454 s390_function_arg_size (enum machine_mode mode
, const_tree type
)
8457 return int_size_in_bytes (type
);
8459 /* No type info available for some library calls ... */
8460 if (mode
!= BLKmode
)
8461 return GET_MODE_SIZE (mode
);
8463 /* If we have neither type nor mode, abort */
8467 /* Return true if a function argument of type TYPE and mode MODE
8468 is to be passed in a floating-point register, if available. */
8471 s390_function_arg_float (enum machine_mode mode
, const_tree type
)
8473 int size
= s390_function_arg_size (mode
, type
);
8477 /* Soft-float changes the ABI: no floating-point registers are used. */
8478 if (TARGET_SOFT_FLOAT
)
8481 /* No type info available for some library calls ... */
8483 return mode
== SFmode
|| mode
== DFmode
|| mode
== SDmode
|| mode
== DDmode
;
8485 /* The ABI says that record types with a single member are treated
8486 just like that member would be. */
8487 while (TREE_CODE (type
) == RECORD_TYPE
)
8489 tree field
, single
= NULL_TREE
;
8491 for (field
= TYPE_FIELDS (type
); field
; field
= DECL_CHAIN (field
))
8493 if (TREE_CODE (field
) != FIELD_DECL
)
8496 if (single
== NULL_TREE
)
8497 single
= TREE_TYPE (field
);
8502 if (single
== NULL_TREE
)
8508 return TREE_CODE (type
) == REAL_TYPE
;
8511 /* Return true if a function argument of type TYPE and mode MODE
8512 is to be passed in an integer register, or a pair of integer
8513 registers, if available. */
8516 s390_function_arg_integer (enum machine_mode mode
, const_tree type
)
8518 int size
= s390_function_arg_size (mode
, type
);
8522 /* No type info available for some library calls ... */
8524 return GET_MODE_CLASS (mode
) == MODE_INT
8525 || (TARGET_SOFT_FLOAT
&& SCALAR_FLOAT_MODE_P (mode
));
8527 /* We accept small integral (and similar) types. */
8528 if (INTEGRAL_TYPE_P (type
)
8529 || POINTER_TYPE_P (type
)
8530 || TREE_CODE (type
) == NULLPTR_TYPE
8531 || TREE_CODE (type
) == OFFSET_TYPE
8532 || (TARGET_SOFT_FLOAT
&& TREE_CODE (type
) == REAL_TYPE
))
8535 /* We also accept structs of size 1, 2, 4, 8 that are not
8536 passed in floating-point registers. */
8537 if (AGGREGATE_TYPE_P (type
)
8538 && exact_log2 (size
) >= 0
8539 && !s390_function_arg_float (mode
, type
))
8545 /* Return 1 if a function argument of type TYPE and mode MODE
8546 is to be passed by reference. The ABI specifies that only
8547 structures of size 1, 2, 4, or 8 bytes are passed by value,
8548 all other structures (and complex numbers) are passed by
8552 s390_pass_by_reference (cumulative_args_t ca ATTRIBUTE_UNUSED
,
8553 enum machine_mode mode
, const_tree type
,
8554 bool named ATTRIBUTE_UNUSED
)
8556 int size
= s390_function_arg_size (mode
, type
);
8562 if (AGGREGATE_TYPE_P (type
) && exact_log2 (size
) < 0)
8565 if (TREE_CODE (type
) == COMPLEX_TYPE
8566 || TREE_CODE (type
) == VECTOR_TYPE
)
8573 /* Update the data in CUM to advance over an argument of mode MODE and
8574 data type TYPE. (TYPE is null for libcalls where that information
8575 may not be available.). The boolean NAMED specifies whether the
8576 argument is a named argument (as opposed to an unnamed argument
8577 matching an ellipsis). */
8580 s390_function_arg_advance (cumulative_args_t cum_v
, enum machine_mode mode
,
8581 const_tree type
, bool named ATTRIBUTE_UNUSED
)
8583 CUMULATIVE_ARGS
*cum
= get_cumulative_args (cum_v
);
8585 if (s390_function_arg_float (mode
, type
))
8589 else if (s390_function_arg_integer (mode
, type
))
8591 int size
= s390_function_arg_size (mode
, type
);
8592 cum
->gprs
+= ((size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
);
8598 /* Define where to put the arguments to a function.
8599 Value is zero to push the argument on the stack,
8600 or a hard register in which to store the argument.
8602 MODE is the argument's machine mode.
8603 TYPE is the data type of the argument (as a tree).
8604 This is null for libcalls where that information may
8606 CUM is a variable of type CUMULATIVE_ARGS which gives info about
8607 the preceding args and about the function being called.
8608 NAMED is nonzero if this argument is a named parameter
8609 (otherwise it is an extra parameter matching an ellipsis).
8611 On S/390, we use general purpose registers 2 through 6 to
8612 pass integer, pointer, and certain structure arguments, and
8613 floating point registers 0 and 2 (0, 2, 4, and 6 on 64-bit)
8614 to pass floating point arguments. All remaining arguments
8615 are pushed to the stack. */
8618 s390_function_arg (cumulative_args_t cum_v
, enum machine_mode mode
,
8619 const_tree type
, bool named ATTRIBUTE_UNUSED
)
8621 CUMULATIVE_ARGS
*cum
= get_cumulative_args (cum_v
);
8623 if (s390_function_arg_float (mode
, type
))
8625 if (cum
->fprs
+ 1 > FP_ARG_NUM_REG
)
8628 return gen_rtx_REG (mode
, cum
->fprs
+ 16);
8630 else if (s390_function_arg_integer (mode
, type
))
8632 int size
= s390_function_arg_size (mode
, type
);
8633 int n_gprs
= (size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
;
8635 if (cum
->gprs
+ n_gprs
> GP_ARG_NUM_REG
)
8637 else if (n_gprs
== 1 || UNITS_PER_WORD
== UNITS_PER_LONG
)
8638 return gen_rtx_REG (mode
, cum
->gprs
+ 2);
8639 else if (n_gprs
== 2)
8641 rtvec p
= rtvec_alloc (2);
8644 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, cum
->gprs
+ 2),
8647 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, cum
->gprs
+ 3),
8650 return gen_rtx_PARALLEL (mode
, p
);
8654 /* After the real arguments, expand_call calls us once again
8655 with a void_type_node type. Whatever we return here is
8656 passed as operand 2 to the call expanders.
8658 We don't need this feature ... */
8659 else if (type
== void_type_node
)
8665 /* Return true if return values of type TYPE should be returned
8666 in a memory buffer whose address is passed by the caller as
8667 hidden first argument. */
8670 s390_return_in_memory (const_tree type
, const_tree fundecl ATTRIBUTE_UNUSED
)
8672 /* We accept small integral (and similar) types. */
8673 if (INTEGRAL_TYPE_P (type
)
8674 || POINTER_TYPE_P (type
)
8675 || TREE_CODE (type
) == OFFSET_TYPE
8676 || TREE_CODE (type
) == REAL_TYPE
)
8677 return int_size_in_bytes (type
) > 8;
8679 /* Aggregates and similar constructs are always returned
8681 if (AGGREGATE_TYPE_P (type
)
8682 || TREE_CODE (type
) == COMPLEX_TYPE
8683 || TREE_CODE (type
) == VECTOR_TYPE
)
8686 /* ??? We get called on all sorts of random stuff from
8687 aggregate_value_p. We can't abort, but it's not clear
8688 what's safe to return. Pretend it's a struct I guess. */
8692 /* Function arguments and return values are promoted to word size. */
8694 static enum machine_mode
8695 s390_promote_function_mode (const_tree type
, enum machine_mode mode
,
8697 const_tree fntype ATTRIBUTE_UNUSED
,
8698 int for_return ATTRIBUTE_UNUSED
)
8700 if (INTEGRAL_MODE_P (mode
)
8701 && GET_MODE_SIZE (mode
) < UNITS_PER_LONG
)
8703 if (type
!= NULL_TREE
&& POINTER_TYPE_P (type
))
8704 *punsignedp
= POINTERS_EXTEND_UNSIGNED
;
8711 /* Define where to return a (scalar) value of type RET_TYPE.
8712 If RET_TYPE is null, define where to return a (scalar)
8713 value of mode MODE from a libcall. */
8716 s390_function_and_libcall_value (enum machine_mode mode
,
8717 const_tree ret_type
,
8718 const_tree fntype_or_decl
,
8719 bool outgoing ATTRIBUTE_UNUSED
)
8721 /* For normal functions perform the promotion as
8722 promote_function_mode would do. */
8725 int unsignedp
= TYPE_UNSIGNED (ret_type
);
8726 mode
= promote_function_mode (ret_type
, mode
, &unsignedp
,
8730 gcc_assert (GET_MODE_CLASS (mode
) == MODE_INT
|| SCALAR_FLOAT_MODE_P (mode
));
8731 gcc_assert (GET_MODE_SIZE (mode
) <= 8);
8733 if (TARGET_HARD_FLOAT
&& SCALAR_FLOAT_MODE_P (mode
))
8734 return gen_rtx_REG (mode
, 16);
8735 else if (GET_MODE_SIZE (mode
) <= UNITS_PER_LONG
8736 || UNITS_PER_LONG
== UNITS_PER_WORD
)
8737 return gen_rtx_REG (mode
, 2);
8738 else if (GET_MODE_SIZE (mode
) == 2 * UNITS_PER_LONG
)
8740 /* This case is triggered when returning a 64 bit value with
8741 -m31 -mzarch. Although the value would fit into a single
8742 register it has to be forced into a 32 bit register pair in
8743 order to match the ABI. */
8744 rtvec p
= rtvec_alloc (2);
8747 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, 2), const0_rtx
);
8749 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, 3), GEN_INT (4));
8751 return gen_rtx_PARALLEL (mode
, p
);
8757 /* Define where to return a scalar return value of type RET_TYPE. */
8760 s390_function_value (const_tree ret_type
, const_tree fn_decl_or_type
,
8763 return s390_function_and_libcall_value (TYPE_MODE (ret_type
), ret_type
,
8764 fn_decl_or_type
, outgoing
);
8767 /* Define where to return a scalar libcall return value of mode
8771 s390_libcall_value (enum machine_mode mode
, const_rtx fun ATTRIBUTE_UNUSED
)
8773 return s390_function_and_libcall_value (mode
, NULL_TREE
,
8778 /* Create and return the va_list datatype.
8780 On S/390, va_list is an array type equivalent to
8782 typedef struct __va_list_tag
8786 void *__overflow_arg_area;
8787 void *__reg_save_area;
8790 where __gpr and __fpr hold the number of general purpose
8791 or floating point arguments used up to now, respectively,
8792 __overflow_arg_area points to the stack location of the
8793 next argument passed on the stack, and __reg_save_area
8794 always points to the start of the register area in the
8795 call frame of the current function. The function prologue
8796 saves all registers used for argument passing into this
8797 area if the function uses variable arguments. */
8800 s390_build_builtin_va_list (void)
8802 tree f_gpr
, f_fpr
, f_ovf
, f_sav
, record
, type_decl
;
8804 record
= lang_hooks
.types
.make_type (RECORD_TYPE
);
8807 build_decl (BUILTINS_LOCATION
,
8808 TYPE_DECL
, get_identifier ("__va_list_tag"), record
);
8810 f_gpr
= build_decl (BUILTINS_LOCATION
,
8811 FIELD_DECL
, get_identifier ("__gpr"),
8812 long_integer_type_node
);
8813 f_fpr
= build_decl (BUILTINS_LOCATION
,
8814 FIELD_DECL
, get_identifier ("__fpr"),
8815 long_integer_type_node
);
8816 f_ovf
= build_decl (BUILTINS_LOCATION
,
8817 FIELD_DECL
, get_identifier ("__overflow_arg_area"),
8819 f_sav
= build_decl (BUILTINS_LOCATION
,
8820 FIELD_DECL
, get_identifier ("__reg_save_area"),
8823 va_list_gpr_counter_field
= f_gpr
;
8824 va_list_fpr_counter_field
= f_fpr
;
8826 DECL_FIELD_CONTEXT (f_gpr
) = record
;
8827 DECL_FIELD_CONTEXT (f_fpr
) = record
;
8828 DECL_FIELD_CONTEXT (f_ovf
) = record
;
8829 DECL_FIELD_CONTEXT (f_sav
) = record
;
8831 TYPE_STUB_DECL (record
) = type_decl
;
8832 TYPE_NAME (record
) = type_decl
;
8833 TYPE_FIELDS (record
) = f_gpr
;
8834 DECL_CHAIN (f_gpr
) = f_fpr
;
8835 DECL_CHAIN (f_fpr
) = f_ovf
;
8836 DECL_CHAIN (f_ovf
) = f_sav
;
8838 layout_type (record
);
8840 /* The correct type is an array type of one element. */
8841 return build_array_type (record
, build_index_type (size_zero_node
));
8844 /* Implement va_start by filling the va_list structure VALIST.
8845 STDARG_P is always true, and ignored.
8846 NEXTARG points to the first anonymous stack argument.
8848 The following global variables are used to initialize
8849 the va_list structure:
8852 holds number of gprs and fprs used for named arguments.
8853 crtl->args.arg_offset_rtx:
8854 holds the offset of the first anonymous stack argument
8855 (relative to the virtual arg pointer). */
8858 s390_va_start (tree valist
, rtx nextarg ATTRIBUTE_UNUSED
)
8860 HOST_WIDE_INT n_gpr
, n_fpr
;
8862 tree f_gpr
, f_fpr
, f_ovf
, f_sav
;
8863 tree gpr
, fpr
, ovf
, sav
, t
;
8865 f_gpr
= TYPE_FIELDS (TREE_TYPE (va_list_type_node
));
8866 f_fpr
= DECL_CHAIN (f_gpr
);
8867 f_ovf
= DECL_CHAIN (f_fpr
);
8868 f_sav
= DECL_CHAIN (f_ovf
);
8870 valist
= build_simple_mem_ref (valist
);
8871 gpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_gpr
), valist
, f_gpr
, NULL_TREE
);
8872 fpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_fpr
), valist
, f_fpr
, NULL_TREE
);
8873 ovf
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovf
), valist
, f_ovf
, NULL_TREE
);
8874 sav
= build3 (COMPONENT_REF
, TREE_TYPE (f_sav
), valist
, f_sav
, NULL_TREE
);
8876 /* Count number of gp and fp argument registers used. */
8878 n_gpr
= crtl
->args
.info
.gprs
;
8879 n_fpr
= crtl
->args
.info
.fprs
;
8881 if (cfun
->va_list_gpr_size
)
8883 t
= build2 (MODIFY_EXPR
, TREE_TYPE (gpr
), gpr
,
8884 build_int_cst (NULL_TREE
, n_gpr
));
8885 TREE_SIDE_EFFECTS (t
) = 1;
8886 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8889 if (cfun
->va_list_fpr_size
)
8891 t
= build2 (MODIFY_EXPR
, TREE_TYPE (fpr
), fpr
,
8892 build_int_cst (NULL_TREE
, n_fpr
));
8893 TREE_SIDE_EFFECTS (t
) = 1;
8894 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8897 /* Find the overflow area. */
8898 if (n_gpr
+ cfun
->va_list_gpr_size
> GP_ARG_NUM_REG
8899 || n_fpr
+ cfun
->va_list_fpr_size
> FP_ARG_NUM_REG
)
8901 t
= make_tree (TREE_TYPE (ovf
), virtual_incoming_args_rtx
);
8903 off
= INTVAL (crtl
->args
.arg_offset_rtx
);
8904 off
= off
< 0 ? 0 : off
;
8905 if (TARGET_DEBUG_ARG
)
8906 fprintf (stderr
, "va_start: n_gpr = %d, n_fpr = %d off %d\n",
8907 (int)n_gpr
, (int)n_fpr
, off
);
8909 t
= fold_build_pointer_plus_hwi (t
, off
);
8911 t
= build2 (MODIFY_EXPR
, TREE_TYPE (ovf
), ovf
, t
);
8912 TREE_SIDE_EFFECTS (t
) = 1;
8913 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8916 /* Find the register save area. */
8917 if ((cfun
->va_list_gpr_size
&& n_gpr
< GP_ARG_NUM_REG
)
8918 || (cfun
->va_list_fpr_size
&& n_fpr
< FP_ARG_NUM_REG
))
8920 t
= make_tree (TREE_TYPE (sav
), return_address_pointer_rtx
);
8921 t
= fold_build_pointer_plus_hwi (t
, -RETURN_REGNUM
* UNITS_PER_LONG
);
8923 t
= build2 (MODIFY_EXPR
, TREE_TYPE (sav
), sav
, t
);
8924 TREE_SIDE_EFFECTS (t
) = 1;
8925 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8929 /* Implement va_arg by updating the va_list structure
8930 VALIST as required to retrieve an argument of type
8931 TYPE, and returning that argument.
8933 Generates code equivalent to:
8935 if (integral value) {
8936 if (size <= 4 && args.gpr < 5 ||
8937 size > 4 && args.gpr < 4 )
8938 ret = args.reg_save_area[args.gpr+8]
8940 ret = *args.overflow_arg_area++;
8941 } else if (float value) {
8943 ret = args.reg_save_area[args.fpr+64]
8945 ret = *args.overflow_arg_area++;
8946 } else if (aggregate value) {
8948 ret = *args.reg_save_area[args.gpr]
8950 ret = **args.overflow_arg_area++;
8954 s390_gimplify_va_arg (tree valist
, tree type
, gimple_seq
*pre_p
,
8955 gimple_seq
*post_p ATTRIBUTE_UNUSED
)
8957 tree f_gpr
, f_fpr
, f_ovf
, f_sav
;
8958 tree gpr
, fpr
, ovf
, sav
, reg
, t
, u
;
8959 int indirect_p
, size
, n_reg
, sav_ofs
, sav_scale
, max_reg
;
8960 tree lab_false
, lab_over
, addr
;
8962 f_gpr
= TYPE_FIELDS (TREE_TYPE (va_list_type_node
));
8963 f_fpr
= DECL_CHAIN (f_gpr
);
8964 f_ovf
= DECL_CHAIN (f_fpr
);
8965 f_sav
= DECL_CHAIN (f_ovf
);
8967 valist
= build_va_arg_indirect_ref (valist
);
8968 gpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_gpr
), valist
, f_gpr
, NULL_TREE
);
8969 fpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_fpr
), valist
, f_fpr
, NULL_TREE
);
8970 sav
= build3 (COMPONENT_REF
, TREE_TYPE (f_sav
), valist
, f_sav
, NULL_TREE
);
8972 /* The tree for args* cannot be shared between gpr/fpr and ovf since
8973 both appear on a lhs. */
8974 valist
= unshare_expr (valist
);
8975 ovf
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovf
), valist
, f_ovf
, NULL_TREE
);
8977 size
= int_size_in_bytes (type
);
8979 if (pass_by_reference (NULL
, TYPE_MODE (type
), type
, false))
8981 if (TARGET_DEBUG_ARG
)
8983 fprintf (stderr
, "va_arg: aggregate type");
8987 /* Aggregates are passed by reference. */
8992 /* kernel stack layout on 31 bit: It is assumed here that no padding
8993 will be added by s390_frame_info because for va_args always an even
8994 number of gprs has to be saved r15-r2 = 14 regs. */
8995 sav_ofs
= 2 * UNITS_PER_LONG
;
8996 sav_scale
= UNITS_PER_LONG
;
8997 size
= UNITS_PER_LONG
;
8998 max_reg
= GP_ARG_NUM_REG
- n_reg
;
9000 else if (s390_function_arg_float (TYPE_MODE (type
), type
))
9002 if (TARGET_DEBUG_ARG
)
9004 fprintf (stderr
, "va_arg: float type");
9008 /* FP args go in FP registers, if present. */
9012 sav_ofs
= 16 * UNITS_PER_LONG
;
9014 max_reg
= FP_ARG_NUM_REG
- n_reg
;
9018 if (TARGET_DEBUG_ARG
)
9020 fprintf (stderr
, "va_arg: other type");
9024 /* Otherwise into GP registers. */
9027 n_reg
= (size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
;
9029 /* kernel stack layout on 31 bit: It is assumed here that no padding
9030 will be added by s390_frame_info because for va_args always an even
9031 number of gprs has to be saved r15-r2 = 14 regs. */
9032 sav_ofs
= 2 * UNITS_PER_LONG
;
9034 if (size
< UNITS_PER_LONG
)
9035 sav_ofs
+= UNITS_PER_LONG
- size
;
9037 sav_scale
= UNITS_PER_LONG
;
9038 max_reg
= GP_ARG_NUM_REG
- n_reg
;
9041 /* Pull the value out of the saved registers ... */
9043 lab_false
= create_artificial_label (UNKNOWN_LOCATION
);
9044 lab_over
= create_artificial_label (UNKNOWN_LOCATION
);
9045 addr
= create_tmp_var (ptr_type_node
, "addr");
9046 mark_sym_for_renaming (addr
);
9048 t
= fold_convert (TREE_TYPE (reg
), size_int (max_reg
));
9049 t
= build2 (GT_EXPR
, boolean_type_node
, reg
, t
);
9050 u
= build1 (GOTO_EXPR
, void_type_node
, lab_false
);
9051 t
= build3 (COND_EXPR
, void_type_node
, t
, u
, NULL_TREE
);
9052 gimplify_and_add (t
, pre_p
);
9054 t
= fold_build_pointer_plus_hwi (sav
, sav_ofs
);
9055 u
= build2 (MULT_EXPR
, TREE_TYPE (reg
), reg
,
9056 fold_convert (TREE_TYPE (reg
), size_int (sav_scale
)));
9057 t
= fold_build_pointer_plus (t
, u
);
9059 gimplify_assign (addr
, t
, pre_p
);
9061 gimple_seq_add_stmt (pre_p
, gimple_build_goto (lab_over
));
9063 gimple_seq_add_stmt (pre_p
, gimple_build_label (lab_false
));
9066 /* ... Otherwise out of the overflow area. */
9069 if (size
< UNITS_PER_LONG
)
9070 t
= fold_build_pointer_plus_hwi (t
, UNITS_PER_LONG
- size
);
9072 gimplify_expr (&t
, pre_p
, NULL
, is_gimple_val
, fb_rvalue
);
9074 gimplify_assign (addr
, t
, pre_p
);
9076 t
= fold_build_pointer_plus_hwi (t
, size
);
9077 gimplify_assign (ovf
, t
, pre_p
);
9079 gimple_seq_add_stmt (pre_p
, gimple_build_label (lab_over
));
9082 /* Increment register save count. */
9084 u
= build2 (PREINCREMENT_EXPR
, TREE_TYPE (reg
), reg
,
9085 fold_convert (TREE_TYPE (reg
), size_int (n_reg
)));
9086 gimplify_and_add (u
, pre_p
);
9090 t
= build_pointer_type_for_mode (build_pointer_type (type
),
9092 addr
= fold_convert (t
, addr
);
9093 addr
= build_va_arg_indirect_ref (addr
);
9097 t
= build_pointer_type_for_mode (type
, ptr_mode
, true);
9098 addr
= fold_convert (t
, addr
);
9101 return build_va_arg_indirect_ref (addr
);
9109 S390_BUILTIN_THREAD_POINTER
,
9110 S390_BUILTIN_SET_THREAD_POINTER
,
9115 static enum insn_code
const code_for_builtin_64
[S390_BUILTIN_max
] = {
9120 static enum insn_code
const code_for_builtin_31
[S390_BUILTIN_max
] = {
9126 s390_init_builtins (void)
9130 ftype
= build_function_type_list (ptr_type_node
, NULL_TREE
);
9131 add_builtin_function ("__builtin_thread_pointer", ftype
,
9132 S390_BUILTIN_THREAD_POINTER
, BUILT_IN_MD
,
9135 ftype
= build_function_type_list (void_type_node
, ptr_type_node
, NULL_TREE
);
9136 add_builtin_function ("__builtin_set_thread_pointer", ftype
,
9137 S390_BUILTIN_SET_THREAD_POINTER
, BUILT_IN_MD
,
9141 /* Expand an expression EXP that calls a built-in function,
9142 with result going to TARGET if that's convenient
9143 (and in mode MODE if that's convenient).
9144 SUBTARGET may be used as the target for computing one of EXP's operands.
9145 IGNORE is nonzero if the value is to be ignored. */
9148 s390_expand_builtin (tree exp
, rtx target
, rtx subtarget ATTRIBUTE_UNUSED
,
9149 enum machine_mode mode ATTRIBUTE_UNUSED
,
9150 int ignore ATTRIBUTE_UNUSED
)
9154 enum insn_code
const *code_for_builtin
=
9155 TARGET_64BIT
? code_for_builtin_64
: code_for_builtin_31
;
9157 tree fndecl
= TREE_OPERAND (CALL_EXPR_FN (exp
), 0);
9158 unsigned int fcode
= DECL_FUNCTION_CODE (fndecl
);
9159 enum insn_code icode
;
9160 rtx op
[MAX_ARGS
], pat
;
9164 call_expr_arg_iterator iter
;
9166 if (fcode
>= S390_BUILTIN_max
)
9167 internal_error ("bad builtin fcode");
9168 icode
= code_for_builtin
[fcode
];
9170 internal_error ("bad builtin fcode");
9172 nonvoid
= TREE_TYPE (TREE_TYPE (fndecl
)) != void_type_node
;
9175 FOR_EACH_CALL_EXPR_ARG (arg
, iter
, exp
)
9177 const struct insn_operand_data
*insn_op
;
9179 if (arg
== error_mark_node
)
9181 if (arity
> MAX_ARGS
)
9184 insn_op
= &insn_data
[icode
].operand
[arity
+ nonvoid
];
9186 op
[arity
] = expand_expr (arg
, NULL_RTX
, insn_op
->mode
, EXPAND_NORMAL
);
9188 if (!(*insn_op
->predicate
) (op
[arity
], insn_op
->mode
))
9189 op
[arity
] = copy_to_mode_reg (insn_op
->mode
, op
[arity
]);
9195 enum machine_mode tmode
= insn_data
[icode
].operand
[0].mode
;
9197 || GET_MODE (target
) != tmode
9198 || !(*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
9199 target
= gen_reg_rtx (tmode
);
9205 pat
= GEN_FCN (icode
) (target
);
9209 pat
= GEN_FCN (icode
) (target
, op
[0]);
9211 pat
= GEN_FCN (icode
) (op
[0]);
9214 pat
= GEN_FCN (icode
) (target
, op
[0], op
[1]);
9230 /* Output assembly code for the trampoline template to
9233 On S/390, we use gpr 1 internally in the trampoline code;
9234 gpr 0 is used to hold the static chain. */
9237 s390_asm_trampoline_template (FILE *file
)
9240 op
[0] = gen_rtx_REG (Pmode
, 0);
9241 op
[1] = gen_rtx_REG (Pmode
, 1);
9245 output_asm_insn ("basr\t%1,0", op
); /* 2 byte */
9246 output_asm_insn ("lmg\t%0,%1,14(%1)", op
); /* 6 byte */
9247 output_asm_insn ("br\t%1", op
); /* 2 byte */
9248 ASM_OUTPUT_SKIP (file
, (HOST_WIDE_INT
)(TRAMPOLINE_SIZE
- 10));
9252 output_asm_insn ("basr\t%1,0", op
); /* 2 byte */
9253 output_asm_insn ("lm\t%0,%1,6(%1)", op
); /* 4 byte */
9254 output_asm_insn ("br\t%1", op
); /* 2 byte */
9255 ASM_OUTPUT_SKIP (file
, (HOST_WIDE_INT
)(TRAMPOLINE_SIZE
- 8));
9259 /* Emit RTL insns to initialize the variable parts of a trampoline.
9260 FNADDR is an RTX for the address of the function's pure code.
9261 CXT is an RTX for the static chain value for the function. */
9264 s390_trampoline_init (rtx m_tramp
, tree fndecl
, rtx cxt
)
9266 rtx fnaddr
= XEXP (DECL_RTL (fndecl
), 0);
9269 emit_block_move (m_tramp
, assemble_trampoline_template (),
9270 GEN_INT (2 * UNITS_PER_LONG
), BLOCK_OP_NORMAL
);
9272 mem
= adjust_address (m_tramp
, Pmode
, 2 * UNITS_PER_LONG
);
9273 emit_move_insn (mem
, cxt
);
9274 mem
= adjust_address (m_tramp
, Pmode
, 3 * UNITS_PER_LONG
);
9275 emit_move_insn (mem
, fnaddr
);
9278 /* Output assembler code to FILE to increment profiler label # LABELNO
9279 for profiling a function entry. */
9282 s390_function_profiler (FILE *file
, int labelno
)
9287 ASM_GENERATE_INTERNAL_LABEL (label
, "LP", labelno
);
9289 fprintf (file
, "# function profiler \n");
9291 op
[0] = gen_rtx_REG (Pmode
, RETURN_REGNUM
);
9292 op
[1] = gen_rtx_REG (Pmode
, STACK_POINTER_REGNUM
);
9293 op
[1] = gen_rtx_MEM (Pmode
, plus_constant (Pmode
, op
[1], UNITS_PER_LONG
));
9295 op
[2] = gen_rtx_REG (Pmode
, 1);
9296 op
[3] = gen_rtx_SYMBOL_REF (Pmode
, label
);
9297 SYMBOL_REF_FLAGS (op
[3]) = SYMBOL_FLAG_LOCAL
;
9299 op
[4] = gen_rtx_SYMBOL_REF (Pmode
, "_mcount");
9302 op
[4] = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op
[4]), UNSPEC_PLT
);
9303 op
[4] = gen_rtx_CONST (Pmode
, op
[4]);
9308 output_asm_insn ("stg\t%0,%1", op
);
9309 output_asm_insn ("larl\t%2,%3", op
);
9310 output_asm_insn ("brasl\t%0,%4", op
);
9311 output_asm_insn ("lg\t%0,%1", op
);
9315 op
[6] = gen_label_rtx ();
9317 output_asm_insn ("st\t%0,%1", op
);
9318 output_asm_insn ("bras\t%2,%l6", op
);
9319 output_asm_insn (".long\t%4", op
);
9320 output_asm_insn (".long\t%3", op
);
9321 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[6]));
9322 output_asm_insn ("l\t%0,0(%2)", op
);
9323 output_asm_insn ("l\t%2,4(%2)", op
);
9324 output_asm_insn ("basr\t%0,%0", op
);
9325 output_asm_insn ("l\t%0,%1", op
);
9329 op
[5] = gen_label_rtx ();
9330 op
[6] = gen_label_rtx ();
9332 output_asm_insn ("st\t%0,%1", op
);
9333 output_asm_insn ("bras\t%2,%l6", op
);
9334 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[5]));
9335 output_asm_insn (".long\t%4-%l5", op
);
9336 output_asm_insn (".long\t%3-%l5", op
);
9337 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[6]));
9338 output_asm_insn ("lr\t%0,%2", op
);
9339 output_asm_insn ("a\t%0,0(%2)", op
);
9340 output_asm_insn ("a\t%2,4(%2)", op
);
9341 output_asm_insn ("basr\t%0,%0", op
);
9342 output_asm_insn ("l\t%0,%1", op
);
9346 /* Encode symbol attributes (local vs. global, tls model) of a SYMBOL_REF
9347 into its SYMBOL_REF_FLAGS. */
9350 s390_encode_section_info (tree decl
, rtx rtl
, int first
)
9352 default_encode_section_info (decl
, rtl
, first
);
9354 if (TREE_CODE (decl
) == VAR_DECL
)
9356 /* If a variable has a forced alignment to < 2 bytes, mark it
9357 with SYMBOL_FLAG_ALIGN1 to prevent it from being used as LARL
9359 if (DECL_USER_ALIGN (decl
) && DECL_ALIGN (decl
) < 16)
9360 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_ALIGN1
;
9361 if (!DECL_SIZE (decl
)
9362 || !DECL_ALIGN (decl
)
9363 || !host_integerp (DECL_SIZE (decl
), 0)
9364 || (DECL_ALIGN (decl
) <= 64
9365 && DECL_ALIGN (decl
) != tree_low_cst (DECL_SIZE (decl
), 0)))
9366 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_NOT_NATURALLY_ALIGNED
;
9369 /* Literal pool references don't have a decl so they are handled
9370 differently here. We rely on the information in the MEM_ALIGN
9371 entry to decide upon natural alignment. */
9373 && GET_CODE (XEXP (rtl
, 0)) == SYMBOL_REF
9374 && TREE_CONSTANT_POOL_ADDRESS_P (XEXP (rtl
, 0))
9375 && (MEM_ALIGN (rtl
) == 0
9376 || GET_MODE_BITSIZE (GET_MODE (rtl
)) == 0
9377 || MEM_ALIGN (rtl
) < GET_MODE_BITSIZE (GET_MODE (rtl
))))
9378 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_NOT_NATURALLY_ALIGNED
;
9381 /* Output thunk to FILE that implements a C++ virtual function call (with
9382 multiple inheritance) to FUNCTION. The thunk adjusts the this pointer
9383 by DELTA, and unless VCALL_OFFSET is zero, applies an additional adjustment
9384 stored at VCALL_OFFSET in the vtable whose address is located at offset 0
9385 relative to the resulting this pointer. */
9388 s390_output_mi_thunk (FILE *file
, tree thunk ATTRIBUTE_UNUSED
,
9389 HOST_WIDE_INT delta
, HOST_WIDE_INT vcall_offset
,
9395 /* Make sure unwind info is emitted for the thunk if needed. */
9396 final_start_function (emit_barrier (), file
, 1);
9398 /* Operand 0 is the target function. */
9399 op
[0] = XEXP (DECL_RTL (function
), 0);
9400 if (flag_pic
&& !SYMBOL_REF_LOCAL_P (op
[0]))
9403 op
[0] = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op
[0]),
9404 TARGET_64BIT
? UNSPEC_PLT
: UNSPEC_GOT
);
9405 op
[0] = gen_rtx_CONST (Pmode
, op
[0]);
9408 /* Operand 1 is the 'this' pointer. */
9409 if (aggregate_value_p (TREE_TYPE (TREE_TYPE (function
)), function
))
9410 op
[1] = gen_rtx_REG (Pmode
, 3);
9412 op
[1] = gen_rtx_REG (Pmode
, 2);
9414 /* Operand 2 is the delta. */
9415 op
[2] = GEN_INT (delta
);
9417 /* Operand 3 is the vcall_offset. */
9418 op
[3] = GEN_INT (vcall_offset
);
9420 /* Operand 4 is the temporary register. */
9421 op
[4] = gen_rtx_REG (Pmode
, 1);
9423 /* Operands 5 to 8 can be used as labels. */
9429 /* Operand 9 can be used for temporary register. */
9432 /* Generate code. */
9435 /* Setup literal pool pointer if required. */
9436 if ((!DISP_IN_RANGE (delta
)
9437 && !CONST_OK_FOR_K (delta
)
9438 && !CONST_OK_FOR_Os (delta
))
9439 || (!DISP_IN_RANGE (vcall_offset
)
9440 && !CONST_OK_FOR_K (vcall_offset
)
9441 && !CONST_OK_FOR_Os (vcall_offset
)))
9443 op
[5] = gen_label_rtx ();
9444 output_asm_insn ("larl\t%4,%5", op
);
9447 /* Add DELTA to this pointer. */
9450 if (CONST_OK_FOR_J (delta
))
9451 output_asm_insn ("la\t%1,%2(%1)", op
);
9452 else if (DISP_IN_RANGE (delta
))
9453 output_asm_insn ("lay\t%1,%2(%1)", op
);
9454 else if (CONST_OK_FOR_K (delta
))
9455 output_asm_insn ("aghi\t%1,%2", op
);
9456 else if (CONST_OK_FOR_Os (delta
))
9457 output_asm_insn ("agfi\t%1,%2", op
);
9460 op
[6] = gen_label_rtx ();
9461 output_asm_insn ("agf\t%1,%6-%5(%4)", op
);
9465 /* Perform vcall adjustment. */
9468 if (DISP_IN_RANGE (vcall_offset
))
9470 output_asm_insn ("lg\t%4,0(%1)", op
);
9471 output_asm_insn ("ag\t%1,%3(%4)", op
);
9473 else if (CONST_OK_FOR_K (vcall_offset
))
9475 output_asm_insn ("lghi\t%4,%3", op
);
9476 output_asm_insn ("ag\t%4,0(%1)", op
);
9477 output_asm_insn ("ag\t%1,0(%4)", op
);
9479 else if (CONST_OK_FOR_Os (vcall_offset
))
9481 output_asm_insn ("lgfi\t%4,%3", op
);
9482 output_asm_insn ("ag\t%4,0(%1)", op
);
9483 output_asm_insn ("ag\t%1,0(%4)", op
);
9487 op
[7] = gen_label_rtx ();
9488 output_asm_insn ("llgf\t%4,%7-%5(%4)", op
);
9489 output_asm_insn ("ag\t%4,0(%1)", op
);
9490 output_asm_insn ("ag\t%1,0(%4)", op
);
9494 /* Jump to target. */
9495 output_asm_insn ("jg\t%0", op
);
9497 /* Output literal pool if required. */
9500 output_asm_insn (".align\t4", op
);
9501 targetm
.asm_out
.internal_label (file
, "L",
9502 CODE_LABEL_NUMBER (op
[5]));
9506 targetm
.asm_out
.internal_label (file
, "L",
9507 CODE_LABEL_NUMBER (op
[6]));
9508 output_asm_insn (".long\t%2", op
);
9512 targetm
.asm_out
.internal_label (file
, "L",
9513 CODE_LABEL_NUMBER (op
[7]));
9514 output_asm_insn (".long\t%3", op
);
9519 /* Setup base pointer if required. */
9521 || (!DISP_IN_RANGE (delta
)
9522 && !CONST_OK_FOR_K (delta
)
9523 && !CONST_OK_FOR_Os (delta
))
9524 || (!DISP_IN_RANGE (delta
)
9525 && !CONST_OK_FOR_K (vcall_offset
)
9526 && !CONST_OK_FOR_Os (vcall_offset
)))
9528 op
[5] = gen_label_rtx ();
9529 output_asm_insn ("basr\t%4,0", op
);
9530 targetm
.asm_out
.internal_label (file
, "L",
9531 CODE_LABEL_NUMBER (op
[5]));
9534 /* Add DELTA to this pointer. */
9537 if (CONST_OK_FOR_J (delta
))
9538 output_asm_insn ("la\t%1,%2(%1)", op
);
9539 else if (DISP_IN_RANGE (delta
))
9540 output_asm_insn ("lay\t%1,%2(%1)", op
);
9541 else if (CONST_OK_FOR_K (delta
))
9542 output_asm_insn ("ahi\t%1,%2", op
);
9543 else if (CONST_OK_FOR_Os (delta
))
9544 output_asm_insn ("afi\t%1,%2", op
);
9547 op
[6] = gen_label_rtx ();
9548 output_asm_insn ("a\t%1,%6-%5(%4)", op
);
9552 /* Perform vcall adjustment. */
9555 if (CONST_OK_FOR_J (vcall_offset
))
9557 output_asm_insn ("l\t%4,0(%1)", op
);
9558 output_asm_insn ("a\t%1,%3(%4)", op
);
9560 else if (DISP_IN_RANGE (vcall_offset
))
9562 output_asm_insn ("l\t%4,0(%1)", op
);
9563 output_asm_insn ("ay\t%1,%3(%4)", op
);
9565 else if (CONST_OK_FOR_K (vcall_offset
))
9567 output_asm_insn ("lhi\t%4,%3", op
);
9568 output_asm_insn ("a\t%4,0(%1)", op
);
9569 output_asm_insn ("a\t%1,0(%4)", op
);
9571 else if (CONST_OK_FOR_Os (vcall_offset
))
9573 output_asm_insn ("iilf\t%4,%3", op
);
9574 output_asm_insn ("a\t%4,0(%1)", op
);
9575 output_asm_insn ("a\t%1,0(%4)", op
);
9579 op
[7] = gen_label_rtx ();
9580 output_asm_insn ("l\t%4,%7-%5(%4)", op
);
9581 output_asm_insn ("a\t%4,0(%1)", op
);
9582 output_asm_insn ("a\t%1,0(%4)", op
);
9585 /* We had to clobber the base pointer register.
9586 Re-setup the base pointer (with a different base). */
9587 op
[5] = gen_label_rtx ();
9588 output_asm_insn ("basr\t%4,0", op
);
9589 targetm
.asm_out
.internal_label (file
, "L",
9590 CODE_LABEL_NUMBER (op
[5]));
9593 /* Jump to target. */
9594 op
[8] = gen_label_rtx ();
9597 output_asm_insn ("l\t%4,%8-%5(%4)", op
);
9599 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9600 /* We cannot call through .plt, since .plt requires %r12 loaded. */
9601 else if (flag_pic
== 1)
9603 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9604 output_asm_insn ("l\t%4,%0(%4)", op
);
9606 else if (flag_pic
== 2)
9608 op
[9] = gen_rtx_REG (Pmode
, 0);
9609 output_asm_insn ("l\t%9,%8-4-%5(%4)", op
);
9610 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9611 output_asm_insn ("ar\t%4,%9", op
);
9612 output_asm_insn ("l\t%4,0(%4)", op
);
9615 output_asm_insn ("br\t%4", op
);
9617 /* Output literal pool. */
9618 output_asm_insn (".align\t4", op
);
9620 if (nonlocal
&& flag_pic
== 2)
9621 output_asm_insn (".long\t%0", op
);
9624 op
[0] = gen_rtx_SYMBOL_REF (Pmode
, "_GLOBAL_OFFSET_TABLE_");
9625 SYMBOL_REF_FLAGS (op
[0]) = SYMBOL_FLAG_LOCAL
;
9628 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[8]));
9630 output_asm_insn (".long\t%0", op
);
9632 output_asm_insn (".long\t%0-%5", op
);
9636 targetm
.asm_out
.internal_label (file
, "L",
9637 CODE_LABEL_NUMBER (op
[6]));
9638 output_asm_insn (".long\t%2", op
);
9642 targetm
.asm_out
.internal_label (file
, "L",
9643 CODE_LABEL_NUMBER (op
[7]));
9644 output_asm_insn (".long\t%3", op
);
9647 final_end_function ();
9651 s390_valid_pointer_mode (enum machine_mode mode
)
9653 return (mode
== SImode
|| (TARGET_64BIT
&& mode
== DImode
));
9656 /* Checks whether the given CALL_EXPR would use a caller
9657 saved register. This is used to decide whether sibling call
9658 optimization could be performed on the respective function
9662 s390_call_saved_register_used (tree call_expr
)
9664 CUMULATIVE_ARGS cum_v
;
9665 cumulative_args_t cum
;
9667 enum machine_mode mode
;
9672 INIT_CUMULATIVE_ARGS (cum_v
, NULL
, NULL
, 0, 0);
9673 cum
= pack_cumulative_args (&cum_v
);
9675 for (i
= 0; i
< call_expr_nargs (call_expr
); i
++)
9677 parameter
= CALL_EXPR_ARG (call_expr
, i
);
9678 gcc_assert (parameter
);
9680 /* For an undeclared variable passed as parameter we will get
9681 an ERROR_MARK node here. */
9682 if (TREE_CODE (parameter
) == ERROR_MARK
)
9685 type
= TREE_TYPE (parameter
);
9688 mode
= TYPE_MODE (type
);
9691 if (pass_by_reference (&cum_v
, mode
, type
, true))
9694 type
= build_pointer_type (type
);
9697 parm_rtx
= s390_function_arg (cum
, mode
, type
, 0);
9699 s390_function_arg_advance (cum
, mode
, type
, 0);
9704 if (REG_P (parm_rtx
))
9707 reg
< HARD_REGNO_NREGS (REGNO (parm_rtx
), GET_MODE (parm_rtx
));
9709 if (!call_used_regs
[reg
+ REGNO (parm_rtx
)])
9713 if (GET_CODE (parm_rtx
) == PARALLEL
)
9717 for (i
= 0; i
< XVECLEN (parm_rtx
, 0); i
++)
9719 rtx r
= XEXP (XVECEXP (parm_rtx
, 0, i
), 0);
9721 gcc_assert (REG_P (r
));
9724 reg
< HARD_REGNO_NREGS (REGNO (r
), GET_MODE (r
));
9726 if (!call_used_regs
[reg
+ REGNO (r
)])
9735 /* Return true if the given call expression can be
9736 turned into a sibling call.
9737 DECL holds the declaration of the function to be called whereas
9738 EXP is the call expression itself. */
9741 s390_function_ok_for_sibcall (tree decl
, tree exp
)
9743 /* The TPF epilogue uses register 1. */
9744 if (TARGET_TPF_PROFILING
)
9747 /* The 31 bit PLT code uses register 12 (GOT pointer - caller saved)
9748 which would have to be restored before the sibcall. */
9749 if (!TARGET_64BIT
&& flag_pic
&& decl
&& !targetm
.binds_local_p (decl
))
9752 /* Register 6 on s390 is available as an argument register but unfortunately
9753 "caller saved". This makes functions needing this register for arguments
9754 not suitable for sibcalls. */
9755 return !s390_call_saved_register_used (exp
);
9758 /* Return the fixed registers used for condition codes. */
9761 s390_fixed_condition_code_regs (unsigned int *p1
, unsigned int *p2
)
9764 *p2
= INVALID_REGNUM
;
9769 /* This function is used by the call expanders of the machine description.
9770 It emits the call insn itself together with the necessary operations
9771 to adjust the target address and returns the emitted insn.
9772 ADDR_LOCATION is the target address rtx
9773 TLS_CALL the location of the thread-local symbol
9774 RESULT_REG the register where the result of the call should be stored
9775 RETADDR_REG the register where the return address should be stored
9776 If this parameter is NULL_RTX the call is considered
9777 to be a sibling call. */
9780 s390_emit_call (rtx addr_location
, rtx tls_call
, rtx result_reg
,
9783 bool plt_call
= false;
9789 /* Direct function calls need special treatment. */
9790 if (GET_CODE (addr_location
) == SYMBOL_REF
)
9792 /* When calling a global routine in PIC mode, we must
9793 replace the symbol itself with the PLT stub. */
9794 if (flag_pic
&& !SYMBOL_REF_LOCAL_P (addr_location
))
9796 if (retaddr_reg
!= NULL_RTX
)
9798 addr_location
= gen_rtx_UNSPEC (Pmode
,
9799 gen_rtvec (1, addr_location
),
9801 addr_location
= gen_rtx_CONST (Pmode
, addr_location
);
9805 /* For -fpic code the PLT entries might use r12 which is
9806 call-saved. Therefore we cannot do a sibcall when
9807 calling directly using a symbol ref. When reaching
9808 this point we decided (in s390_function_ok_for_sibcall)
9809 to do a sibcall for a function pointer but one of the
9810 optimizers was able to get rid of the function pointer
9811 by propagating the symbol ref into the call. This
9812 optimization is illegal for S/390 so we turn the direct
9813 call into a indirect call again. */
9814 addr_location
= force_reg (Pmode
, addr_location
);
9817 /* Unless we can use the bras(l) insn, force the
9818 routine address into a register. */
9819 if (!TARGET_SMALL_EXEC
&& !TARGET_CPU_ZARCH
)
9822 addr_location
= legitimize_pic_address (addr_location
, 0);
9824 addr_location
= force_reg (Pmode
, addr_location
);
9828 /* If it is already an indirect call or the code above moved the
9829 SYMBOL_REF to somewhere else make sure the address can be found in
9831 if (retaddr_reg
== NULL_RTX
9832 && GET_CODE (addr_location
) != SYMBOL_REF
9835 emit_move_insn (gen_rtx_REG (Pmode
, SIBCALL_REGNUM
), addr_location
);
9836 addr_location
= gen_rtx_REG (Pmode
, SIBCALL_REGNUM
);
9839 addr_location
= gen_rtx_MEM (QImode
, addr_location
);
9840 call
= gen_rtx_CALL (VOIDmode
, addr_location
, const0_rtx
);
9842 if (result_reg
!= NULL_RTX
)
9843 call
= gen_rtx_SET (VOIDmode
, result_reg
, call
);
9845 if (retaddr_reg
!= NULL_RTX
)
9847 clobber
= gen_rtx_CLOBBER (VOIDmode
, retaddr_reg
);
9849 if (tls_call
!= NULL_RTX
)
9850 vec
= gen_rtvec (3, call
, clobber
,
9851 gen_rtx_USE (VOIDmode
, tls_call
));
9853 vec
= gen_rtvec (2, call
, clobber
);
9855 call
= gen_rtx_PARALLEL (VOIDmode
, vec
);
9858 insn
= emit_call_insn (call
);
9860 /* 31-bit PLT stubs and tls calls use the GOT register implicitly. */
9861 if ((!TARGET_64BIT
&& plt_call
) || tls_call
!= NULL_RTX
)
9863 /* s390_function_ok_for_sibcall should
9864 have denied sibcalls in this case. */
9865 gcc_assert (retaddr_reg
!= NULL_RTX
);
9866 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), gen_rtx_REG (Pmode
, 12));
9871 /* Implement TARGET_CONDITIONAL_REGISTER_USAGE. */
9874 s390_conditional_register_usage (void)
9880 fixed_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
9881 call_used_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
9883 if (TARGET_CPU_ZARCH
)
9885 fixed_regs
[BASE_REGNUM
] = 0;
9886 call_used_regs
[BASE_REGNUM
] = 0;
9887 fixed_regs
[RETURN_REGNUM
] = 0;
9888 call_used_regs
[RETURN_REGNUM
] = 0;
9892 for (i
= 24; i
< 32; i
++)
9893 call_used_regs
[i
] = call_really_used_regs
[i
] = 0;
9897 for (i
= 18; i
< 20; i
++)
9898 call_used_regs
[i
] = call_really_used_regs
[i
] = 0;
9901 if (TARGET_SOFT_FLOAT
)
9903 for (i
= 16; i
< 32; i
++)
9904 call_used_regs
[i
] = fixed_regs
[i
] = 1;
9908 /* Corresponding function to eh_return expander. */
9910 static GTY(()) rtx s390_tpf_eh_return_symbol
;
9912 s390_emit_tpf_eh_return (rtx target
)
9916 if (!s390_tpf_eh_return_symbol
)
9917 s390_tpf_eh_return_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "__tpf_eh_return");
9919 reg
= gen_rtx_REG (Pmode
, 2);
9921 emit_move_insn (reg
, target
);
9922 insn
= s390_emit_call (s390_tpf_eh_return_symbol
, NULL_RTX
, reg
,
9923 gen_rtx_REG (Pmode
, RETURN_REGNUM
));
9924 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), reg
);
9926 emit_move_insn (EH_RETURN_HANDLER_RTX
, reg
);
9929 /* Rework the prologue/epilogue to avoid saving/restoring
9930 registers unnecessarily. */
9933 s390_optimize_prologue (void)
9935 rtx insn
, new_insn
, next_insn
;
9937 /* Do a final recompute of the frame-related data. */
9939 s390_update_frame_layout ();
9941 /* If all special registers are in fact used, there's nothing we
9942 can do, so no point in walking the insn list. */
9944 if (cfun_frame_layout
.first_save_gpr
<= BASE_REGNUM
9945 && cfun_frame_layout
.last_save_gpr
>= BASE_REGNUM
9946 && (TARGET_CPU_ZARCH
9947 || (cfun_frame_layout
.first_save_gpr
<= RETURN_REGNUM
9948 && cfun_frame_layout
.last_save_gpr
>= RETURN_REGNUM
)))
9951 /* Search for prologue/epilogue insns and replace them. */
9953 for (insn
= get_insns (); insn
; insn
= next_insn
)
9955 int first
, last
, off
;
9956 rtx set
, base
, offset
;
9958 next_insn
= NEXT_INSN (insn
);
9960 if (GET_CODE (insn
) != INSN
)
9963 if (GET_CODE (PATTERN (insn
)) == PARALLEL
9964 && store_multiple_operation (PATTERN (insn
), VOIDmode
))
9966 set
= XVECEXP (PATTERN (insn
), 0, 0);
9967 first
= REGNO (SET_SRC (set
));
9968 last
= first
+ XVECLEN (PATTERN (insn
), 0) - 1;
9969 offset
= const0_rtx
;
9970 base
= eliminate_constant_term (XEXP (SET_DEST (set
), 0), &offset
);
9971 off
= INTVAL (offset
);
9973 if (GET_CODE (base
) != REG
|| off
< 0)
9975 if (cfun_frame_layout
.first_save_gpr
!= -1
9976 && (cfun_frame_layout
.first_save_gpr
< first
9977 || cfun_frame_layout
.last_save_gpr
> last
))
9979 if (REGNO (base
) != STACK_POINTER_REGNUM
9980 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
9982 if (first
> BASE_REGNUM
|| last
< BASE_REGNUM
)
9985 if (cfun_frame_layout
.first_save_gpr
!= -1)
9987 new_insn
= save_gprs (base
,
9988 off
+ (cfun_frame_layout
.first_save_gpr
9989 - first
) * UNITS_PER_LONG
,
9990 cfun_frame_layout
.first_save_gpr
,
9991 cfun_frame_layout
.last_save_gpr
);
9992 new_insn
= emit_insn_before (new_insn
, insn
);
9993 INSN_ADDRESSES_NEW (new_insn
, -1);
10000 if (cfun_frame_layout
.first_save_gpr
== -1
10001 && GET_CODE (PATTERN (insn
)) == SET
10002 && GET_CODE (SET_SRC (PATTERN (insn
))) == REG
10003 && (REGNO (SET_SRC (PATTERN (insn
))) == BASE_REGNUM
10004 || (!TARGET_CPU_ZARCH
10005 && REGNO (SET_SRC (PATTERN (insn
))) == RETURN_REGNUM
))
10006 && GET_CODE (SET_DEST (PATTERN (insn
))) == MEM
)
10008 set
= PATTERN (insn
);
10009 first
= REGNO (SET_SRC (set
));
10010 offset
= const0_rtx
;
10011 base
= eliminate_constant_term (XEXP (SET_DEST (set
), 0), &offset
);
10012 off
= INTVAL (offset
);
10014 if (GET_CODE (base
) != REG
|| off
< 0)
10016 if (REGNO (base
) != STACK_POINTER_REGNUM
10017 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10020 remove_insn (insn
);
10024 if (GET_CODE (PATTERN (insn
)) == PARALLEL
10025 && load_multiple_operation (PATTERN (insn
), VOIDmode
))
10027 set
= XVECEXP (PATTERN (insn
), 0, 0);
10028 first
= REGNO (SET_DEST (set
));
10029 last
= first
+ XVECLEN (PATTERN (insn
), 0) - 1;
10030 offset
= const0_rtx
;
10031 base
= eliminate_constant_term (XEXP (SET_SRC (set
), 0), &offset
);
10032 off
= INTVAL (offset
);
10034 if (GET_CODE (base
) != REG
|| off
< 0)
10036 if (cfun_frame_layout
.first_restore_gpr
!= -1
10037 && (cfun_frame_layout
.first_restore_gpr
< first
10038 || cfun_frame_layout
.last_restore_gpr
> last
))
10040 if (REGNO (base
) != STACK_POINTER_REGNUM
10041 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10043 if (first
> BASE_REGNUM
|| last
< BASE_REGNUM
)
10046 if (cfun_frame_layout
.first_restore_gpr
!= -1)
10048 new_insn
= restore_gprs (base
,
10049 off
+ (cfun_frame_layout
.first_restore_gpr
10050 - first
) * UNITS_PER_LONG
,
10051 cfun_frame_layout
.first_restore_gpr
,
10052 cfun_frame_layout
.last_restore_gpr
);
10053 new_insn
= emit_insn_before (new_insn
, insn
);
10054 INSN_ADDRESSES_NEW (new_insn
, -1);
10057 remove_insn (insn
);
10061 if (cfun_frame_layout
.first_restore_gpr
== -1
10062 && GET_CODE (PATTERN (insn
)) == SET
10063 && GET_CODE (SET_DEST (PATTERN (insn
))) == REG
10064 && (REGNO (SET_DEST (PATTERN (insn
))) == BASE_REGNUM
10065 || (!TARGET_CPU_ZARCH
10066 && REGNO (SET_DEST (PATTERN (insn
))) == RETURN_REGNUM
))
10067 && GET_CODE (SET_SRC (PATTERN (insn
))) == MEM
)
10069 set
= PATTERN (insn
);
10070 first
= REGNO (SET_DEST (set
));
10071 offset
= const0_rtx
;
10072 base
= eliminate_constant_term (XEXP (SET_SRC (set
), 0), &offset
);
10073 off
= INTVAL (offset
);
10075 if (GET_CODE (base
) != REG
|| off
< 0)
10077 if (REGNO (base
) != STACK_POINTER_REGNUM
10078 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10081 remove_insn (insn
);
10087 /* On z10 and later the dynamic branch prediction must see the
10088 backward jump within a certain windows. If not it falls back to
10089 the static prediction. This function rearranges the loop backward
10090 branch in a way which makes the static prediction always correct.
10091 The function returns true if it added an instruction. */
10093 s390_fix_long_loop_prediction (rtx insn
)
10095 rtx set
= single_set (insn
);
10096 rtx code_label
, label_ref
, new_label
;
10102 /* This will exclude branch on count and branch on index patterns
10103 since these are correctly statically predicted. */
10105 || SET_DEST (set
) != pc_rtx
10106 || GET_CODE (SET_SRC(set
)) != IF_THEN_ELSE
)
10109 label_ref
= (GET_CODE (XEXP (SET_SRC (set
), 1)) == LABEL_REF
?
10110 XEXP (SET_SRC (set
), 1) : XEXP (SET_SRC (set
), 2));
10112 gcc_assert (GET_CODE (label_ref
) == LABEL_REF
);
10114 code_label
= XEXP (label_ref
, 0);
10116 if (INSN_ADDRESSES (INSN_UID (code_label
)) == -1
10117 || INSN_ADDRESSES (INSN_UID (insn
)) == -1
10118 || (INSN_ADDRESSES (INSN_UID (insn
))
10119 - INSN_ADDRESSES (INSN_UID (code_label
)) < PREDICT_DISTANCE
))
10122 for (distance
= 0, cur_insn
= PREV_INSN (insn
);
10123 distance
< PREDICT_DISTANCE
- 6;
10124 distance
+= get_attr_length (cur_insn
), cur_insn
= PREV_INSN (cur_insn
))
10125 if (!cur_insn
|| JUMP_P (cur_insn
) || LABEL_P (cur_insn
))
10128 new_label
= gen_label_rtx ();
10129 uncond_jump
= emit_jump_insn_after (
10130 gen_rtx_SET (VOIDmode
, pc_rtx
,
10131 gen_rtx_LABEL_REF (VOIDmode
, code_label
)),
10133 emit_label_after (new_label
, uncond_jump
);
10135 tmp
= XEXP (SET_SRC (set
), 1);
10136 XEXP (SET_SRC (set
), 1) = XEXP (SET_SRC (set
), 2);
10137 XEXP (SET_SRC (set
), 2) = tmp
;
10138 INSN_CODE (insn
) = -1;
10140 XEXP (label_ref
, 0) = new_label
;
10141 JUMP_LABEL (insn
) = new_label
;
10142 JUMP_LABEL (uncond_jump
) = code_label
;
10147 /* Returns 1 if INSN reads the value of REG for purposes not related
10148 to addressing of memory, and 0 otherwise. */
10150 s390_non_addr_reg_read_p (rtx reg
, rtx insn
)
10152 return reg_referenced_p (reg
, PATTERN (insn
))
10153 && !reg_used_in_mem_p (REGNO (reg
), PATTERN (insn
));
10156 /* Starting from INSN find_cond_jump looks downwards in the insn
10157 stream for a single jump insn which is the last user of the
10158 condition code set in INSN. */
10160 find_cond_jump (rtx insn
)
10162 for (; insn
; insn
= NEXT_INSN (insn
))
10166 if (LABEL_P (insn
))
10169 if (!JUMP_P (insn
))
10171 if (reg_mentioned_p (gen_rtx_REG (CCmode
, CC_REGNUM
), insn
))
10176 /* This will be triggered by a return. */
10177 if (GET_CODE (PATTERN (insn
)) != SET
)
10180 gcc_assert (SET_DEST (PATTERN (insn
)) == pc_rtx
);
10181 ite
= SET_SRC (PATTERN (insn
));
10183 if (GET_CODE (ite
) != IF_THEN_ELSE
)
10186 cc
= XEXP (XEXP (ite
, 0), 0);
10187 if (!REG_P (cc
) || !CC_REGNO_P (REGNO (cc
)))
10190 if (find_reg_note (insn
, REG_DEAD
, cc
))
10198 /* Swap the condition in COND and the operands in OP0 and OP1 so that
10199 the semantics does not change. If NULL_RTX is passed as COND the
10200 function tries to find the conditional jump starting with INSN. */
10202 s390_swap_cmp (rtx cond
, rtx
*op0
, rtx
*op1
, rtx insn
)
10206 if (cond
== NULL_RTX
)
10208 rtx jump
= find_cond_jump (NEXT_INSN (insn
));
10209 jump
= jump
? single_set (jump
) : NULL_RTX
;
10211 if (jump
== NULL_RTX
)
10214 cond
= XEXP (XEXP (jump
, 1), 0);
10219 PUT_CODE (cond
, swap_condition (GET_CODE (cond
)));
10222 /* On z10, instructions of the compare-and-branch family have the
10223 property to access the register occurring as second operand with
10224 its bits complemented. If such a compare is grouped with a second
10225 instruction that accesses the same register non-complemented, and
10226 if that register's value is delivered via a bypass, then the
10227 pipeline recycles, thereby causing significant performance decline.
10228 This function locates such situations and exchanges the two
10229 operands of the compare. The function return true whenever it
10232 s390_z10_optimize_cmp (rtx insn
)
10234 rtx prev_insn
, next_insn
;
10235 bool insn_added_p
= false;
10236 rtx cond
, *op0
, *op1
;
10238 if (GET_CODE (PATTERN (insn
)) == PARALLEL
)
10240 /* Handle compare and branch and branch on count
10242 rtx pattern
= single_set (insn
);
10245 || SET_DEST (pattern
) != pc_rtx
10246 || GET_CODE (SET_SRC (pattern
)) != IF_THEN_ELSE
)
10249 cond
= XEXP (SET_SRC (pattern
), 0);
10250 op0
= &XEXP (cond
, 0);
10251 op1
= &XEXP (cond
, 1);
10253 else if (GET_CODE (PATTERN (insn
)) == SET
)
10257 /* Handle normal compare instructions. */
10258 src
= SET_SRC (PATTERN (insn
));
10259 dest
= SET_DEST (PATTERN (insn
));
10262 || !CC_REGNO_P (REGNO (dest
))
10263 || GET_CODE (src
) != COMPARE
)
10266 /* s390_swap_cmp will try to find the conditional
10267 jump when passing NULL_RTX as condition. */
10269 op0
= &XEXP (src
, 0);
10270 op1
= &XEXP (src
, 1);
10275 if (!REG_P (*op0
) || !REG_P (*op1
))
10278 if (GET_MODE_CLASS (GET_MODE (*op0
)) != MODE_INT
)
10281 /* Swap the COMPARE arguments and its mask if there is a
10282 conflicting access in the previous insn. */
10283 prev_insn
= prev_active_insn (insn
);
10284 if (prev_insn
!= NULL_RTX
&& INSN_P (prev_insn
)
10285 && reg_referenced_p (*op1
, PATTERN (prev_insn
)))
10286 s390_swap_cmp (cond
, op0
, op1
, insn
);
10288 /* Check if there is a conflict with the next insn. If there
10289 was no conflict with the previous insn, then swap the
10290 COMPARE arguments and its mask. If we already swapped
10291 the operands, or if swapping them would cause a conflict
10292 with the previous insn, issue a NOP after the COMPARE in
10293 order to separate the two instuctions. */
10294 next_insn
= next_active_insn (insn
);
10295 if (next_insn
!= NULL_RTX
&& INSN_P (next_insn
)
10296 && s390_non_addr_reg_read_p (*op1
, next_insn
))
10298 if (prev_insn
!= NULL_RTX
&& INSN_P (prev_insn
)
10299 && s390_non_addr_reg_read_p (*op0
, prev_insn
))
10301 if (REGNO (*op1
) == 0)
10302 emit_insn_after (gen_nop1 (), insn
);
10304 emit_insn_after (gen_nop (), insn
);
10305 insn_added_p
= true;
10308 s390_swap_cmp (cond
, op0
, op1
, insn
);
10310 return insn_added_p
;
10313 /* Perform machine-dependent processing. */
10318 bool pool_overflow
= false;
10320 /* Make sure all splits have been performed; splits after
10321 machine_dependent_reorg might confuse insn length counts. */
10322 split_all_insns_noflow ();
10324 /* Install the main literal pool and the associated base
10325 register load insns.
10327 In addition, there are two problematic situations we need
10330 - the literal pool might be > 4096 bytes in size, so that
10331 some of its elements cannot be directly accessed
10333 - a branch target might be > 64K away from the branch, so that
10334 it is not possible to use a PC-relative instruction.
10336 To fix those, we split the single literal pool into multiple
10337 pool chunks, reloading the pool base register at various
10338 points throughout the function to ensure it always points to
10339 the pool chunk the following code expects, and / or replace
10340 PC-relative branches by absolute branches.
10342 However, the two problems are interdependent: splitting the
10343 literal pool can move a branch further away from its target,
10344 causing the 64K limit to overflow, and on the other hand,
10345 replacing a PC-relative branch by an absolute branch means
10346 we need to put the branch target address into the literal
10347 pool, possibly causing it to overflow.
10349 So, we loop trying to fix up both problems until we manage
10350 to satisfy both conditions at the same time. Note that the
10351 loop is guaranteed to terminate as every pass of the loop
10352 strictly decreases the total number of PC-relative branches
10353 in the function. (This is not completely true as there
10354 might be branch-over-pool insns introduced by chunkify_start.
10355 Those never need to be split however.) */
10359 struct constant_pool
*pool
= NULL
;
10361 /* Collect the literal pool. */
10362 if (!pool_overflow
)
10364 pool
= s390_mainpool_start ();
10366 pool_overflow
= true;
10369 /* If literal pool overflowed, start to chunkify it. */
10371 pool
= s390_chunkify_start ();
10373 /* Split out-of-range branches. If this has created new
10374 literal pool entries, cancel current chunk list and
10375 recompute it. zSeries machines have large branch
10376 instructions, so we never need to split a branch. */
10377 if (!TARGET_CPU_ZARCH
&& s390_split_branches ())
10380 s390_chunkify_cancel (pool
);
10382 s390_mainpool_cancel (pool
);
10387 /* If we made it up to here, both conditions are satisfied.
10388 Finish up literal pool related changes. */
10390 s390_chunkify_finish (pool
);
10392 s390_mainpool_finish (pool
);
10394 /* We're done splitting branches. */
10395 cfun
->machine
->split_branches_pending_p
= false;
10399 /* Generate out-of-pool execute target insns. */
10400 if (TARGET_CPU_ZARCH
)
10402 rtx insn
, label
, target
;
10404 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
10406 label
= s390_execute_label (insn
);
10410 gcc_assert (label
!= const0_rtx
);
10412 target
= emit_label (XEXP (label
, 0));
10413 INSN_ADDRESSES_NEW (target
, -1);
10415 target
= emit_insn (s390_execute_target (insn
));
10416 INSN_ADDRESSES_NEW (target
, -1);
10420 /* Try to optimize prologue and epilogue further. */
10421 s390_optimize_prologue ();
10423 /* Walk over the insns and do some >=z10 specific changes. */
10424 if (s390_tune
== PROCESSOR_2097_Z10
10425 || s390_tune
== PROCESSOR_2817_Z196
)
10428 bool insn_added_p
= false;
10430 /* The insn lengths and addresses have to be up to date for the
10431 following manipulations. */
10432 shorten_branches (get_insns ());
10434 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
10436 if (!INSN_P (insn
) || INSN_CODE (insn
) <= 0)
10440 insn_added_p
|= s390_fix_long_loop_prediction (insn
);
10442 if ((GET_CODE (PATTERN (insn
)) == PARALLEL
10443 || GET_CODE (PATTERN (insn
)) == SET
)
10444 && s390_tune
== PROCESSOR_2097_Z10
)
10445 insn_added_p
|= s390_z10_optimize_cmp (insn
);
10448 /* Adjust branches if we added new instructions. */
10450 shorten_branches (get_insns ());
10454 /* Return true if INSN is a fp load insn writing register REGNO. */
10456 s390_fpload_toreg (rtx insn
, unsigned int regno
)
10459 enum attr_type flag
= s390_safe_attr_type (insn
);
10461 if (flag
!= TYPE_FLOADSF
&& flag
!= TYPE_FLOADDF
)
10464 set
= single_set (insn
);
10466 if (set
== NULL_RTX
)
10469 if (!REG_P (SET_DEST (set
)) || !MEM_P (SET_SRC (set
)))
10472 if (REGNO (SET_DEST (set
)) != regno
)
10478 /* This value describes the distance to be avoided between an
10479 aritmetic fp instruction and an fp load writing the same register.
10480 Z10_EARLYLOAD_DISTANCE - 1 as well as Z10_EARLYLOAD_DISTANCE + 1 is
10481 fine but the exact value has to be avoided. Otherwise the FP
10482 pipeline will throw an exception causing a major penalty. */
10483 #define Z10_EARLYLOAD_DISTANCE 7
10485 /* Rearrange the ready list in order to avoid the situation described
10486 for Z10_EARLYLOAD_DISTANCE. A problematic load instruction is
10487 moved to the very end of the ready list. */
10489 s390_z10_prevent_earlyload_conflicts (rtx
*ready
, int *nready_p
)
10491 unsigned int regno
;
10492 int nready
= *nready_p
;
10497 enum attr_type flag
;
10500 /* Skip DISTANCE - 1 active insns. */
10501 for (insn
= last_scheduled_insn
, distance
= Z10_EARLYLOAD_DISTANCE
- 1;
10502 distance
> 0 && insn
!= NULL_RTX
;
10503 distance
--, insn
= prev_active_insn (insn
))
10504 if (CALL_P (insn
) || JUMP_P (insn
))
10507 if (insn
== NULL_RTX
)
10510 set
= single_set (insn
);
10512 if (set
== NULL_RTX
|| !REG_P (SET_DEST (set
))
10513 || GET_MODE_CLASS (GET_MODE (SET_DEST (set
))) != MODE_FLOAT
)
10516 flag
= s390_safe_attr_type (insn
);
10518 if (flag
== TYPE_FLOADSF
|| flag
== TYPE_FLOADDF
)
10521 regno
= REGNO (SET_DEST (set
));
10524 while (!s390_fpload_toreg (ready
[i
], regno
) && i
> 0)
10531 memmove (&ready
[1], &ready
[0], sizeof (rtx
) * i
);
10535 /* This function is called via hook TARGET_SCHED_REORDER before
10536 issuing one insn from list READY which contains *NREADYP entries.
10537 For target z10 it reorders load instructions to avoid early load
10538 conflicts in the floating point pipeline */
10540 s390_sched_reorder (FILE *file ATTRIBUTE_UNUSED
, int verbose ATTRIBUTE_UNUSED
,
10541 rtx
*ready
, int *nreadyp
, int clock ATTRIBUTE_UNUSED
)
10543 if (s390_tune
== PROCESSOR_2097_Z10
)
10544 if (reload_completed
&& *nreadyp
> 1)
10545 s390_z10_prevent_earlyload_conflicts (ready
, nreadyp
);
10547 return s390_issue_rate ();
10550 /* This function is called via hook TARGET_SCHED_VARIABLE_ISSUE after
10551 the scheduler has issued INSN. It stores the last issued insn into
10552 last_scheduled_insn in order to make it available for
10553 s390_sched_reorder. */
10555 s390_sched_variable_issue (FILE *file ATTRIBUTE_UNUSED
,
10556 int verbose ATTRIBUTE_UNUSED
,
10557 rtx insn
, int more
)
10559 last_scheduled_insn
= insn
;
10561 if (GET_CODE (PATTERN (insn
)) != USE
10562 && GET_CODE (PATTERN (insn
)) != CLOBBER
)
10569 s390_sched_init (FILE *file ATTRIBUTE_UNUSED
,
10570 int verbose ATTRIBUTE_UNUSED
,
10571 int max_ready ATTRIBUTE_UNUSED
)
10573 last_scheduled_insn
= NULL_RTX
;
10576 /* This function checks the whole of insn X for memory references. The
10577 function always returns zero because the framework it is called
10578 from would stop recursively analyzing the insn upon a return value
10579 other than zero. The real result of this function is updating
10580 counter variable MEM_COUNT. */
10582 check_dpu (rtx
*x
, unsigned *mem_count
)
10584 if (*x
!= NULL_RTX
&& MEM_P (*x
))
10589 /* This target hook implementation for TARGET_LOOP_UNROLL_ADJUST calculates
10590 a new number struct loop *loop should be unrolled if tuned for cpus with
10591 a built-in stride prefetcher.
10592 The loop is analyzed for memory accesses by calling check_dpu for
10593 each rtx of the loop. Depending on the loop_depth and the amount of
10594 memory accesses a new number <=nunroll is returned to improve the
10595 behaviour of the hardware prefetch unit. */
10597 s390_loop_unroll_adjust (unsigned nunroll
, struct loop
*loop
)
10602 unsigned mem_count
= 0;
10604 if (s390_tune
!= PROCESSOR_2097_Z10
&& s390_tune
!= PROCESSOR_2817_Z196
)
10607 /* Count the number of memory references within the loop body. */
10608 bbs
= get_loop_body (loop
);
10609 for (i
= 0; i
< loop
->num_nodes
; i
++)
10611 for (insn
= BB_HEAD (bbs
[i
]); insn
!= BB_END (bbs
[i
]); insn
= NEXT_INSN (insn
))
10612 if (INSN_P (insn
) && INSN_CODE (insn
) != -1)
10613 for_each_rtx (&insn
, (rtx_function
) check_dpu
, &mem_count
);
10617 /* Prevent division by zero, and we do not need to adjust nunroll in this case. */
10618 if (mem_count
== 0)
10621 switch (loop_depth(loop
))
10624 return MIN (nunroll
, 28 / mem_count
);
10626 return MIN (nunroll
, 22 / mem_count
);
10628 return MIN (nunroll
, 16 / mem_count
);
10632 /* Initialize GCC target structure. */
10634 #undef TARGET_ASM_ALIGNED_HI_OP
10635 #define TARGET_ASM_ALIGNED_HI_OP "\t.word\t"
10636 #undef TARGET_ASM_ALIGNED_DI_OP
10637 #define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t"
10638 #undef TARGET_ASM_INTEGER
10639 #define TARGET_ASM_INTEGER s390_assemble_integer
10641 #undef TARGET_ASM_OPEN_PAREN
10642 #define TARGET_ASM_OPEN_PAREN ""
10644 #undef TARGET_ASM_CLOSE_PAREN
10645 #define TARGET_ASM_CLOSE_PAREN ""
10647 #undef TARGET_OPTION_OVERRIDE
10648 #define TARGET_OPTION_OVERRIDE s390_option_override
10650 #undef TARGET_ENCODE_SECTION_INFO
10651 #define TARGET_ENCODE_SECTION_INFO s390_encode_section_info
10653 #undef TARGET_SCALAR_MODE_SUPPORTED_P
10654 #define TARGET_SCALAR_MODE_SUPPORTED_P s390_scalar_mode_supported_p
10657 #undef TARGET_HAVE_TLS
10658 #define TARGET_HAVE_TLS true
10660 #undef TARGET_CANNOT_FORCE_CONST_MEM
10661 #define TARGET_CANNOT_FORCE_CONST_MEM s390_cannot_force_const_mem
10663 #undef TARGET_DELEGITIMIZE_ADDRESS
10664 #define TARGET_DELEGITIMIZE_ADDRESS s390_delegitimize_address
10666 #undef TARGET_LEGITIMIZE_ADDRESS
10667 #define TARGET_LEGITIMIZE_ADDRESS s390_legitimize_address
10669 #undef TARGET_RETURN_IN_MEMORY
10670 #define TARGET_RETURN_IN_MEMORY s390_return_in_memory
10672 #undef TARGET_INIT_BUILTINS
10673 #define TARGET_INIT_BUILTINS s390_init_builtins
10674 #undef TARGET_EXPAND_BUILTIN
10675 #define TARGET_EXPAND_BUILTIN s390_expand_builtin
10677 #undef TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA
10678 #define TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA s390_output_addr_const_extra
10680 #undef TARGET_ASM_OUTPUT_MI_THUNK
10681 #define TARGET_ASM_OUTPUT_MI_THUNK s390_output_mi_thunk
10682 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
10683 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true
10685 #undef TARGET_SCHED_ADJUST_PRIORITY
10686 #define TARGET_SCHED_ADJUST_PRIORITY s390_adjust_priority
10687 #undef TARGET_SCHED_ISSUE_RATE
10688 #define TARGET_SCHED_ISSUE_RATE s390_issue_rate
10689 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
10690 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD s390_first_cycle_multipass_dfa_lookahead
10692 #undef TARGET_SCHED_VARIABLE_ISSUE
10693 #define TARGET_SCHED_VARIABLE_ISSUE s390_sched_variable_issue
10694 #undef TARGET_SCHED_REORDER
10695 #define TARGET_SCHED_REORDER s390_sched_reorder
10696 #undef TARGET_SCHED_INIT
10697 #define TARGET_SCHED_INIT s390_sched_init
10699 #undef TARGET_CANNOT_COPY_INSN_P
10700 #define TARGET_CANNOT_COPY_INSN_P s390_cannot_copy_insn_p
10701 #undef TARGET_RTX_COSTS
10702 #define TARGET_RTX_COSTS s390_rtx_costs
10703 #undef TARGET_ADDRESS_COST
10704 #define TARGET_ADDRESS_COST s390_address_cost
10705 #undef TARGET_REGISTER_MOVE_COST
10706 #define TARGET_REGISTER_MOVE_COST s390_register_move_cost
10707 #undef TARGET_MEMORY_MOVE_COST
10708 #define TARGET_MEMORY_MOVE_COST s390_memory_move_cost
10710 #undef TARGET_MACHINE_DEPENDENT_REORG
10711 #define TARGET_MACHINE_DEPENDENT_REORG s390_reorg
10713 #undef TARGET_VALID_POINTER_MODE
10714 #define TARGET_VALID_POINTER_MODE s390_valid_pointer_mode
10716 #undef TARGET_BUILD_BUILTIN_VA_LIST
10717 #define TARGET_BUILD_BUILTIN_VA_LIST s390_build_builtin_va_list
10718 #undef TARGET_EXPAND_BUILTIN_VA_START
10719 #define TARGET_EXPAND_BUILTIN_VA_START s390_va_start
10720 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
10721 #define TARGET_GIMPLIFY_VA_ARG_EXPR s390_gimplify_va_arg
10723 #undef TARGET_PROMOTE_FUNCTION_MODE
10724 #define TARGET_PROMOTE_FUNCTION_MODE s390_promote_function_mode
10725 #undef TARGET_PASS_BY_REFERENCE
10726 #define TARGET_PASS_BY_REFERENCE s390_pass_by_reference
10728 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
10729 #define TARGET_FUNCTION_OK_FOR_SIBCALL s390_function_ok_for_sibcall
10730 #undef TARGET_FUNCTION_ARG
10731 #define TARGET_FUNCTION_ARG s390_function_arg
10732 #undef TARGET_FUNCTION_ARG_ADVANCE
10733 #define TARGET_FUNCTION_ARG_ADVANCE s390_function_arg_advance
10734 #undef TARGET_FUNCTION_VALUE
10735 #define TARGET_FUNCTION_VALUE s390_function_value
10736 #undef TARGET_LIBCALL_VALUE
10737 #define TARGET_LIBCALL_VALUE s390_libcall_value
10739 #undef TARGET_FIXED_CONDITION_CODE_REGS
10740 #define TARGET_FIXED_CONDITION_CODE_REGS s390_fixed_condition_code_regs
10742 #undef TARGET_CC_MODES_COMPATIBLE
10743 #define TARGET_CC_MODES_COMPATIBLE s390_cc_modes_compatible
10745 #undef TARGET_INVALID_WITHIN_DOLOOP
10746 #define TARGET_INVALID_WITHIN_DOLOOP hook_constcharptr_const_rtx_null
10749 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
10750 #define TARGET_ASM_OUTPUT_DWARF_DTPREL s390_output_dwarf_dtprel
10753 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
10754 #undef TARGET_MANGLE_TYPE
10755 #define TARGET_MANGLE_TYPE s390_mangle_type
10758 #undef TARGET_SCALAR_MODE_SUPPORTED_P
10759 #define TARGET_SCALAR_MODE_SUPPORTED_P s390_scalar_mode_supported_p
10761 #undef TARGET_PREFERRED_RELOAD_CLASS
10762 #define TARGET_PREFERRED_RELOAD_CLASS s390_preferred_reload_class
10764 #undef TARGET_SECONDARY_RELOAD
10765 #define TARGET_SECONDARY_RELOAD s390_secondary_reload
10767 #undef TARGET_LIBGCC_CMP_RETURN_MODE
10768 #define TARGET_LIBGCC_CMP_RETURN_MODE s390_libgcc_cmp_return_mode
10770 #undef TARGET_LIBGCC_SHIFT_COUNT_MODE
10771 #define TARGET_LIBGCC_SHIFT_COUNT_MODE s390_libgcc_shift_count_mode
10773 #undef TARGET_LEGITIMATE_ADDRESS_P
10774 #define TARGET_LEGITIMATE_ADDRESS_P s390_legitimate_address_p
10776 #undef TARGET_LEGITIMATE_CONSTANT_P
10777 #define TARGET_LEGITIMATE_CONSTANT_P s390_legitimate_constant_p
10779 #undef TARGET_CAN_ELIMINATE
10780 #define TARGET_CAN_ELIMINATE s390_can_eliminate
10782 #undef TARGET_CONDITIONAL_REGISTER_USAGE
10783 #define TARGET_CONDITIONAL_REGISTER_USAGE s390_conditional_register_usage
10785 #undef TARGET_LOOP_UNROLL_ADJUST
10786 #define TARGET_LOOP_UNROLL_ADJUST s390_loop_unroll_adjust
10788 #undef TARGET_ASM_TRAMPOLINE_TEMPLATE
10789 #define TARGET_ASM_TRAMPOLINE_TEMPLATE s390_asm_trampoline_template
10790 #undef TARGET_TRAMPOLINE_INIT
10791 #define TARGET_TRAMPOLINE_INIT s390_trampoline_init
10793 #undef TARGET_UNWIND_WORD_MODE
10794 #define TARGET_UNWIND_WORD_MODE s390_unwind_word_mode
10796 struct gcc_target targetm
= TARGET_INITIALIZER
;
10798 #include "gt-s390.h"