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 Free Software Foundation, Inc.
4 Contributed by Hartmut Penner (hpenner@de.ibm.com) and
5 Ulrich Weigand (uweigand@de.ibm.com) and
6 Andreas Krebbel (Andreas.Krebbel@de.ibm.com).
8 This file is part of GCC.
10 GCC is free software; you can redistribute it and/or modify it under
11 the terms of the GNU General Public License as published by the Free
12 Software Foundation; either version 3, or (at your option) any later
15 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
16 WARRANTY; without even the implied warranty of MERCHANTABILITY or
17 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
20 You should have received a copy of the GNU General Public License
21 along with GCC; see the file COPYING3. If not see
22 <http://www.gnu.org/licenses/>. */
26 #include "coretypes.h"
32 #include "hard-reg-set.h"
33 #include "insn-config.h"
34 #include "conditions.h"
36 #include "insn-attr.h"
43 #include "diagnostic-core.h"
44 #include "basic-block.h"
45 #include "integrate.h"
48 #include "target-def.h"
50 #include "langhooks.h"
58 /* Define the specific costs for a given cpu. */
60 struct processor_costs
63 const int m
; /* cost of an M instruction. */
64 const int mghi
; /* cost of an MGHI instruction. */
65 const int mh
; /* cost of an MH instruction. */
66 const int mhi
; /* cost of an MHI instruction. */
67 const int ml
; /* cost of an ML instruction. */
68 const int mr
; /* cost of an MR instruction. */
69 const int ms
; /* cost of an MS instruction. */
70 const int msg
; /* cost of an MSG instruction. */
71 const int msgf
; /* cost of an MSGF instruction. */
72 const int msgfr
; /* cost of an MSGFR instruction. */
73 const int msgr
; /* cost of an MSGR instruction. */
74 const int msr
; /* cost of an MSR instruction. */
75 const int mult_df
; /* cost of multiplication in DFmode. */
78 const int sqxbr
; /* cost of square root in TFmode. */
79 const int sqdbr
; /* cost of square root in DFmode. */
80 const int sqebr
; /* cost of square root in SFmode. */
81 /* multiply and add */
82 const int madbr
; /* cost of multiply and add in DFmode. */
83 const int maebr
; /* cost of multiply and add in SFmode. */
95 const struct processor_costs
*s390_cost
;
98 struct processor_costs z900_cost
=
100 COSTS_N_INSNS (5), /* M */
101 COSTS_N_INSNS (10), /* MGHI */
102 COSTS_N_INSNS (5), /* MH */
103 COSTS_N_INSNS (4), /* MHI */
104 COSTS_N_INSNS (5), /* ML */
105 COSTS_N_INSNS (5), /* MR */
106 COSTS_N_INSNS (4), /* MS */
107 COSTS_N_INSNS (15), /* MSG */
108 COSTS_N_INSNS (7), /* MSGF */
109 COSTS_N_INSNS (7), /* MSGFR */
110 COSTS_N_INSNS (10), /* MSGR */
111 COSTS_N_INSNS (4), /* MSR */
112 COSTS_N_INSNS (7), /* multiplication in DFmode */
113 COSTS_N_INSNS (13), /* MXBR */
114 COSTS_N_INSNS (136), /* SQXBR */
115 COSTS_N_INSNS (44), /* SQDBR */
116 COSTS_N_INSNS (35), /* SQEBR */
117 COSTS_N_INSNS (18), /* MADBR */
118 COSTS_N_INSNS (13), /* MAEBR */
119 COSTS_N_INSNS (134), /* DXBR */
120 COSTS_N_INSNS (30), /* DDBR */
121 COSTS_N_INSNS (27), /* DEBR */
122 COSTS_N_INSNS (220), /* DLGR */
123 COSTS_N_INSNS (34), /* DLR */
124 COSTS_N_INSNS (34), /* DR */
125 COSTS_N_INSNS (32), /* DSGFR */
126 COSTS_N_INSNS (32), /* DSGR */
130 struct processor_costs z990_cost
=
132 COSTS_N_INSNS (4), /* M */
133 COSTS_N_INSNS (2), /* MGHI */
134 COSTS_N_INSNS (2), /* MH */
135 COSTS_N_INSNS (2), /* MHI */
136 COSTS_N_INSNS (4), /* ML */
137 COSTS_N_INSNS (4), /* MR */
138 COSTS_N_INSNS (5), /* MS */
139 COSTS_N_INSNS (6), /* MSG */
140 COSTS_N_INSNS (4), /* MSGF */
141 COSTS_N_INSNS (4), /* MSGFR */
142 COSTS_N_INSNS (4), /* MSGR */
143 COSTS_N_INSNS (4), /* MSR */
144 COSTS_N_INSNS (1), /* multiplication in DFmode */
145 COSTS_N_INSNS (28), /* MXBR */
146 COSTS_N_INSNS (130), /* SQXBR */
147 COSTS_N_INSNS (66), /* SQDBR */
148 COSTS_N_INSNS (38), /* SQEBR */
149 COSTS_N_INSNS (1), /* MADBR */
150 COSTS_N_INSNS (1), /* MAEBR */
151 COSTS_N_INSNS (60), /* DXBR */
152 COSTS_N_INSNS (40), /* DDBR */
153 COSTS_N_INSNS (26), /* DEBR */
154 COSTS_N_INSNS (176), /* DLGR */
155 COSTS_N_INSNS (31), /* DLR */
156 COSTS_N_INSNS (31), /* DR */
157 COSTS_N_INSNS (31), /* DSGFR */
158 COSTS_N_INSNS (31), /* DSGR */
162 struct processor_costs z9_109_cost
=
164 COSTS_N_INSNS (4), /* M */
165 COSTS_N_INSNS (2), /* MGHI */
166 COSTS_N_INSNS (2), /* MH */
167 COSTS_N_INSNS (2), /* MHI */
168 COSTS_N_INSNS (4), /* ML */
169 COSTS_N_INSNS (4), /* MR */
170 COSTS_N_INSNS (5), /* MS */
171 COSTS_N_INSNS (6), /* MSG */
172 COSTS_N_INSNS (4), /* MSGF */
173 COSTS_N_INSNS (4), /* MSGFR */
174 COSTS_N_INSNS (4), /* MSGR */
175 COSTS_N_INSNS (4), /* MSR */
176 COSTS_N_INSNS (1), /* multiplication in DFmode */
177 COSTS_N_INSNS (28), /* MXBR */
178 COSTS_N_INSNS (130), /* SQXBR */
179 COSTS_N_INSNS (66), /* SQDBR */
180 COSTS_N_INSNS (38), /* SQEBR */
181 COSTS_N_INSNS (1), /* MADBR */
182 COSTS_N_INSNS (1), /* MAEBR */
183 COSTS_N_INSNS (60), /* DXBR */
184 COSTS_N_INSNS (40), /* DDBR */
185 COSTS_N_INSNS (26), /* DEBR */
186 COSTS_N_INSNS (30), /* DLGR */
187 COSTS_N_INSNS (23), /* DLR */
188 COSTS_N_INSNS (23), /* DR */
189 COSTS_N_INSNS (24), /* DSGFR */
190 COSTS_N_INSNS (24), /* DSGR */
194 struct processor_costs z10_cost
=
196 COSTS_N_INSNS (10), /* M */
197 COSTS_N_INSNS (10), /* MGHI */
198 COSTS_N_INSNS (10), /* MH */
199 COSTS_N_INSNS (10), /* MHI */
200 COSTS_N_INSNS (10), /* ML */
201 COSTS_N_INSNS (10), /* MR */
202 COSTS_N_INSNS (10), /* MS */
203 COSTS_N_INSNS (10), /* MSG */
204 COSTS_N_INSNS (10), /* MSGF */
205 COSTS_N_INSNS (10), /* MSGFR */
206 COSTS_N_INSNS (10), /* MSGR */
207 COSTS_N_INSNS (10), /* MSR */
208 COSTS_N_INSNS (1) , /* multiplication in DFmode */
209 COSTS_N_INSNS (50), /* MXBR */
210 COSTS_N_INSNS (120), /* SQXBR */
211 COSTS_N_INSNS (52), /* SQDBR */
212 COSTS_N_INSNS (38), /* SQEBR */
213 COSTS_N_INSNS (1), /* MADBR */
214 COSTS_N_INSNS (1), /* MAEBR */
215 COSTS_N_INSNS (111), /* DXBR */
216 COSTS_N_INSNS (39), /* DDBR */
217 COSTS_N_INSNS (32), /* DEBR */
218 COSTS_N_INSNS (160), /* DLGR */
219 COSTS_N_INSNS (71), /* DLR */
220 COSTS_N_INSNS (71), /* DR */
221 COSTS_N_INSNS (71), /* DSGFR */
222 COSTS_N_INSNS (71), /* DSGR */
226 struct processor_costs z196_cost
=
228 COSTS_N_INSNS (7), /* M */
229 COSTS_N_INSNS (5), /* MGHI */
230 COSTS_N_INSNS (5), /* MH */
231 COSTS_N_INSNS (5), /* MHI */
232 COSTS_N_INSNS (7), /* ML */
233 COSTS_N_INSNS (7), /* MR */
234 COSTS_N_INSNS (6), /* MS */
235 COSTS_N_INSNS (8), /* MSG */
236 COSTS_N_INSNS (6), /* MSGF */
237 COSTS_N_INSNS (6), /* MSGFR */
238 COSTS_N_INSNS (8), /* MSGR */
239 COSTS_N_INSNS (6), /* MSR */
240 COSTS_N_INSNS (1) , /* multiplication in DFmode */
241 COSTS_N_INSNS (40), /* MXBR B+40 */
242 COSTS_N_INSNS (100), /* SQXBR B+100 */
243 COSTS_N_INSNS (42), /* SQDBR B+42 */
244 COSTS_N_INSNS (28), /* SQEBR B+28 */
245 COSTS_N_INSNS (1), /* MADBR B */
246 COSTS_N_INSNS (1), /* MAEBR B */
247 COSTS_N_INSNS (101), /* DXBR B+101 */
248 COSTS_N_INSNS (29), /* DDBR */
249 COSTS_N_INSNS (22), /* DEBR */
250 COSTS_N_INSNS (160), /* DLGR cracked */
251 COSTS_N_INSNS (160), /* DLR cracked */
252 COSTS_N_INSNS (160), /* DR expanded */
253 COSTS_N_INSNS (160), /* DSGFR cracked */
254 COSTS_N_INSNS (160), /* DSGR cracked */
257 extern int reload_completed
;
259 /* Kept up to date using the SCHED_VARIABLE_ISSUE hook. */
260 static rtx last_scheduled_insn
;
262 /* Structure used to hold the components of a S/390 memory
263 address. A legitimate address on S/390 is of the general
265 base + index + displacement
266 where any of the components is optional.
268 base and index are registers of the class ADDR_REGS,
269 displacement is an unsigned 12-bit immediate constant. */
280 /* Which cpu are we tuning for. */
281 enum processor_type s390_tune
= PROCESSOR_max
;
283 /* Which instruction set architecture to use. */
284 enum processor_type s390_arch
;
287 HOST_WIDE_INT s390_warn_framesize
= 0;
288 HOST_WIDE_INT s390_stack_size
= 0;
289 HOST_WIDE_INT s390_stack_guard
= 0;
291 /* The following structure is embedded in the machine
292 specific part of struct function. */
294 struct GTY (()) s390_frame_layout
296 /* Offset within stack frame. */
297 HOST_WIDE_INT gprs_offset
;
298 HOST_WIDE_INT f0_offset
;
299 HOST_WIDE_INT f4_offset
;
300 HOST_WIDE_INT f8_offset
;
301 HOST_WIDE_INT backchain_offset
;
303 /* Number of first and last gpr where slots in the register
304 save area are reserved for. */
305 int first_save_gpr_slot
;
306 int last_save_gpr_slot
;
308 /* Number of first and last gpr to be saved, restored. */
310 int first_restore_gpr
;
312 int last_restore_gpr
;
314 /* Bits standing for floating point registers. Set, if the
315 respective register has to be saved. Starting with reg 16 (f0)
316 at the rightmost bit.
317 Bit 15 - 8 7 6 5 4 3 2 1 0
318 fpr 15 - 8 7 5 3 1 6 4 2 0
319 reg 31 - 24 23 22 21 20 19 18 17 16 */
320 unsigned int fpr_bitmap
;
322 /* Number of floating point registers f8-f15 which must be saved. */
325 /* Set if return address needs to be saved.
326 This flag is set by s390_return_addr_rtx if it could not use
327 the initial value of r14 and therefore depends on r14 saved
329 bool save_return_addr_p
;
331 /* Size of stack frame. */
332 HOST_WIDE_INT frame_size
;
335 /* Define the structure for the machine field in struct function. */
337 struct GTY(()) machine_function
339 struct s390_frame_layout frame_layout
;
341 /* Literal pool base register. */
344 /* True if we may need to perform branch splitting. */
345 bool split_branches_pending_p
;
347 /* Some local-dynamic TLS symbol name. */
348 const char *some_ld_name
;
350 bool has_landing_pad_p
;
353 /* Few accessor macros for struct cfun->machine->s390_frame_layout. */
355 #define cfun_frame_layout (cfun->machine->frame_layout)
356 #define cfun_save_high_fprs_p (!!cfun_frame_layout.high_fprs)
357 #define cfun_gprs_save_area_size ((cfun_frame_layout.last_save_gpr_slot - \
358 cfun_frame_layout.first_save_gpr_slot + 1) * UNITS_PER_LONG)
359 #define cfun_set_fpr_bit(BITNUM) (cfun->machine->frame_layout.fpr_bitmap |= \
361 #define cfun_fpr_bit_p(BITNUM) (!!(cfun->machine->frame_layout.fpr_bitmap & \
364 /* Number of GPRs and FPRs used for argument passing. */
365 #define GP_ARG_NUM_REG 5
366 #define FP_ARG_NUM_REG (TARGET_64BIT? 4 : 2)
368 /* A couple of shortcuts. */
369 #define CONST_OK_FOR_J(x) \
370 CONST_OK_FOR_CONSTRAINT_P((x), 'J', "J")
371 #define CONST_OK_FOR_K(x) \
372 CONST_OK_FOR_CONSTRAINT_P((x), 'K', "K")
373 #define CONST_OK_FOR_Os(x) \
374 CONST_OK_FOR_CONSTRAINT_P((x), 'O', "Os")
375 #define CONST_OK_FOR_Op(x) \
376 CONST_OK_FOR_CONSTRAINT_P((x), 'O', "Op")
377 #define CONST_OK_FOR_On(x) \
378 CONST_OK_FOR_CONSTRAINT_P((x), 'O', "On")
380 #define REGNO_PAIR_OK(REGNO, MODE) \
381 (HARD_REGNO_NREGS ((REGNO), (MODE)) == 1 || !((REGNO) & 1))
383 /* That's the read ahead of the dynamic branch prediction unit in
384 bytes on a z10 (or higher) CPU. */
385 #define PREDICT_DISTANCE (TARGET_Z10 ? 384 : 2048)
387 /* Return the alignment for LABEL. We default to the -falign-labels
388 value except for the literal pool base label. */
390 s390_label_align (rtx label
)
392 rtx prev_insn
= prev_active_insn (label
);
394 if (prev_insn
== NULL_RTX
)
397 prev_insn
= single_set (prev_insn
);
399 if (prev_insn
== NULL_RTX
)
402 prev_insn
= SET_SRC (prev_insn
);
404 /* Don't align literal pool base labels. */
405 if (GET_CODE (prev_insn
) == UNSPEC
406 && XINT (prev_insn
, 1) == UNSPEC_MAIN_BASE
)
410 return align_labels_log
;
413 static enum machine_mode
414 s390_libgcc_cmp_return_mode (void)
416 return TARGET_64BIT
? DImode
: SImode
;
419 static enum machine_mode
420 s390_libgcc_shift_count_mode (void)
422 return TARGET_64BIT
? DImode
: SImode
;
425 static enum machine_mode
426 s390_unwind_word_mode (void)
428 return TARGET_64BIT
? DImode
: SImode
;
431 /* Return true if the back end supports mode MODE. */
433 s390_scalar_mode_supported_p (enum machine_mode mode
)
435 /* In contrast to the default implementation reject TImode constants on 31bit
436 TARGET_ZARCH for ABI compliance. */
437 if (!TARGET_64BIT
&& TARGET_ZARCH
&& mode
== TImode
)
440 if (DECIMAL_FLOAT_MODE_P (mode
))
441 return default_decimal_float_supported_p ();
443 return default_scalar_mode_supported_p (mode
);
446 /* Set the has_landing_pad_p flag in struct machine_function to VALUE. */
449 s390_set_has_landing_pad_p (bool value
)
451 cfun
->machine
->has_landing_pad_p
= value
;
454 /* If two condition code modes are compatible, return a condition code
455 mode which is compatible with both. Otherwise, return
458 static enum machine_mode
459 s390_cc_modes_compatible (enum machine_mode m1
, enum machine_mode m2
)
467 if (m2
== CCUmode
|| m2
== CCTmode
|| m2
== CCZ1mode
468 || m2
== CCSmode
|| m2
== CCSRmode
|| m2
== CCURmode
)
489 /* Return true if SET either doesn't set the CC register, or else
490 the source and destination have matching CC modes and that
491 CC mode is at least as constrained as REQ_MODE. */
494 s390_match_ccmode_set (rtx set
, enum machine_mode req_mode
)
496 enum machine_mode set_mode
;
498 gcc_assert (GET_CODE (set
) == SET
);
500 if (GET_CODE (SET_DEST (set
)) != REG
|| !CC_REGNO_P (REGNO (SET_DEST (set
))))
503 set_mode
= GET_MODE (SET_DEST (set
));
517 if (req_mode
!= set_mode
)
522 if (req_mode
!= CCSmode
&& req_mode
!= CCUmode
&& req_mode
!= CCTmode
523 && req_mode
!= CCSRmode
&& req_mode
!= CCURmode
)
529 if (req_mode
!= CCAmode
)
537 return (GET_MODE (SET_SRC (set
)) == set_mode
);
540 /* Return true if every SET in INSN that sets the CC register
541 has source and destination with matching CC modes and that
542 CC mode is at least as constrained as REQ_MODE.
543 If REQ_MODE is VOIDmode, always return false. */
546 s390_match_ccmode (rtx insn
, enum machine_mode req_mode
)
550 /* s390_tm_ccmode returns VOIDmode to indicate failure. */
551 if (req_mode
== VOIDmode
)
554 if (GET_CODE (PATTERN (insn
)) == SET
)
555 return s390_match_ccmode_set (PATTERN (insn
), req_mode
);
557 if (GET_CODE (PATTERN (insn
)) == PARALLEL
)
558 for (i
= 0; i
< XVECLEN (PATTERN (insn
), 0); i
++)
560 rtx set
= XVECEXP (PATTERN (insn
), 0, i
);
561 if (GET_CODE (set
) == SET
)
562 if (!s390_match_ccmode_set (set
, req_mode
))
569 /* If a test-under-mask instruction can be used to implement
570 (compare (and ... OP1) OP2), return the CC mode required
571 to do that. Otherwise, return VOIDmode.
572 MIXED is true if the instruction can distinguish between
573 CC1 and CC2 for mixed selected bits (TMxx), it is false
574 if the instruction cannot (TM). */
577 s390_tm_ccmode (rtx op1
, rtx op2
, bool mixed
)
581 /* ??? Fixme: should work on CONST_DOUBLE as well. */
582 if (GET_CODE (op1
) != CONST_INT
|| GET_CODE (op2
) != CONST_INT
)
585 /* Selected bits all zero: CC0.
586 e.g.: int a; if ((a & (16 + 128)) == 0) */
587 if (INTVAL (op2
) == 0)
590 /* Selected bits all one: CC3.
591 e.g.: int a; if ((a & (16 + 128)) == 16 + 128) */
592 if (INTVAL (op2
) == INTVAL (op1
))
595 /* Exactly two bits selected, mixed zeroes and ones: CC1 or CC2. e.g.:
597 if ((a & (16 + 128)) == 16) -> CCT1
598 if ((a & (16 + 128)) == 128) -> CCT2 */
601 bit1
= exact_log2 (INTVAL (op2
));
602 bit0
= exact_log2 (INTVAL (op1
) ^ INTVAL (op2
));
603 if (bit0
!= -1 && bit1
!= -1)
604 return bit0
> bit1
? CCT1mode
: CCT2mode
;
610 /* Given a comparison code OP (EQ, NE, etc.) and the operands
611 OP0 and OP1 of a COMPARE, return the mode to be used for the
615 s390_select_ccmode (enum rtx_code code
, rtx op0
, rtx op1
)
621 if ((GET_CODE (op0
) == NEG
|| GET_CODE (op0
) == ABS
)
622 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
624 if (GET_CODE (op0
) == PLUS
&& GET_CODE (XEXP (op0
, 1)) == CONST_INT
625 && CONST_OK_FOR_K (INTVAL (XEXP (op0
, 1))))
627 if ((GET_CODE (op0
) == PLUS
|| GET_CODE (op0
) == MINUS
628 || GET_CODE (op1
) == NEG
)
629 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
632 if (GET_CODE (op0
) == AND
)
634 /* Check whether we can potentially do it via TM. */
635 enum machine_mode ccmode
;
636 ccmode
= s390_tm_ccmode (XEXP (op0
, 1), op1
, 1);
637 if (ccmode
!= VOIDmode
)
639 /* Relax CCTmode to CCZmode to allow fall-back to AND
640 if that turns out to be beneficial. */
641 return ccmode
== CCTmode
? CCZmode
: ccmode
;
645 if (register_operand (op0
, HImode
)
646 && GET_CODE (op1
) == CONST_INT
647 && (INTVAL (op1
) == -1 || INTVAL (op1
) == 65535))
649 if (register_operand (op0
, QImode
)
650 && GET_CODE (op1
) == CONST_INT
651 && (INTVAL (op1
) == -1 || INTVAL (op1
) == 255))
660 /* The only overflow condition of NEG and ABS happens when
661 -INT_MAX is used as parameter, which stays negative. So
662 we have an overflow from a positive value to a negative.
663 Using CCAP mode the resulting cc can be used for comparisons. */
664 if ((GET_CODE (op0
) == NEG
|| GET_CODE (op0
) == ABS
)
665 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
668 /* If constants are involved in an add instruction it is possible to use
669 the resulting cc for comparisons with zero. Knowing the sign of the
670 constant the overflow behavior gets predictable. e.g.:
671 int a, b; if ((b = a + c) > 0)
672 with c as a constant value: c < 0 -> CCAN and c >= 0 -> CCAP */
673 if (GET_CODE (op0
) == PLUS
&& GET_CODE (XEXP (op0
, 1)) == CONST_INT
674 && CONST_OK_FOR_K (INTVAL (XEXP (op0
, 1))))
676 if (INTVAL (XEXP((op0
), 1)) < 0)
690 if ((GET_CODE (op0
) == SIGN_EXTEND
|| GET_CODE (op0
) == ZERO_EXTEND
)
691 && GET_CODE (op1
) != CONST_INT
)
697 if (GET_CODE (op0
) == PLUS
698 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
701 if ((GET_CODE (op0
) == SIGN_EXTEND
|| GET_CODE (op0
) == ZERO_EXTEND
)
702 && GET_CODE (op1
) != CONST_INT
)
708 if (GET_CODE (op0
) == MINUS
709 && GET_MODE_CLASS (GET_MODE (op0
)) == MODE_INT
)
712 if ((GET_CODE (op0
) == SIGN_EXTEND
|| GET_CODE (op0
) == ZERO_EXTEND
)
713 && GET_CODE (op1
) != CONST_INT
)
722 /* Replace the comparison OP0 CODE OP1 by a semantically equivalent one
723 that we can implement more efficiently. */
726 s390_canonicalize_comparison (enum rtx_code
*code
, rtx
*op0
, rtx
*op1
)
728 /* Convert ZERO_EXTRACT back to AND to enable TM patterns. */
729 if ((*code
== EQ
|| *code
== NE
)
730 && *op1
== const0_rtx
731 && GET_CODE (*op0
) == ZERO_EXTRACT
732 && GET_CODE (XEXP (*op0
, 1)) == CONST_INT
733 && GET_CODE (XEXP (*op0
, 2)) == CONST_INT
734 && SCALAR_INT_MODE_P (GET_MODE (XEXP (*op0
, 0))))
736 rtx inner
= XEXP (*op0
, 0);
737 HOST_WIDE_INT modesize
= GET_MODE_BITSIZE (GET_MODE (inner
));
738 HOST_WIDE_INT len
= INTVAL (XEXP (*op0
, 1));
739 HOST_WIDE_INT pos
= INTVAL (XEXP (*op0
, 2));
741 if (len
> 0 && len
< modesize
742 && pos
>= 0 && pos
+ len
<= modesize
743 && modesize
<= HOST_BITS_PER_WIDE_INT
)
745 unsigned HOST_WIDE_INT block
;
746 block
= ((unsigned HOST_WIDE_INT
) 1 << len
) - 1;
747 block
<<= modesize
- pos
- len
;
749 *op0
= gen_rtx_AND (GET_MODE (inner
), inner
,
750 gen_int_mode (block
, GET_MODE (inner
)));
754 /* Narrow AND of memory against immediate to enable TM. */
755 if ((*code
== EQ
|| *code
== NE
)
756 && *op1
== const0_rtx
757 && GET_CODE (*op0
) == AND
758 && GET_CODE (XEXP (*op0
, 1)) == CONST_INT
759 && SCALAR_INT_MODE_P (GET_MODE (XEXP (*op0
, 0))))
761 rtx inner
= XEXP (*op0
, 0);
762 rtx mask
= XEXP (*op0
, 1);
764 /* Ignore paradoxical SUBREGs if all extra bits are masked out. */
765 if (GET_CODE (inner
) == SUBREG
766 && SCALAR_INT_MODE_P (GET_MODE (SUBREG_REG (inner
)))
767 && (GET_MODE_SIZE (GET_MODE (inner
))
768 >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner
))))
770 & GET_MODE_MASK (GET_MODE (inner
))
771 & ~GET_MODE_MASK (GET_MODE (SUBREG_REG (inner
))))
773 inner
= SUBREG_REG (inner
);
775 /* Do not change volatile MEMs. */
776 if (MEM_P (inner
) && !MEM_VOLATILE_P (inner
))
778 int part
= s390_single_part (XEXP (*op0
, 1),
779 GET_MODE (inner
), QImode
, 0);
782 mask
= gen_int_mode (s390_extract_part (mask
, QImode
, 0), QImode
);
783 inner
= adjust_address_nv (inner
, QImode
, part
);
784 *op0
= gen_rtx_AND (QImode
, inner
, mask
);
789 /* Narrow comparisons against 0xffff to HImode if possible. */
790 if ((*code
== EQ
|| *code
== NE
)
791 && GET_CODE (*op1
) == CONST_INT
792 && INTVAL (*op1
) == 0xffff
793 && SCALAR_INT_MODE_P (GET_MODE (*op0
))
794 && (nonzero_bits (*op0
, GET_MODE (*op0
))
795 & ~(unsigned HOST_WIDE_INT
) 0xffff) == 0)
797 *op0
= gen_lowpart (HImode
, *op0
);
801 /* Remove redundant UNSPEC_CCU_TO_INT conversions if possible. */
802 if (GET_CODE (*op0
) == UNSPEC
803 && XINT (*op0
, 1) == UNSPEC_CCU_TO_INT
804 && XVECLEN (*op0
, 0) == 1
805 && GET_MODE (XVECEXP (*op0
, 0, 0)) == CCUmode
806 && GET_CODE (XVECEXP (*op0
, 0, 0)) == REG
807 && REGNO (XVECEXP (*op0
, 0, 0)) == CC_REGNUM
808 && *op1
== const0_rtx
)
810 enum rtx_code new_code
= UNKNOWN
;
813 case EQ
: new_code
= EQ
; break;
814 case NE
: new_code
= NE
; break;
815 case LT
: new_code
= GTU
; break;
816 case GT
: new_code
= LTU
; break;
817 case LE
: new_code
= GEU
; break;
818 case GE
: new_code
= LEU
; break;
822 if (new_code
!= UNKNOWN
)
824 *op0
= XVECEXP (*op0
, 0, 0);
829 /* Remove redundant UNSPEC_CCZ_TO_INT conversions if possible. */
830 if (GET_CODE (*op0
) == UNSPEC
831 && XINT (*op0
, 1) == UNSPEC_CCZ_TO_INT
832 && XVECLEN (*op0
, 0) == 1
833 && GET_MODE (XVECEXP (*op0
, 0, 0)) == CCZmode
834 && GET_CODE (XVECEXP (*op0
, 0, 0)) == REG
835 && REGNO (XVECEXP (*op0
, 0, 0)) == CC_REGNUM
836 && *op1
== const0_rtx
)
838 enum rtx_code new_code
= UNKNOWN
;
841 case EQ
: new_code
= EQ
; break;
842 case NE
: new_code
= NE
; break;
846 if (new_code
!= UNKNOWN
)
848 *op0
= XVECEXP (*op0
, 0, 0);
853 /* Simplify cascaded EQ, NE with const0_rtx. */
854 if ((*code
== NE
|| *code
== EQ
)
855 && (GET_CODE (*op0
) == EQ
|| GET_CODE (*op0
) == NE
)
856 && GET_MODE (*op0
) == SImode
857 && GET_MODE (XEXP (*op0
, 0)) == CCZ1mode
858 && REG_P (XEXP (*op0
, 0))
859 && XEXP (*op0
, 1) == const0_rtx
860 && *op1
== const0_rtx
)
862 if ((*code
== EQ
&& GET_CODE (*op0
) == NE
)
863 || (*code
== NE
&& GET_CODE (*op0
) == EQ
))
867 *op0
= XEXP (*op0
, 0);
870 /* Prefer register over memory as first operand. */
871 if (MEM_P (*op0
) && REG_P (*op1
))
873 rtx tem
= *op0
; *op0
= *op1
; *op1
= tem
;
874 *code
= swap_condition (*code
);
878 /* Emit a compare instruction suitable to implement the comparison
879 OP0 CODE OP1. Return the correct condition RTL to be placed in
880 the IF_THEN_ELSE of the conditional branch testing the result. */
883 s390_emit_compare (enum rtx_code code
, rtx op0
, rtx op1
)
885 enum machine_mode mode
= s390_select_ccmode (code
, op0
, op1
);
888 /* Do not output a redundant compare instruction if a compare_and_swap
889 pattern already computed the result and the machine modes are compatible. */
890 if (GET_MODE_CLASS (GET_MODE (op0
)) == MODE_CC
)
892 gcc_assert (s390_cc_modes_compatible (GET_MODE (op0
), mode
)
898 cc
= gen_rtx_REG (mode
, CC_REGNUM
);
899 emit_insn (gen_rtx_SET (VOIDmode
, cc
, gen_rtx_COMPARE (mode
, op0
, op1
)));
902 return gen_rtx_fmt_ee (code
, VOIDmode
, cc
, const0_rtx
);
905 /* Emit a SImode compare and swap instruction setting MEM to NEW_RTX if OLD
907 Return the correct condition RTL to be placed in the IF_THEN_ELSE of the
908 conditional branch testing the result. */
911 s390_emit_compare_and_swap (enum rtx_code code
, rtx old
, rtx mem
, rtx cmp
, rtx new_rtx
)
913 emit_insn (gen_sync_compare_and_swapsi (old
, mem
, cmp
, new_rtx
));
914 return s390_emit_compare (code
, gen_rtx_REG (CCZ1mode
, CC_REGNUM
), const0_rtx
);
917 /* Emit a jump instruction to TARGET. If COND is NULL_RTX, emit an
918 unconditional jump, else a conditional jump under condition COND. */
921 s390_emit_jump (rtx target
, rtx cond
)
925 target
= gen_rtx_LABEL_REF (VOIDmode
, target
);
927 target
= gen_rtx_IF_THEN_ELSE (VOIDmode
, cond
, target
, pc_rtx
);
929 insn
= gen_rtx_SET (VOIDmode
, pc_rtx
, target
);
930 emit_jump_insn (insn
);
933 /* Return branch condition mask to implement a branch
934 specified by CODE. Return -1 for invalid comparisons. */
937 s390_branch_condition_mask (rtx code
)
939 const int CC0
= 1 << 3;
940 const int CC1
= 1 << 2;
941 const int CC2
= 1 << 1;
942 const int CC3
= 1 << 0;
944 gcc_assert (GET_CODE (XEXP (code
, 0)) == REG
);
945 gcc_assert (REGNO (XEXP (code
, 0)) == CC_REGNUM
);
946 gcc_assert (XEXP (code
, 1) == const0_rtx
);
948 switch (GET_MODE (XEXP (code
, 0)))
952 switch (GET_CODE (code
))
955 case NE
: return CC1
| CC2
| CC3
;
961 switch (GET_CODE (code
))
964 case NE
: return CC0
| CC2
| CC3
;
970 switch (GET_CODE (code
))
973 case NE
: return CC0
| CC1
| CC3
;
979 switch (GET_CODE (code
))
982 case NE
: return CC0
| CC1
| CC2
;
988 switch (GET_CODE (code
))
990 case EQ
: return CC0
| CC2
;
991 case NE
: return CC1
| CC3
;
997 switch (GET_CODE (code
))
999 case LTU
: return CC2
| CC3
; /* carry */
1000 case GEU
: return CC0
| CC1
; /* no carry */
1006 switch (GET_CODE (code
))
1008 case GTU
: return CC0
| CC1
; /* borrow */
1009 case LEU
: return CC2
| CC3
; /* no borrow */
1015 switch (GET_CODE (code
))
1017 case EQ
: return CC0
| CC2
;
1018 case NE
: return CC1
| CC3
;
1019 case LTU
: return CC1
;
1020 case GTU
: return CC3
;
1021 case LEU
: return CC1
| CC2
;
1022 case GEU
: return CC2
| CC3
;
1027 switch (GET_CODE (code
))
1029 case EQ
: return CC0
;
1030 case NE
: return CC1
| CC2
| CC3
;
1031 case LTU
: return CC1
;
1032 case GTU
: return CC2
;
1033 case LEU
: return CC0
| CC1
;
1034 case GEU
: return CC0
| CC2
;
1040 switch (GET_CODE (code
))
1042 case EQ
: return CC0
;
1043 case NE
: return CC2
| CC1
| CC3
;
1044 case LTU
: return CC2
;
1045 case GTU
: return CC1
;
1046 case LEU
: return CC0
| CC2
;
1047 case GEU
: return CC0
| CC1
;
1053 switch (GET_CODE (code
))
1055 case EQ
: return CC0
;
1056 case NE
: return CC1
| CC2
| CC3
;
1057 case LT
: return CC1
| CC3
;
1058 case GT
: return CC2
;
1059 case LE
: return CC0
| CC1
| CC3
;
1060 case GE
: return CC0
| CC2
;
1066 switch (GET_CODE (code
))
1068 case EQ
: return CC0
;
1069 case NE
: return CC1
| CC2
| CC3
;
1070 case LT
: return CC1
;
1071 case GT
: return CC2
| CC3
;
1072 case LE
: return CC0
| CC1
;
1073 case GE
: return CC0
| CC2
| CC3
;
1079 switch (GET_CODE (code
))
1081 case EQ
: return CC0
;
1082 case NE
: return CC1
| CC2
| CC3
;
1083 case LT
: return CC1
;
1084 case GT
: return CC2
;
1085 case LE
: return CC0
| CC1
;
1086 case GE
: return CC0
| CC2
;
1087 case UNORDERED
: return CC3
;
1088 case ORDERED
: return CC0
| CC1
| CC2
;
1089 case UNEQ
: return CC0
| CC3
;
1090 case UNLT
: return CC1
| CC3
;
1091 case UNGT
: return CC2
| CC3
;
1092 case UNLE
: return CC0
| CC1
| CC3
;
1093 case UNGE
: return CC0
| CC2
| CC3
;
1094 case LTGT
: return CC1
| CC2
;
1100 switch (GET_CODE (code
))
1102 case EQ
: return CC0
;
1103 case NE
: return CC2
| CC1
| CC3
;
1104 case LT
: return CC2
;
1105 case GT
: return CC1
;
1106 case LE
: return CC0
| CC2
;
1107 case GE
: return CC0
| CC1
;
1108 case UNORDERED
: return CC3
;
1109 case ORDERED
: return CC0
| CC2
| CC1
;
1110 case UNEQ
: return CC0
| CC3
;
1111 case UNLT
: return CC2
| CC3
;
1112 case UNGT
: return CC1
| CC3
;
1113 case UNLE
: return CC0
| CC2
| CC3
;
1114 case UNGE
: return CC0
| CC1
| CC3
;
1115 case LTGT
: return CC2
| CC1
;
1126 /* Return branch condition mask to implement a compare and branch
1127 specified by CODE. Return -1 for invalid comparisons. */
1130 s390_compare_and_branch_condition_mask (rtx code
)
1132 const int CC0
= 1 << 3;
1133 const int CC1
= 1 << 2;
1134 const int CC2
= 1 << 1;
1136 switch (GET_CODE (code
))
1160 /* If INV is false, return assembler mnemonic string to implement
1161 a branch specified by CODE. If INV is true, return mnemonic
1162 for the corresponding inverted branch. */
1165 s390_branch_condition_mnemonic (rtx code
, int inv
)
1169 static const char *const mnemonic
[16] =
1171 NULL
, "o", "h", "nle",
1172 "l", "nhe", "lh", "ne",
1173 "e", "nlh", "he", "nl",
1174 "le", "nh", "no", NULL
1177 if (GET_CODE (XEXP (code
, 0)) == REG
1178 && REGNO (XEXP (code
, 0)) == CC_REGNUM
1179 && XEXP (code
, 1) == const0_rtx
)
1180 mask
= s390_branch_condition_mask (code
);
1182 mask
= s390_compare_and_branch_condition_mask (code
);
1184 gcc_assert (mask
>= 0);
1189 gcc_assert (mask
>= 1 && mask
<= 14);
1191 return mnemonic
[mask
];
1194 /* Return the part of op which has a value different from def.
1195 The size of the part is determined by mode.
1196 Use this function only if you already know that op really
1197 contains such a part. */
1199 unsigned HOST_WIDE_INT
1200 s390_extract_part (rtx op
, enum machine_mode mode
, int def
)
1202 unsigned HOST_WIDE_INT value
= 0;
1203 int max_parts
= HOST_BITS_PER_WIDE_INT
/ GET_MODE_BITSIZE (mode
);
1204 int part_bits
= GET_MODE_BITSIZE (mode
);
1205 unsigned HOST_WIDE_INT part_mask
1206 = ((unsigned HOST_WIDE_INT
)1 << part_bits
) - 1;
1209 for (i
= 0; i
< max_parts
; i
++)
1212 value
= (unsigned HOST_WIDE_INT
) INTVAL (op
);
1214 value
>>= part_bits
;
1216 if ((value
& part_mask
) != (def
& part_mask
))
1217 return value
& part_mask
;
1223 /* If OP is an integer constant of mode MODE with exactly one
1224 part of mode PART_MODE unequal to DEF, return the number of that
1225 part. Otherwise, return -1. */
1228 s390_single_part (rtx op
,
1229 enum machine_mode mode
,
1230 enum machine_mode part_mode
,
1233 unsigned HOST_WIDE_INT value
= 0;
1234 int n_parts
= GET_MODE_SIZE (mode
) / GET_MODE_SIZE (part_mode
);
1235 unsigned HOST_WIDE_INT part_mask
1236 = ((unsigned HOST_WIDE_INT
)1 << GET_MODE_BITSIZE (part_mode
)) - 1;
1239 if (GET_CODE (op
) != CONST_INT
)
1242 for (i
= 0; i
< n_parts
; i
++)
1245 value
= (unsigned HOST_WIDE_INT
) INTVAL (op
);
1247 value
>>= GET_MODE_BITSIZE (part_mode
);
1249 if ((value
& part_mask
) != (def
& part_mask
))
1257 return part
== -1 ? -1 : n_parts
- 1 - part
;
1260 /* Return true if IN contains a contiguous bitfield in the lower SIZE
1261 bits and no other bits are set in IN. POS and LENGTH can be used
1262 to obtain the start position and the length of the bitfield.
1264 POS gives the position of the first bit of the bitfield counting
1265 from the lowest order bit starting with zero. In order to use this
1266 value for S/390 instructions this has to be converted to "bits big
1270 s390_contiguous_bitmask_p (unsigned HOST_WIDE_INT in
, int size
,
1271 int *pos
, int *length
)
1276 unsigned HOST_WIDE_INT mask
= 1ULL;
1277 bool contiguous
= false;
1279 for (i
= 0; i
< size
; mask
<<= 1, i
++)
1303 /* Calculate a mask for all bits beyond the contiguous bits. */
1304 mask
= (-1LL & ~(((1ULL << (tmp_length
+ tmp_pos
- 1)) << 1) - 1));
1309 if (tmp_length
+ tmp_pos
- 1 > size
)
1313 *length
= tmp_length
;
1321 /* Check whether we can (and want to) split a double-word
1322 move in mode MODE from SRC to DST into two single-word
1323 moves, moving the subword FIRST_SUBWORD first. */
1326 s390_split_ok_p (rtx dst
, rtx src
, enum machine_mode mode
, int first_subword
)
1328 /* Floating point registers cannot be split. */
1329 if (FP_REG_P (src
) || FP_REG_P (dst
))
1332 /* We don't need to split if operands are directly accessible. */
1333 if (s_operand (src
, mode
) || s_operand (dst
, mode
))
1336 /* Non-offsettable memory references cannot be split. */
1337 if ((GET_CODE (src
) == MEM
&& !offsettable_memref_p (src
))
1338 || (GET_CODE (dst
) == MEM
&& !offsettable_memref_p (dst
)))
1341 /* Moving the first subword must not clobber a register
1342 needed to move the second subword. */
1343 if (register_operand (dst
, mode
))
1345 rtx subreg
= operand_subword (dst
, first_subword
, 0, mode
);
1346 if (reg_overlap_mentioned_p (subreg
, src
))
1353 /* Return true if it can be proven that [MEM1, MEM1 + SIZE]
1354 and [MEM2, MEM2 + SIZE] do overlap and false
1358 s390_overlap_p (rtx mem1
, rtx mem2
, HOST_WIDE_INT size
)
1360 rtx addr1
, addr2
, addr_delta
;
1361 HOST_WIDE_INT delta
;
1363 if (GET_CODE (mem1
) != MEM
|| GET_CODE (mem2
) != MEM
)
1369 addr1
= XEXP (mem1
, 0);
1370 addr2
= XEXP (mem2
, 0);
1372 addr_delta
= simplify_binary_operation (MINUS
, Pmode
, addr2
, addr1
);
1374 /* This overlapping check is used by peepholes merging memory block operations.
1375 Overlapping operations would otherwise be recognized by the S/390 hardware
1376 and would fall back to a slower implementation. Allowing overlapping
1377 operations would lead to slow code but not to wrong code. Therefore we are
1378 somewhat optimistic if we cannot prove that the memory blocks are
1380 That's why we return false here although this may accept operations on
1381 overlapping memory areas. */
1382 if (!addr_delta
|| GET_CODE (addr_delta
) != CONST_INT
)
1385 delta
= INTVAL (addr_delta
);
1388 || (delta
> 0 && delta
< size
)
1389 || (delta
< 0 && -delta
< size
))
1395 /* Check whether the address of memory reference MEM2 equals exactly
1396 the address of memory reference MEM1 plus DELTA. Return true if
1397 we can prove this to be the case, false otherwise. */
1400 s390_offset_p (rtx mem1
, rtx mem2
, rtx delta
)
1402 rtx addr1
, addr2
, addr_delta
;
1404 if (GET_CODE (mem1
) != MEM
|| GET_CODE (mem2
) != MEM
)
1407 addr1
= XEXP (mem1
, 0);
1408 addr2
= XEXP (mem2
, 0);
1410 addr_delta
= simplify_binary_operation (MINUS
, Pmode
, addr2
, addr1
);
1411 if (!addr_delta
|| !rtx_equal_p (addr_delta
, delta
))
1417 /* Expand logical operator CODE in mode MODE with operands OPERANDS. */
1420 s390_expand_logical_operator (enum rtx_code code
, enum machine_mode mode
,
1423 enum machine_mode wmode
= mode
;
1424 rtx dst
= operands
[0];
1425 rtx src1
= operands
[1];
1426 rtx src2
= operands
[2];
1429 /* If we cannot handle the operation directly, use a temp register. */
1430 if (!s390_logical_operator_ok_p (operands
))
1431 dst
= gen_reg_rtx (mode
);
1433 /* QImode and HImode patterns make sense only if we have a destination
1434 in memory. Otherwise perform the operation in SImode. */
1435 if ((mode
== QImode
|| mode
== HImode
) && GET_CODE (dst
) != MEM
)
1438 /* Widen operands if required. */
1441 if (GET_CODE (dst
) == SUBREG
1442 && (tem
= simplify_subreg (wmode
, dst
, mode
, 0)) != 0)
1444 else if (REG_P (dst
))
1445 dst
= gen_rtx_SUBREG (wmode
, dst
, 0);
1447 dst
= gen_reg_rtx (wmode
);
1449 if (GET_CODE (src1
) == SUBREG
1450 && (tem
= simplify_subreg (wmode
, src1
, mode
, 0)) != 0)
1452 else if (GET_MODE (src1
) != VOIDmode
)
1453 src1
= gen_rtx_SUBREG (wmode
, force_reg (mode
, src1
), 0);
1455 if (GET_CODE (src2
) == SUBREG
1456 && (tem
= simplify_subreg (wmode
, src2
, mode
, 0)) != 0)
1458 else if (GET_MODE (src2
) != VOIDmode
)
1459 src2
= gen_rtx_SUBREG (wmode
, force_reg (mode
, src2
), 0);
1462 /* Emit the instruction. */
1463 op
= gen_rtx_SET (VOIDmode
, dst
, gen_rtx_fmt_ee (code
, wmode
, src1
, src2
));
1464 clob
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
1465 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, op
, clob
)));
1467 /* Fix up the destination if needed. */
1468 if (dst
!= operands
[0])
1469 emit_move_insn (operands
[0], gen_lowpart (mode
, dst
));
1472 /* Check whether OPERANDS are OK for a logical operation (AND, IOR, XOR). */
1475 s390_logical_operator_ok_p (rtx
*operands
)
1477 /* If the destination operand is in memory, it needs to coincide
1478 with one of the source operands. After reload, it has to be
1479 the first source operand. */
1480 if (GET_CODE (operands
[0]) == MEM
)
1481 return rtx_equal_p (operands
[0], operands
[1])
1482 || (!reload_completed
&& rtx_equal_p (operands
[0], operands
[2]));
1487 /* Narrow logical operation CODE of memory operand MEMOP with immediate
1488 operand IMMOP to switch from SS to SI type instructions. */
1491 s390_narrow_logical_operator (enum rtx_code code
, rtx
*memop
, rtx
*immop
)
1493 int def
= code
== AND
? -1 : 0;
1497 gcc_assert (GET_CODE (*memop
) == MEM
);
1498 gcc_assert (!MEM_VOLATILE_P (*memop
));
1500 mask
= s390_extract_part (*immop
, QImode
, def
);
1501 part
= s390_single_part (*immop
, GET_MODE (*memop
), QImode
, def
);
1502 gcc_assert (part
>= 0);
1504 *memop
= adjust_address (*memop
, QImode
, part
);
1505 *immop
= gen_int_mode (mask
, QImode
);
1509 /* How to allocate a 'struct machine_function'. */
1511 static struct machine_function
*
1512 s390_init_machine_status (void)
1514 return ggc_alloc_cleared_machine_function ();
1517 /* Change optimizations to be performed, depending on the
1518 optimization level. */
1520 static const struct default_options s390_option_optimization_table
[] =
1522 { OPT_LEVELS_1_PLUS
, OPT_fomit_frame_pointer
, NULL
, 1 },
1524 /* ??? There are apparently still problems with -fcaller-saves. */
1525 { OPT_LEVELS_ALL
, OPT_fcaller_saves
, NULL
, 0 },
1527 /* Use MVCLE instructions to decrease code size if requested. */
1528 { OPT_LEVELS_SIZE
, OPT_mmvcle
, NULL
, 1 },
1530 { OPT_LEVELS_NONE
, 0, NULL
, 0 }
1533 /* Implement TARGET_OPTION_INIT_STRUCT. */
1536 s390_option_init_struct (struct gcc_options
*opts
)
1538 /* By default, always emit DWARF-2 unwind info. This allows debugging
1539 without maintaining a stack frame back-chain. */
1540 opts
->x_flag_asynchronous_unwind_tables
= 1;
1543 /* Return true if ARG is the name of a processor. Set *TYPE and *FLAGS
1544 to the associated processor_type and processor_flags if so. */
1547 s390_handle_arch_option (const char *arg
,
1548 enum processor_type
*type
,
1553 const char *const name
; /* processor name or nickname. */
1554 const enum processor_type processor
;
1555 const int flags
; /* From enum processor_flags. */
1557 const processor_alias_table
[] =
1559 {"g5", PROCESSOR_9672_G5
, PF_IEEE_FLOAT
},
1560 {"g6", PROCESSOR_9672_G6
, PF_IEEE_FLOAT
},
1561 {"z900", PROCESSOR_2064_Z900
, PF_IEEE_FLOAT
| PF_ZARCH
},
1562 {"z990", PROCESSOR_2084_Z990
, PF_IEEE_FLOAT
| PF_ZARCH
1563 | PF_LONG_DISPLACEMENT
},
1564 {"z9-109", PROCESSOR_2094_Z9_109
, PF_IEEE_FLOAT
| PF_ZARCH
1565 | PF_LONG_DISPLACEMENT
| PF_EXTIMM
},
1566 {"z9-ec", PROCESSOR_2094_Z9_109
, PF_IEEE_FLOAT
| PF_ZARCH
1567 | PF_LONG_DISPLACEMENT
| PF_EXTIMM
| PF_DFP
},
1568 {"z10", PROCESSOR_2097_Z10
, PF_IEEE_FLOAT
| PF_ZARCH
1569 | PF_LONG_DISPLACEMENT
| PF_EXTIMM
| PF_DFP
| PF_Z10
},
1570 {"z196", PROCESSOR_2817_Z196
, PF_IEEE_FLOAT
| PF_ZARCH
1571 | PF_LONG_DISPLACEMENT
| PF_EXTIMM
| PF_DFP
| PF_Z10
| PF_Z196
},
1575 for (i
= 0; i
< ARRAY_SIZE (processor_alias_table
); i
++)
1576 if (strcmp (arg
, processor_alias_table
[i
].name
) == 0)
1578 *type
= processor_alias_table
[i
].processor
;
1579 *flags
= processor_alias_table
[i
].flags
;
1583 *type
= PROCESSOR_max
;
1588 /* Implement TARGET_HANDLE_OPTION. */
1591 s390_handle_option (size_t code
, const char *arg
, int value ATTRIBUTE_UNUSED
)
1596 return s390_handle_arch_option (arg
, &s390_arch
, &s390_arch_flags
);
1598 case OPT_mstack_guard_
:
1599 if (sscanf (arg
, HOST_WIDE_INT_PRINT_DEC
, &s390_stack_guard
) != 1)
1601 if (exact_log2 (s390_stack_guard
) == -1)
1602 error ("stack guard value must be an exact power of 2");
1605 case OPT_mstack_size_
:
1606 if (sscanf (arg
, HOST_WIDE_INT_PRINT_DEC
, &s390_stack_size
) != 1)
1608 if (exact_log2 (s390_stack_size
) == -1)
1609 error ("stack size must be an exact power of 2");
1613 return s390_handle_arch_option (arg
, &s390_tune
, &s390_tune_flags
);
1615 case OPT_mwarn_framesize_
:
1616 return sscanf (arg
, HOST_WIDE_INT_PRINT_DEC
, &s390_warn_framesize
) == 1;
1624 s390_option_override (void)
1626 /* Set up function hooks. */
1627 init_machine_status
= s390_init_machine_status
;
1629 /* Architecture mode defaults according to ABI. */
1630 if (!(target_flags_explicit
& MASK_ZARCH
))
1633 target_flags
|= MASK_ZARCH
;
1635 target_flags
&= ~MASK_ZARCH
;
1638 /* Determine processor architectural level. */
1639 if (!s390_arch_string
)
1641 s390_arch_string
= TARGET_ZARCH
? "z900" : "g5";
1642 s390_handle_arch_option (s390_arch_string
, &s390_arch
, &s390_arch_flags
);
1645 /* This check is triggered when the user specified a wrong -march=
1646 string and prevents subsequent error messages from being
1648 if (s390_arch
== PROCESSOR_max
)
1651 /* Determine processor to tune for. */
1652 if (s390_tune
== PROCESSOR_max
)
1654 s390_tune
= s390_arch
;
1655 s390_tune_flags
= s390_arch_flags
;
1658 /* Sanity checks. */
1659 if (TARGET_ZARCH
&& !TARGET_CPU_ZARCH
)
1660 error ("z/Architecture mode not supported on %s", s390_arch_string
);
1661 if (TARGET_64BIT
&& !TARGET_ZARCH
)
1662 error ("64-bit ABI not supported in ESA/390 mode");
1664 if (TARGET_HARD_DFP
&& !TARGET_DFP
)
1666 if (target_flags_explicit
& MASK_HARD_DFP
)
1668 if (!TARGET_CPU_DFP
)
1669 error ("hardware decimal floating point instructions"
1670 " not available on %s", s390_arch_string
);
1672 error ("hardware decimal floating point instructions"
1673 " not available in ESA/390 mode");
1676 target_flags
&= ~MASK_HARD_DFP
;
1679 if ((target_flags_explicit
& MASK_SOFT_FLOAT
) && TARGET_SOFT_FLOAT
)
1681 if ((target_flags_explicit
& MASK_HARD_DFP
) && TARGET_HARD_DFP
)
1682 error ("-mhard-dfp can%'t be used in conjunction with -msoft-float");
1684 target_flags
&= ~MASK_HARD_DFP
;
1687 /* Set processor cost function. */
1690 case PROCESSOR_2084_Z990
:
1691 s390_cost
= &z990_cost
;
1693 case PROCESSOR_2094_Z9_109
:
1694 s390_cost
= &z9_109_cost
;
1696 case PROCESSOR_2097_Z10
:
1697 s390_cost
= &z10_cost
;
1698 case PROCESSOR_2817_Z196
:
1699 s390_cost
= &z196_cost
;
1702 s390_cost
= &z900_cost
;
1705 if (TARGET_BACKCHAIN
&& TARGET_PACKED_STACK
&& TARGET_HARD_FLOAT
)
1706 error ("-mbackchain -mpacked-stack -mhard-float are not supported "
1709 if (s390_stack_size
)
1711 if (s390_stack_guard
>= s390_stack_size
)
1712 error ("stack size must be greater than the stack guard value");
1713 else if (s390_stack_size
> 1 << 16)
1714 error ("stack size must not be greater than 64k");
1716 else if (s390_stack_guard
)
1717 error ("-mstack-guard implies use of -mstack-size");
1719 #ifdef TARGET_DEFAULT_LONG_DOUBLE_128
1720 if (!(target_flags_explicit
& MASK_LONG_DOUBLE_128
))
1721 target_flags
|= MASK_LONG_DOUBLE_128
;
1724 if (s390_tune
== PROCESSOR_2097_Z10
1725 || s390_tune
== PROCESSOR_2817_Z196
)
1727 maybe_set_param_value (PARAM_MAX_UNROLLED_INSNS
, 100,
1728 global_options
.x_param_values
,
1729 global_options_set
.x_param_values
);
1730 maybe_set_param_value (PARAM_MAX_UNROLL_TIMES
, 32,
1731 global_options
.x_param_values
,
1732 global_options_set
.x_param_values
);
1733 maybe_set_param_value (PARAM_MAX_COMPLETELY_PEELED_INSNS
, 2000,
1734 global_options
.x_param_values
,
1735 global_options_set
.x_param_values
);
1736 maybe_set_param_value (PARAM_MAX_COMPLETELY_PEEL_TIMES
, 64,
1737 global_options
.x_param_values
,
1738 global_options_set
.x_param_values
);
1741 maybe_set_param_value (PARAM_MAX_PENDING_LIST_LENGTH
, 256,
1742 global_options
.x_param_values
,
1743 global_options_set
.x_param_values
);
1744 /* values for loop prefetching */
1745 maybe_set_param_value (PARAM_L1_CACHE_LINE_SIZE
, 256,
1746 global_options
.x_param_values
,
1747 global_options_set
.x_param_values
);
1748 maybe_set_param_value (PARAM_L1_CACHE_SIZE
, 128,
1749 global_options
.x_param_values
,
1750 global_options_set
.x_param_values
);
1751 /* s390 has more than 2 levels and the size is much larger. Since
1752 we are always running virtualized assume that we only get a small
1753 part of the caches above l1. */
1754 maybe_set_param_value (PARAM_L2_CACHE_SIZE
, 1500,
1755 global_options
.x_param_values
,
1756 global_options_set
.x_param_values
);
1757 maybe_set_param_value (PARAM_PREFETCH_MIN_INSN_TO_MEM_RATIO
, 2,
1758 global_options
.x_param_values
,
1759 global_options_set
.x_param_values
);
1760 maybe_set_param_value (PARAM_SIMULTANEOUS_PREFETCHES
, 6,
1761 global_options
.x_param_values
,
1762 global_options_set
.x_param_values
);
1764 /* This cannot reside in s390_option_optimization_table since HAVE_prefetch
1765 requires the arch flags to be evaluated already. Since prefetching
1766 is beneficial on s390, we enable it if available. */
1767 if (flag_prefetch_loop_arrays
< 0 && HAVE_prefetch
&& optimize
>= 3)
1768 flag_prefetch_loop_arrays
= 1;
1771 /* Map for smallest class containing reg regno. */
1773 const enum reg_class regclass_map
[FIRST_PSEUDO_REGISTER
] =
1774 { GENERAL_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1775 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1776 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1777 ADDR_REGS
, ADDR_REGS
, ADDR_REGS
, ADDR_REGS
,
1778 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1779 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1780 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1781 FP_REGS
, FP_REGS
, FP_REGS
, FP_REGS
,
1782 ADDR_REGS
, CC_REGS
, ADDR_REGS
, ADDR_REGS
,
1783 ACCESS_REGS
, ACCESS_REGS
1786 /* Return attribute type of insn. */
1788 static enum attr_type
1789 s390_safe_attr_type (rtx insn
)
1791 if (recog_memoized (insn
) >= 0)
1792 return get_attr_type (insn
);
1797 /* Return true if DISP is a valid short displacement. */
1800 s390_short_displacement (rtx disp
)
1802 /* No displacement is OK. */
1806 /* Without the long displacement facility we don't need to
1807 distingiush between long and short displacement. */
1808 if (!TARGET_LONG_DISPLACEMENT
)
1811 /* Integer displacement in range. */
1812 if (GET_CODE (disp
) == CONST_INT
)
1813 return INTVAL (disp
) >= 0 && INTVAL (disp
) < 4096;
1815 /* GOT offset is not OK, the GOT can be large. */
1816 if (GET_CODE (disp
) == CONST
1817 && GET_CODE (XEXP (disp
, 0)) == UNSPEC
1818 && (XINT (XEXP (disp
, 0), 1) == UNSPEC_GOT
1819 || XINT (XEXP (disp
, 0), 1) == UNSPEC_GOTNTPOFF
))
1822 /* All other symbolic constants are literal pool references,
1823 which are OK as the literal pool must be small. */
1824 if (GET_CODE (disp
) == CONST
)
1830 /* Decompose a RTL expression ADDR for a memory address into
1831 its components, returned in OUT.
1833 Returns false if ADDR is not a valid memory address, true
1834 otherwise. If OUT is NULL, don't return the components,
1835 but check for validity only.
1837 Note: Only addresses in canonical form are recognized.
1838 LEGITIMIZE_ADDRESS should convert non-canonical forms to the
1839 canonical form so that they will be recognized. */
1842 s390_decompose_address (rtx addr
, struct s390_address
*out
)
1844 HOST_WIDE_INT offset
= 0;
1845 rtx base
= NULL_RTX
;
1846 rtx indx
= NULL_RTX
;
1847 rtx disp
= NULL_RTX
;
1849 bool pointer
= false;
1850 bool base_ptr
= false;
1851 bool indx_ptr
= false;
1852 bool literal_pool
= false;
1854 /* We may need to substitute the literal pool base register into the address
1855 below. However, at this point we do not know which register is going to
1856 be used as base, so we substitute the arg pointer register. This is going
1857 to be treated as holding a pointer below -- it shouldn't be used for any
1859 rtx fake_pool_base
= gen_rtx_REG (Pmode
, ARG_POINTER_REGNUM
);
1861 /* Decompose address into base + index + displacement. */
1863 if (GET_CODE (addr
) == REG
|| GET_CODE (addr
) == UNSPEC
)
1866 else if (GET_CODE (addr
) == PLUS
)
1868 rtx op0
= XEXP (addr
, 0);
1869 rtx op1
= XEXP (addr
, 1);
1870 enum rtx_code code0
= GET_CODE (op0
);
1871 enum rtx_code code1
= GET_CODE (op1
);
1873 if (code0
== REG
|| code0
== UNSPEC
)
1875 if (code1
== REG
|| code1
== UNSPEC
)
1877 indx
= op0
; /* index + base */
1883 base
= op0
; /* base + displacement */
1888 else if (code0
== PLUS
)
1890 indx
= XEXP (op0
, 0); /* index + base + disp */
1891 base
= XEXP (op0
, 1);
1902 disp
= addr
; /* displacement */
1904 /* Extract integer part of displacement. */
1908 if (GET_CODE (disp
) == CONST_INT
)
1910 offset
= INTVAL (disp
);
1913 else if (GET_CODE (disp
) == CONST
1914 && GET_CODE (XEXP (disp
, 0)) == PLUS
1915 && GET_CODE (XEXP (XEXP (disp
, 0), 1)) == CONST_INT
)
1917 offset
= INTVAL (XEXP (XEXP (disp
, 0), 1));
1918 disp
= XEXP (XEXP (disp
, 0), 0);
1922 /* Strip off CONST here to avoid special case tests later. */
1923 if (disp
&& GET_CODE (disp
) == CONST
)
1924 disp
= XEXP (disp
, 0);
1926 /* We can convert literal pool addresses to
1927 displacements by basing them off the base register. */
1928 if (disp
&& GET_CODE (disp
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (disp
))
1930 /* Either base or index must be free to hold the base register. */
1932 base
= fake_pool_base
, literal_pool
= true;
1934 indx
= fake_pool_base
, literal_pool
= true;
1938 /* Mark up the displacement. */
1939 disp
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, disp
),
1940 UNSPEC_LTREL_OFFSET
);
1943 /* Validate base register. */
1946 if (GET_CODE (base
) == UNSPEC
)
1947 switch (XINT (base
, 1))
1951 disp
= gen_rtx_UNSPEC (Pmode
,
1952 gen_rtvec (1, XVECEXP (base
, 0, 0)),
1953 UNSPEC_LTREL_OFFSET
);
1957 base
= XVECEXP (base
, 0, 1);
1960 case UNSPEC_LTREL_BASE
:
1961 if (XVECLEN (base
, 0) == 1)
1962 base
= fake_pool_base
, literal_pool
= true;
1964 base
= XVECEXP (base
, 0, 1);
1972 || (GET_MODE (base
) != SImode
1973 && GET_MODE (base
) != Pmode
))
1976 if (REGNO (base
) == STACK_POINTER_REGNUM
1977 || REGNO (base
) == FRAME_POINTER_REGNUM
1978 || ((reload_completed
|| reload_in_progress
)
1979 && frame_pointer_needed
1980 && REGNO (base
) == HARD_FRAME_POINTER_REGNUM
)
1981 || REGNO (base
) == ARG_POINTER_REGNUM
1983 && REGNO (base
) == PIC_OFFSET_TABLE_REGNUM
))
1984 pointer
= base_ptr
= true;
1986 if ((reload_completed
|| reload_in_progress
)
1987 && base
== cfun
->machine
->base_reg
)
1988 pointer
= base_ptr
= literal_pool
= true;
1991 /* Validate index register. */
1994 if (GET_CODE (indx
) == UNSPEC
)
1995 switch (XINT (indx
, 1))
1999 disp
= gen_rtx_UNSPEC (Pmode
,
2000 gen_rtvec (1, XVECEXP (indx
, 0, 0)),
2001 UNSPEC_LTREL_OFFSET
);
2005 indx
= XVECEXP (indx
, 0, 1);
2008 case UNSPEC_LTREL_BASE
:
2009 if (XVECLEN (indx
, 0) == 1)
2010 indx
= fake_pool_base
, literal_pool
= true;
2012 indx
= XVECEXP (indx
, 0, 1);
2020 || (GET_MODE (indx
) != SImode
2021 && GET_MODE (indx
) != Pmode
))
2024 if (REGNO (indx
) == STACK_POINTER_REGNUM
2025 || REGNO (indx
) == FRAME_POINTER_REGNUM
2026 || ((reload_completed
|| reload_in_progress
)
2027 && frame_pointer_needed
2028 && REGNO (indx
) == HARD_FRAME_POINTER_REGNUM
)
2029 || REGNO (indx
) == ARG_POINTER_REGNUM
2031 && REGNO (indx
) == PIC_OFFSET_TABLE_REGNUM
))
2032 pointer
= indx_ptr
= true;
2034 if ((reload_completed
|| reload_in_progress
)
2035 && indx
== cfun
->machine
->base_reg
)
2036 pointer
= indx_ptr
= literal_pool
= true;
2039 /* Prefer to use pointer as base, not index. */
2040 if (base
&& indx
&& !base_ptr
2041 && (indx_ptr
|| (!REG_POINTER (base
) && REG_POINTER (indx
))))
2048 /* Validate displacement. */
2051 /* If virtual registers are involved, the displacement will change later
2052 anyway as the virtual registers get eliminated. This could make a
2053 valid displacement invalid, but it is more likely to make an invalid
2054 displacement valid, because we sometimes access the register save area
2055 via negative offsets to one of those registers.
2056 Thus we don't check the displacement for validity here. If after
2057 elimination the displacement turns out to be invalid after all,
2058 this is fixed up by reload in any case. */
2059 if (base
!= arg_pointer_rtx
2060 && indx
!= arg_pointer_rtx
2061 && base
!= return_address_pointer_rtx
2062 && indx
!= return_address_pointer_rtx
2063 && base
!= frame_pointer_rtx
2064 && indx
!= frame_pointer_rtx
2065 && base
!= virtual_stack_vars_rtx
2066 && indx
!= virtual_stack_vars_rtx
)
2067 if (!DISP_IN_RANGE (offset
))
2072 /* All the special cases are pointers. */
2075 /* In the small-PIC case, the linker converts @GOT
2076 and @GOTNTPOFF offsets to possible displacements. */
2077 if (GET_CODE (disp
) == UNSPEC
2078 && (XINT (disp
, 1) == UNSPEC_GOT
2079 || XINT (disp
, 1) == UNSPEC_GOTNTPOFF
)
2085 /* Accept pool label offsets. */
2086 else if (GET_CODE (disp
) == UNSPEC
2087 && XINT (disp
, 1) == UNSPEC_POOL_OFFSET
)
2090 /* Accept literal pool references. */
2091 else if (GET_CODE (disp
) == UNSPEC
2092 && XINT (disp
, 1) == UNSPEC_LTREL_OFFSET
)
2094 /* In case CSE pulled a non literal pool reference out of
2095 the pool we have to reject the address. This is
2096 especially important when loading the GOT pointer on non
2097 zarch CPUs. In this case the literal pool contains an lt
2098 relative offset to the _GLOBAL_OFFSET_TABLE_ label which
2099 will most likely exceed the displacement. */
2100 if (GET_CODE (XVECEXP (disp
, 0, 0)) != SYMBOL_REF
2101 || !CONSTANT_POOL_ADDRESS_P (XVECEXP (disp
, 0, 0)))
2104 orig_disp
= gen_rtx_CONST (Pmode
, disp
);
2107 /* If we have an offset, make sure it does not
2108 exceed the size of the constant pool entry. */
2109 rtx sym
= XVECEXP (disp
, 0, 0);
2110 if (offset
>= GET_MODE_SIZE (get_pool_mode (sym
)))
2113 orig_disp
= plus_constant (orig_disp
, offset
);
2128 out
->disp
= orig_disp
;
2129 out
->pointer
= pointer
;
2130 out
->literal_pool
= literal_pool
;
2136 /* Decompose a RTL expression OP for a shift count into its components,
2137 and return the base register in BASE and the offset in OFFSET.
2139 Return true if OP is a valid shift count, false if not. */
2142 s390_decompose_shift_count (rtx op
, rtx
*base
, HOST_WIDE_INT
*offset
)
2144 HOST_WIDE_INT off
= 0;
2146 /* We can have an integer constant, an address register,
2147 or a sum of the two. */
2148 if (GET_CODE (op
) == CONST_INT
)
2153 if (op
&& GET_CODE (op
) == PLUS
&& GET_CODE (XEXP (op
, 1)) == CONST_INT
)
2155 off
= INTVAL (XEXP (op
, 1));
2158 while (op
&& GET_CODE (op
) == SUBREG
)
2159 op
= SUBREG_REG (op
);
2161 if (op
&& GET_CODE (op
) != REG
)
2173 /* Return true if CODE is a valid address without index. */
2176 s390_legitimate_address_without_index_p (rtx op
)
2178 struct s390_address addr
;
2180 if (!s390_decompose_address (XEXP (op
, 0), &addr
))
2189 /* Return true if ADDR is of kind symbol_ref or symbol_ref + const_int
2190 and return these parts in SYMREF and ADDEND. You can pass NULL in
2191 SYMREF and/or ADDEND if you are not interested in these values.
2192 Literal pool references are *not* considered symbol references. */
2195 s390_symref_operand_p (rtx addr
, rtx
*symref
, HOST_WIDE_INT
*addend
)
2197 HOST_WIDE_INT tmpaddend
= 0;
2199 if (GET_CODE (addr
) == CONST
)
2200 addr
= XEXP (addr
, 0);
2202 if (GET_CODE (addr
) == PLUS
)
2204 if (GET_CODE (XEXP (addr
, 0)) == SYMBOL_REF
2205 && !CONSTANT_POOL_ADDRESS_P (XEXP (addr
, 0))
2206 && CONST_INT_P (XEXP (addr
, 1)))
2208 tmpaddend
= INTVAL (XEXP (addr
, 1));
2209 addr
= XEXP (addr
, 0);
2215 if (GET_CODE (addr
) != SYMBOL_REF
|| CONSTANT_POOL_ADDRESS_P (addr
))
2221 *addend
= tmpaddend
;
2227 /* Return true if the address in OP is valid for constraint letter C
2228 if wrapped in a MEM rtx. Set LIT_POOL_OK to true if it literal
2229 pool MEMs should be accepted. Only the Q, R, S, T constraint
2230 letters are allowed for C. */
2233 s390_check_qrst_address (char c
, rtx op
, bool lit_pool_ok
)
2235 struct s390_address addr
;
2236 bool decomposed
= false;
2238 /* This check makes sure that no symbolic address (except literal
2239 pool references) are accepted by the R or T constraints. */
2240 if (s390_symref_operand_p (op
, NULL
, NULL
))
2243 /* Ensure literal pool references are only accepted if LIT_POOL_OK. */
2246 if (!s390_decompose_address (op
, &addr
))
2248 if (addr
.literal_pool
)
2255 case 'Q': /* no index short displacement */
2256 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2260 if (!s390_short_displacement (addr
.disp
))
2264 case 'R': /* with index short displacement */
2265 if (TARGET_LONG_DISPLACEMENT
)
2267 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2269 if (!s390_short_displacement (addr
.disp
))
2272 /* Any invalid address here will be fixed up by reload,
2273 so accept it for the most generic constraint. */
2276 case 'S': /* no index long displacement */
2277 if (!TARGET_LONG_DISPLACEMENT
)
2279 if (!decomposed
&& !s390_decompose_address (op
, &addr
))
2283 if (s390_short_displacement (addr
.disp
))
2287 case 'T': /* with index long displacement */
2288 if (!TARGET_LONG_DISPLACEMENT
)
2290 /* Any invalid address here will be fixed up by reload,
2291 so accept it for the most generic constraint. */
2292 if ((decomposed
|| s390_decompose_address (op
, &addr
))
2293 && s390_short_displacement (addr
.disp
))
2303 /* Evaluates constraint strings described by the regular expression
2304 ([A|B|Z](Q|R|S|T))|U|W|Y and returns 1 if OP is a valid operand for
2305 the constraint given in STR, or 0 else. */
2308 s390_mem_constraint (const char *str
, rtx op
)
2315 /* Check for offsettable variants of memory constraints. */
2316 if (!MEM_P (op
) || MEM_VOLATILE_P (op
))
2318 if ((reload_completed
|| reload_in_progress
)
2319 ? !offsettable_memref_p (op
) : !offsettable_nonstrict_memref_p (op
))
2321 return s390_check_qrst_address (str
[1], XEXP (op
, 0), true);
2323 /* Check for non-literal-pool variants of memory constraints. */
2326 return s390_check_qrst_address (str
[1], XEXP (op
, 0), false);
2331 if (GET_CODE (op
) != MEM
)
2333 return s390_check_qrst_address (c
, XEXP (op
, 0), true);
2335 return (s390_check_qrst_address ('Q', op
, true)
2336 || s390_check_qrst_address ('R', op
, true));
2338 return (s390_check_qrst_address ('S', op
, true)
2339 || s390_check_qrst_address ('T', op
, true));
2341 /* Simply check for the basic form of a shift count. Reload will
2342 take care of making sure we have a proper base register. */
2343 if (!s390_decompose_shift_count (op
, NULL
, NULL
))
2347 return s390_check_qrst_address (str
[1], op
, true);
2355 /* Evaluates constraint strings starting with letter O. Input
2356 parameter C is the second letter following the "O" in the constraint
2357 string. Returns 1 if VALUE meets the respective constraint and 0
2361 s390_O_constraint_str (const char c
, HOST_WIDE_INT value
)
2369 return trunc_int_for_mode (value
, SImode
) == value
;
2373 || s390_single_part (GEN_INT (value
), DImode
, SImode
, 0) == 1;
2376 return s390_single_part (GEN_INT (value
- 1), DImode
, SImode
, -1) == 1;
2384 /* Evaluates constraint strings starting with letter N. Parameter STR
2385 contains the letters following letter "N" in the constraint string.
2386 Returns true if VALUE matches the constraint. */
2389 s390_N_constraint_str (const char *str
, HOST_WIDE_INT value
)
2391 enum machine_mode mode
, part_mode
;
2393 int part
, part_goal
;
2399 part_goal
= str
[0] - '0';
2443 if (GET_MODE_SIZE (mode
) <= GET_MODE_SIZE (part_mode
))
2446 part
= s390_single_part (GEN_INT (value
), mode
, part_mode
, def
);
2449 if (part_goal
!= -1 && part_goal
!= part
)
2456 /* Returns true if the input parameter VALUE is a float zero. */
2459 s390_float_const_zero_p (rtx value
)
2461 return (GET_MODE_CLASS (GET_MODE (value
)) == MODE_FLOAT
2462 && value
== CONST0_RTX (GET_MODE (value
)));
2465 /* Implement TARGET_REGISTER_MOVE_COST. */
2468 s390_register_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED
,
2469 reg_class_t from
, reg_class_t to
)
2471 /* On s390, copy between fprs and gprs is expensive. */
2472 if ((reg_classes_intersect_p (from
, GENERAL_REGS
)
2473 && reg_classes_intersect_p (to
, FP_REGS
))
2474 || (reg_classes_intersect_p (from
, FP_REGS
)
2475 && reg_classes_intersect_p (to
, GENERAL_REGS
)))
2481 /* Implement TARGET_MEMORY_MOVE_COST. */
2484 s390_memory_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED
,
2485 reg_class_t rclass ATTRIBUTE_UNUSED
,
2486 bool in ATTRIBUTE_UNUSED
)
2491 /* Compute a (partial) cost for rtx X. Return true if the complete
2492 cost has been computed, and false if subexpressions should be
2493 scanned. In either case, *TOTAL contains the cost result.
2494 CODE contains GET_CODE (x), OUTER_CODE contains the code
2495 of the superexpression of x. */
2498 s390_rtx_costs (rtx x
, int code
, int outer_code
, int *total
,
2499 bool speed ATTRIBUTE_UNUSED
)
2522 *total
= COSTS_N_INSNS (1);
2527 *total
= COSTS_N_INSNS (1);
2531 switch (GET_MODE (x
))
2535 rtx left
= XEXP (x
, 0);
2536 rtx right
= XEXP (x
, 1);
2537 if (GET_CODE (right
) == CONST_INT
2538 && CONST_OK_FOR_K (INTVAL (right
)))
2539 *total
= s390_cost
->mhi
;
2540 else if (GET_CODE (left
) == SIGN_EXTEND
)
2541 *total
= s390_cost
->mh
;
2543 *total
= s390_cost
->ms
; /* msr, ms, msy */
2548 rtx left
= XEXP (x
, 0);
2549 rtx right
= XEXP (x
, 1);
2552 if (GET_CODE (right
) == CONST_INT
2553 && CONST_OK_FOR_K (INTVAL (right
)))
2554 *total
= s390_cost
->mghi
;
2555 else if (GET_CODE (left
) == SIGN_EXTEND
)
2556 *total
= s390_cost
->msgf
;
2558 *total
= s390_cost
->msg
; /* msgr, msg */
2560 else /* TARGET_31BIT */
2562 if (GET_CODE (left
) == SIGN_EXTEND
2563 && GET_CODE (right
) == SIGN_EXTEND
)
2564 /* mulsidi case: mr, m */
2565 *total
= s390_cost
->m
;
2566 else if (GET_CODE (left
) == ZERO_EXTEND
2567 && GET_CODE (right
) == ZERO_EXTEND
2568 && TARGET_CPU_ZARCH
)
2569 /* umulsidi case: ml, mlr */
2570 *total
= s390_cost
->ml
;
2572 /* Complex calculation is required. */
2573 *total
= COSTS_N_INSNS (40);
2579 *total
= s390_cost
->mult_df
;
2582 *total
= s390_cost
->mxbr
;
2590 switch (GET_MODE (x
))
2593 *total
= s390_cost
->madbr
;
2596 *total
= s390_cost
->maebr
;
2601 /* Negate in the third argument is free: FMSUB. */
2602 if (GET_CODE (XEXP (x
, 2)) == NEG
)
2604 *total
+= (rtx_cost (XEXP (x
, 0), FMA
, speed
)
2605 + rtx_cost (XEXP (x
, 1), FMA
, speed
)
2606 + rtx_cost (XEXP (XEXP (x
, 2), 0), FMA
, speed
));
2613 if (GET_MODE (x
) == TImode
) /* 128 bit division */
2614 *total
= s390_cost
->dlgr
;
2615 else if (GET_MODE (x
) == DImode
)
2617 rtx right
= XEXP (x
, 1);
2618 if (GET_CODE (right
) == ZERO_EXTEND
) /* 64 by 32 bit division */
2619 *total
= s390_cost
->dlr
;
2620 else /* 64 by 64 bit division */
2621 *total
= s390_cost
->dlgr
;
2623 else if (GET_MODE (x
) == SImode
) /* 32 bit division */
2624 *total
= s390_cost
->dlr
;
2629 if (GET_MODE (x
) == DImode
)
2631 rtx right
= XEXP (x
, 1);
2632 if (GET_CODE (right
) == ZERO_EXTEND
) /* 64 by 32 bit division */
2634 *total
= s390_cost
->dsgfr
;
2636 *total
= s390_cost
->dr
;
2637 else /* 64 by 64 bit division */
2638 *total
= s390_cost
->dsgr
;
2640 else if (GET_MODE (x
) == SImode
) /* 32 bit division */
2641 *total
= s390_cost
->dlr
;
2642 else if (GET_MODE (x
) == SFmode
)
2644 *total
= s390_cost
->debr
;
2646 else if (GET_MODE (x
) == DFmode
)
2648 *total
= s390_cost
->ddbr
;
2650 else if (GET_MODE (x
) == TFmode
)
2652 *total
= s390_cost
->dxbr
;
2657 if (GET_MODE (x
) == SFmode
)
2658 *total
= s390_cost
->sqebr
;
2659 else if (GET_MODE (x
) == DFmode
)
2660 *total
= s390_cost
->sqdbr
;
2662 *total
= s390_cost
->sqxbr
;
2667 if (outer_code
== MULT
|| outer_code
== DIV
|| outer_code
== MOD
2668 || outer_code
== PLUS
|| outer_code
== MINUS
2669 || outer_code
== COMPARE
)
2674 *total
= COSTS_N_INSNS (1);
2675 if (GET_CODE (XEXP (x
, 0)) == AND
2676 && GET_CODE (XEXP (x
, 1)) == CONST_INT
2677 && GET_CODE (XEXP (XEXP (x
, 0), 1)) == CONST_INT
)
2679 rtx op0
= XEXP (XEXP (x
, 0), 0);
2680 rtx op1
= XEXP (XEXP (x
, 0), 1);
2681 rtx op2
= XEXP (x
, 1);
2683 if (memory_operand (op0
, GET_MODE (op0
))
2684 && s390_tm_ccmode (op1
, op2
, 0) != VOIDmode
)
2686 if (register_operand (op0
, GET_MODE (op0
))
2687 && s390_tm_ccmode (op1
, op2
, 1) != VOIDmode
)
2697 /* Return the cost of an address rtx ADDR. */
2700 s390_address_cost (rtx addr
, bool speed ATTRIBUTE_UNUSED
)
2702 struct s390_address ad
;
2703 if (!s390_decompose_address (addr
, &ad
))
2706 return ad
.indx
? COSTS_N_INSNS (1) + 1 : COSTS_N_INSNS (1);
2709 /* If OP is a SYMBOL_REF of a thread-local symbol, return its TLS mode,
2710 otherwise return 0. */
2713 tls_symbolic_operand (rtx op
)
2715 if (GET_CODE (op
) != SYMBOL_REF
)
2717 return SYMBOL_REF_TLS_MODEL (op
);
2720 /* Split DImode access register reference REG (on 64-bit) into its constituent
2721 low and high parts, and store them into LO and HI. Note that gen_lowpart/
2722 gen_highpart cannot be used as they assume all registers are word-sized,
2723 while our access registers have only half that size. */
2726 s390_split_access_reg (rtx reg
, rtx
*lo
, rtx
*hi
)
2728 gcc_assert (TARGET_64BIT
);
2729 gcc_assert (ACCESS_REG_P (reg
));
2730 gcc_assert (GET_MODE (reg
) == DImode
);
2731 gcc_assert (!(REGNO (reg
) & 1));
2733 *lo
= gen_rtx_REG (SImode
, REGNO (reg
) + 1);
2734 *hi
= gen_rtx_REG (SImode
, REGNO (reg
));
2737 /* Return true if OP contains a symbol reference */
2740 symbolic_reference_mentioned_p (rtx op
)
2745 if (GET_CODE (op
) == SYMBOL_REF
|| GET_CODE (op
) == LABEL_REF
)
2748 fmt
= GET_RTX_FORMAT (GET_CODE (op
));
2749 for (i
= GET_RTX_LENGTH (GET_CODE (op
)) - 1; i
>= 0; i
--)
2755 for (j
= XVECLEN (op
, i
) - 1; j
>= 0; j
--)
2756 if (symbolic_reference_mentioned_p (XVECEXP (op
, i
, j
)))
2760 else if (fmt
[i
] == 'e' && symbolic_reference_mentioned_p (XEXP (op
, i
)))
2767 /* Return true if OP contains a reference to a thread-local symbol. */
2770 tls_symbolic_reference_mentioned_p (rtx op
)
2775 if (GET_CODE (op
) == SYMBOL_REF
)
2776 return tls_symbolic_operand (op
);
2778 fmt
= GET_RTX_FORMAT (GET_CODE (op
));
2779 for (i
= GET_RTX_LENGTH (GET_CODE (op
)) - 1; i
>= 0; i
--)
2785 for (j
= XVECLEN (op
, i
) - 1; j
>= 0; j
--)
2786 if (tls_symbolic_reference_mentioned_p (XVECEXP (op
, i
, j
)))
2790 else if (fmt
[i
] == 'e' && tls_symbolic_reference_mentioned_p (XEXP (op
, i
)))
2798 /* Return true if OP is a legitimate general operand when
2799 generating PIC code. It is given that flag_pic is on
2800 and that OP satisfies CONSTANT_P or is a CONST_DOUBLE. */
2803 legitimate_pic_operand_p (rtx op
)
2805 /* Accept all non-symbolic constants. */
2806 if (!SYMBOLIC_CONST (op
))
2809 /* Reject everything else; must be handled
2810 via emit_symbolic_move. */
2814 /* Returns true if the constant value OP is a legitimate general operand.
2815 It is given that OP satisfies CONSTANT_P or is a CONST_DOUBLE. */
2818 legitimate_constant_p (rtx op
)
2820 /* Accept all non-symbolic constants. */
2821 if (!SYMBOLIC_CONST (op
))
2824 /* Accept immediate LARL operands. */
2825 if (TARGET_CPU_ZARCH
&& larl_operand (op
, VOIDmode
))
2828 /* Thread-local symbols are never legal constants. This is
2829 so that emit_call knows that computing such addresses
2830 might require a function call. */
2831 if (TLS_SYMBOLIC_CONST (op
))
2834 /* In the PIC case, symbolic constants must *not* be
2835 forced into the literal pool. We accept them here,
2836 so that they will be handled by emit_symbolic_move. */
2840 /* All remaining non-PIC symbolic constants are
2841 forced into the literal pool. */
2845 /* Determine if it's legal to put X into the constant pool. This
2846 is not possible if X contains the address of a symbol that is
2847 not constant (TLS) or not known at final link time (PIC). */
2850 s390_cannot_force_const_mem (rtx x
)
2852 switch (GET_CODE (x
))
2856 /* Accept all non-symbolic constants. */
2860 /* Labels are OK iff we are non-PIC. */
2861 return flag_pic
!= 0;
2864 /* 'Naked' TLS symbol references are never OK,
2865 non-TLS symbols are OK iff we are non-PIC. */
2866 if (tls_symbolic_operand (x
))
2869 return flag_pic
!= 0;
2872 return s390_cannot_force_const_mem (XEXP (x
, 0));
2875 return s390_cannot_force_const_mem (XEXP (x
, 0))
2876 || s390_cannot_force_const_mem (XEXP (x
, 1));
2879 switch (XINT (x
, 1))
2881 /* Only lt-relative or GOT-relative UNSPECs are OK. */
2882 case UNSPEC_LTREL_OFFSET
:
2890 case UNSPEC_GOTNTPOFF
:
2891 case UNSPEC_INDNTPOFF
:
2894 /* If the literal pool shares the code section, be put
2895 execute template placeholders into the pool as well. */
2897 return TARGET_CPU_ZARCH
;
2909 /* Returns true if the constant value OP is a legitimate general
2910 operand during and after reload. The difference to
2911 legitimate_constant_p is that this function will not accept
2912 a constant that would need to be forced to the literal pool
2913 before it can be used as operand.
2914 This function accepts all constants which can be loaded directly
2918 legitimate_reload_constant_p (rtx op
)
2920 /* Accept la(y) operands. */
2921 if (GET_CODE (op
) == CONST_INT
2922 && DISP_IN_RANGE (INTVAL (op
)))
2925 /* Accept l(g)hi/l(g)fi operands. */
2926 if (GET_CODE (op
) == CONST_INT
2927 && (CONST_OK_FOR_K (INTVAL (op
)) || CONST_OK_FOR_Os (INTVAL (op
))))
2930 /* Accept lliXX operands. */
2932 && GET_CODE (op
) == CONST_INT
2933 && trunc_int_for_mode (INTVAL (op
), word_mode
) == INTVAL (op
)
2934 && s390_single_part (op
, word_mode
, HImode
, 0) >= 0)
2938 && GET_CODE (op
) == CONST_INT
2939 && trunc_int_for_mode (INTVAL (op
), word_mode
) == INTVAL (op
)
2940 && s390_single_part (op
, word_mode
, SImode
, 0) >= 0)
2943 /* Accept larl operands. */
2944 if (TARGET_CPU_ZARCH
2945 && larl_operand (op
, VOIDmode
))
2948 /* Accept floating-point zero operands that fit into a single GPR. */
2949 if (GET_CODE (op
) == CONST_DOUBLE
2950 && s390_float_const_zero_p (op
)
2951 && GET_MODE_SIZE (GET_MODE (op
)) <= UNITS_PER_WORD
)
2954 /* Accept double-word operands that can be split. */
2955 if (GET_CODE (op
) == CONST_INT
2956 && trunc_int_for_mode (INTVAL (op
), word_mode
) != INTVAL (op
))
2958 enum machine_mode dword_mode
= word_mode
== SImode
? DImode
: TImode
;
2959 rtx hi
= operand_subword (op
, 0, 0, dword_mode
);
2960 rtx lo
= operand_subword (op
, 1, 0, dword_mode
);
2961 return legitimate_reload_constant_p (hi
)
2962 && legitimate_reload_constant_p (lo
);
2965 /* Everything else cannot be handled without reload. */
2969 /* Returns true if the constant value OP is a legitimate fp operand
2970 during and after reload.
2971 This function accepts all constants which can be loaded directly
2975 legitimate_reload_fp_constant_p (rtx op
)
2977 /* Accept floating-point zero operands if the load zero instruction
2980 && GET_CODE (op
) == CONST_DOUBLE
2981 && s390_float_const_zero_p (op
))
2987 /* Given an rtx OP being reloaded into a reg required to be in class RCLASS,
2988 return the class of reg to actually use. */
2991 s390_preferred_reload_class (rtx op
, reg_class_t rclass
)
2993 switch (GET_CODE (op
))
2995 /* Constants we cannot reload into general registers
2996 must be forced into the literal pool. */
2999 if (reg_class_subset_p (GENERAL_REGS
, rclass
)
3000 && legitimate_reload_constant_p (op
))
3001 return GENERAL_REGS
;
3002 else if (reg_class_subset_p (ADDR_REGS
, rclass
)
3003 && legitimate_reload_constant_p (op
))
3005 else if (reg_class_subset_p (FP_REGS
, rclass
)
3006 && legitimate_reload_fp_constant_p (op
))
3010 /* If a symbolic constant or a PLUS is reloaded,
3011 it is most likely being used as an address, so
3012 prefer ADDR_REGS. If 'class' is not a superset
3013 of ADDR_REGS, e.g. FP_REGS, reject this reload. */
3018 if (reg_class_subset_p (ADDR_REGS
, rclass
))
3030 /* Return true if ADDR is SYMBOL_REF + addend with addend being a
3031 multiple of ALIGNMENT and the SYMBOL_REF being naturally
3035 s390_check_symref_alignment (rtx addr
, HOST_WIDE_INT alignment
)
3037 HOST_WIDE_INT addend
;
3040 if (!s390_symref_operand_p (addr
, &symref
, &addend
))
3043 return (!SYMBOL_REF_NOT_NATURALLY_ALIGNED_P (symref
)
3044 && !(addend
& (alignment
- 1)));
3047 /* ADDR is moved into REG using larl. If ADDR isn't a valid larl
3048 operand SCRATCH is used to reload the even part of the address and
3052 s390_reload_larl_operand (rtx reg
, rtx addr
, rtx scratch
)
3054 HOST_WIDE_INT addend
;
3057 if (!s390_symref_operand_p (addr
, &symref
, &addend
))
3061 /* Easy case. The addend is even so larl will do fine. */
3062 emit_move_insn (reg
, addr
);
3065 /* We can leave the scratch register untouched if the target
3066 register is a valid base register. */
3067 if (REGNO (reg
) < FIRST_PSEUDO_REGISTER
3068 && REGNO_REG_CLASS (REGNO (reg
)) == ADDR_REGS
)
3071 gcc_assert (REGNO (scratch
) < FIRST_PSEUDO_REGISTER
);
3072 gcc_assert (REGNO_REG_CLASS (REGNO (scratch
)) == ADDR_REGS
);
3075 emit_move_insn (scratch
,
3076 gen_rtx_CONST (Pmode
,
3077 gen_rtx_PLUS (Pmode
, symref
,
3078 GEN_INT (addend
- 1))));
3080 emit_move_insn (scratch
, symref
);
3082 /* Increment the address using la in order to avoid clobbering cc. */
3083 emit_move_insn (reg
, gen_rtx_PLUS (Pmode
, scratch
, const1_rtx
));
3087 /* Generate what is necessary to move between REG and MEM using
3088 SCRATCH. The direction is given by TOMEM. */
3091 s390_reload_symref_address (rtx reg
, rtx mem
, rtx scratch
, bool tomem
)
3093 /* Reload might have pulled a constant out of the literal pool.
3094 Force it back in. */
3095 if (CONST_INT_P (mem
) || GET_CODE (mem
) == CONST_DOUBLE
3096 || GET_CODE (mem
) == CONST
)
3097 mem
= force_const_mem (GET_MODE (reg
), mem
);
3099 gcc_assert (MEM_P (mem
));
3101 /* For a load from memory we can leave the scratch register
3102 untouched if the target register is a valid base register. */
3104 && REGNO (reg
) < FIRST_PSEUDO_REGISTER
3105 && REGNO_REG_CLASS (REGNO (reg
)) == ADDR_REGS
3106 && GET_MODE (reg
) == GET_MODE (scratch
))
3109 /* Load address into scratch register. Since we can't have a
3110 secondary reload for a secondary reload we have to cover the case
3111 where larl would need a secondary reload here as well. */
3112 s390_reload_larl_operand (scratch
, XEXP (mem
, 0), scratch
);
3114 /* Now we can use a standard load/store to do the move. */
3116 emit_move_insn (replace_equiv_address (mem
, scratch
), reg
);
3118 emit_move_insn (reg
, replace_equiv_address (mem
, scratch
));
3121 /* Inform reload about cases where moving X with a mode MODE to a register in
3122 RCLASS requires an extra scratch or immediate register. Return the class
3123 needed for the immediate register. */
3126 s390_secondary_reload (bool in_p
, rtx x
, reg_class_t rclass_i
,
3127 enum machine_mode mode
, secondary_reload_info
*sri
)
3129 enum reg_class rclass
= (enum reg_class
) rclass_i
;
3131 /* Intermediate register needed. */
3132 if (reg_classes_intersect_p (CC_REGS
, rclass
))
3133 return GENERAL_REGS
;
3137 /* On z10 several optimizer steps may generate larl operands with
3140 && s390_symref_operand_p (x
, NULL
, NULL
)
3142 && !s390_check_symref_alignment (x
, 2))
3143 sri
->icode
= ((mode
== DImode
) ? CODE_FOR_reloaddi_larl_odd_addend_z10
3144 : CODE_FOR_reloadsi_larl_odd_addend_z10
);
3146 /* On z10 we need a scratch register when moving QI, TI or floating
3147 point mode values from or to a memory location with a SYMBOL_REF
3148 or if the symref addend of a SI or DI move is not aligned to the
3149 width of the access. */
3151 && s390_symref_operand_p (XEXP (x
, 0), NULL
, NULL
)
3152 && (mode
== QImode
|| mode
== TImode
|| FLOAT_MODE_P (mode
)
3153 || (!TARGET_ZARCH
&& mode
== DImode
)
3154 || ((mode
== HImode
|| mode
== SImode
|| mode
== DImode
)
3155 && (!s390_check_symref_alignment (XEXP (x
, 0),
3156 GET_MODE_SIZE (mode
))))))
3158 #define __SECONDARY_RELOAD_CASE(M,m) \
3161 sri->icode = in_p ? CODE_FOR_reload##m##di_toreg_z10 : \
3162 CODE_FOR_reload##m##di_tomem_z10; \
3164 sri->icode = in_p ? CODE_FOR_reload##m##si_toreg_z10 : \
3165 CODE_FOR_reload##m##si_tomem_z10; \
3168 switch (GET_MODE (x
))
3170 __SECONDARY_RELOAD_CASE (QI
, qi
);
3171 __SECONDARY_RELOAD_CASE (HI
, hi
);
3172 __SECONDARY_RELOAD_CASE (SI
, si
);
3173 __SECONDARY_RELOAD_CASE (DI
, di
);
3174 __SECONDARY_RELOAD_CASE (TI
, ti
);
3175 __SECONDARY_RELOAD_CASE (SF
, sf
);
3176 __SECONDARY_RELOAD_CASE (DF
, df
);
3177 __SECONDARY_RELOAD_CASE (TF
, tf
);
3178 __SECONDARY_RELOAD_CASE (SD
, sd
);
3179 __SECONDARY_RELOAD_CASE (DD
, dd
);
3180 __SECONDARY_RELOAD_CASE (TD
, td
);
3185 #undef __SECONDARY_RELOAD_CASE
3189 /* We need a scratch register when loading a PLUS expression which
3190 is not a legitimate operand of the LOAD ADDRESS instruction. */
3191 if (in_p
&& s390_plus_operand (x
, mode
))
3192 sri
->icode
= (TARGET_64BIT
?
3193 CODE_FOR_reloaddi_plus
: CODE_FOR_reloadsi_plus
);
3195 /* Performing a multiword move from or to memory we have to make sure the
3196 second chunk in memory is addressable without causing a displacement
3197 overflow. If that would be the case we calculate the address in
3198 a scratch register. */
3200 && GET_CODE (XEXP (x
, 0)) == PLUS
3201 && GET_CODE (XEXP (XEXP (x
, 0), 1)) == CONST_INT
3202 && !DISP_IN_RANGE (INTVAL (XEXP (XEXP (x
, 0), 1))
3203 + GET_MODE_SIZE (mode
) - 1))
3205 /* For GENERAL_REGS a displacement overflow is no problem if occurring
3206 in a s_operand address since we may fallback to lm/stm. So we only
3207 have to care about overflows in the b+i+d case. */
3208 if ((reg_classes_intersect_p (GENERAL_REGS
, rclass
)
3209 && s390_class_max_nregs (GENERAL_REGS
, mode
) > 1
3210 && GET_CODE (XEXP (XEXP (x
, 0), 0)) == PLUS
)
3211 /* For FP_REGS no lm/stm is available so this check is triggered
3212 for displacement overflows in b+i+d and b+d like addresses. */
3213 || (reg_classes_intersect_p (FP_REGS
, rclass
)
3214 && s390_class_max_nregs (FP_REGS
, mode
) > 1))
3217 sri
->icode
= (TARGET_64BIT
?
3218 CODE_FOR_reloaddi_nonoffmem_in
:
3219 CODE_FOR_reloadsi_nonoffmem_in
);
3221 sri
->icode
= (TARGET_64BIT
?
3222 CODE_FOR_reloaddi_nonoffmem_out
:
3223 CODE_FOR_reloadsi_nonoffmem_out
);
3227 /* A scratch address register is needed when a symbolic constant is
3228 copied to r0 compiling with -fPIC. In other cases the target
3229 register might be used as temporary (see legitimize_pic_address). */
3230 if (in_p
&& SYMBOLIC_CONST (x
) && flag_pic
== 2 && rclass
!= ADDR_REGS
)
3231 sri
->icode
= (TARGET_64BIT
?
3232 CODE_FOR_reloaddi_PIC_addr
:
3233 CODE_FOR_reloadsi_PIC_addr
);
3235 /* Either scratch or no register needed. */
3239 /* Generate code to load SRC, which is PLUS that is not a
3240 legitimate operand for the LA instruction, into TARGET.
3241 SCRATCH may be used as scratch register. */
3244 s390_expand_plus_operand (rtx target
, rtx src
,
3248 struct s390_address ad
;
3250 /* src must be a PLUS; get its two operands. */
3251 gcc_assert (GET_CODE (src
) == PLUS
);
3252 gcc_assert (GET_MODE (src
) == Pmode
);
3254 /* Check if any of the two operands is already scheduled
3255 for replacement by reload. This can happen e.g. when
3256 float registers occur in an address. */
3257 sum1
= find_replacement (&XEXP (src
, 0));
3258 sum2
= find_replacement (&XEXP (src
, 1));
3259 src
= gen_rtx_PLUS (Pmode
, sum1
, sum2
);
3261 /* If the address is already strictly valid, there's nothing to do. */
3262 if (!s390_decompose_address (src
, &ad
)
3263 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
3264 || (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
))))
3266 /* Otherwise, one of the operands cannot be an address register;
3267 we reload its value into the scratch register. */
3268 if (true_regnum (sum1
) < 1 || true_regnum (sum1
) > 15)
3270 emit_move_insn (scratch
, sum1
);
3273 if (true_regnum (sum2
) < 1 || true_regnum (sum2
) > 15)
3275 emit_move_insn (scratch
, sum2
);
3279 /* According to the way these invalid addresses are generated
3280 in reload.c, it should never happen (at least on s390) that
3281 *neither* of the PLUS components, after find_replacements
3282 was applied, is an address register. */
3283 if (sum1
== scratch
&& sum2
== scratch
)
3289 src
= gen_rtx_PLUS (Pmode
, sum1
, sum2
);
3292 /* Emit the LOAD ADDRESS pattern. Note that reload of PLUS
3293 is only ever performed on addresses, so we can mark the
3294 sum as legitimate for LA in any case. */
3295 s390_load_address (target
, src
);
3299 /* Return true if ADDR is a valid memory address.
3300 STRICT specifies whether strict register checking applies. */
3303 s390_legitimate_address_p (enum machine_mode mode
, rtx addr
, bool strict
)
3305 struct s390_address ad
;
3308 && larl_operand (addr
, VOIDmode
)
3309 && (mode
== VOIDmode
3310 || s390_check_symref_alignment (addr
, GET_MODE_SIZE (mode
))))
3313 if (!s390_decompose_address (addr
, &ad
))
3318 if (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
3321 if (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
)))
3327 && !(REGNO (ad
.base
) >= FIRST_PSEUDO_REGISTER
3328 || REGNO_REG_CLASS (REGNO (ad
.base
)) == ADDR_REGS
))
3332 && !(REGNO (ad
.indx
) >= FIRST_PSEUDO_REGISTER
3333 || REGNO_REG_CLASS (REGNO (ad
.indx
)) == ADDR_REGS
))
3339 /* Return true if OP is a valid operand for the LA instruction.
3340 In 31-bit, we need to prove that the result is used as an
3341 address, as LA performs only a 31-bit addition. */
3344 legitimate_la_operand_p (rtx op
)
3346 struct s390_address addr
;
3347 if (!s390_decompose_address (op
, &addr
))
3350 return (TARGET_64BIT
|| addr
.pointer
);
3353 /* Return true if it is valid *and* preferable to use LA to
3354 compute the sum of OP1 and OP2. */
3357 preferred_la_operand_p (rtx op1
, rtx op2
)
3359 struct s390_address addr
;
3361 if (op2
!= const0_rtx
)
3362 op1
= gen_rtx_PLUS (Pmode
, op1
, op2
);
3364 if (!s390_decompose_address (op1
, &addr
))
3366 if (addr
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (addr
.base
)))
3368 if (addr
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (addr
.indx
)))
3371 /* Avoid LA instructions with index register on z196; it is
3372 preferable to use regular add instructions when possible. */
3373 if (addr
.indx
&& s390_tune
== PROCESSOR_2817_Z196
)
3376 if (!TARGET_64BIT
&& !addr
.pointer
)
3382 if ((addr
.base
&& REG_P (addr
.base
) && REG_POINTER (addr
.base
))
3383 || (addr
.indx
&& REG_P (addr
.indx
) && REG_POINTER (addr
.indx
)))
3389 /* Emit a forced load-address operation to load SRC into DST.
3390 This will use the LOAD ADDRESS instruction even in situations
3391 where legitimate_la_operand_p (SRC) returns false. */
3394 s390_load_address (rtx dst
, rtx src
)
3397 emit_move_insn (dst
, src
);
3399 emit_insn (gen_force_la_31 (dst
, src
));
3402 /* Return a legitimate reference for ORIG (an address) using the
3403 register REG. If REG is 0, a new pseudo is generated.
3405 There are two types of references that must be handled:
3407 1. Global data references must load the address from the GOT, via
3408 the PIC reg. An insn is emitted to do this load, and the reg is
3411 2. Static data references, constant pool addresses, and code labels
3412 compute the address as an offset from the GOT, whose base is in
3413 the PIC reg. Static data objects have SYMBOL_FLAG_LOCAL set to
3414 differentiate them from global data objects. The returned
3415 address is the PIC reg + an unspec constant.
3417 TARGET_LEGITIMIZE_ADDRESS_P rejects symbolic references unless the PIC
3418 reg also appears in the address. */
3421 legitimize_pic_address (rtx orig
, rtx reg
)
3427 gcc_assert (!TLS_SYMBOLIC_CONST (addr
));
3429 if (GET_CODE (addr
) == LABEL_REF
3430 || (GET_CODE (addr
) == SYMBOL_REF
&& SYMBOL_REF_LOCAL_P (addr
)))
3432 /* This is a local symbol. */
3433 if (TARGET_CPU_ZARCH
&& larl_operand (addr
, VOIDmode
))
3435 /* Access local symbols PC-relative via LARL.
3436 This is the same as in the non-PIC case, so it is
3437 handled automatically ... */
3441 /* Access local symbols relative to the GOT. */
3443 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3445 if (reload_in_progress
|| reload_completed
)
3446 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3448 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTOFF
);
3449 addr
= gen_rtx_CONST (Pmode
, addr
);
3450 addr
= force_const_mem (Pmode
, addr
);
3451 emit_move_insn (temp
, addr
);
3453 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3456 s390_load_address (reg
, new_rtx
);
3461 else if (GET_CODE (addr
) == SYMBOL_REF
)
3464 reg
= gen_reg_rtx (Pmode
);
3468 /* Assume GOT offset < 4k. This is handled the same way
3469 in both 31- and 64-bit code (@GOT). */
3471 if (reload_in_progress
|| reload_completed
)
3472 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3474 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOT
);
3475 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3476 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, new_rtx
);
3477 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3478 emit_move_insn (reg
, new_rtx
);
3481 else if (TARGET_CPU_ZARCH
)
3483 /* If the GOT offset might be >= 4k, we determine the position
3484 of the GOT entry via a PC-relative LARL (@GOTENT). */
3486 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3488 gcc_assert (REGNO (temp
) >= FIRST_PSEUDO_REGISTER
3489 || REGNO_REG_CLASS (REGNO (temp
)) == ADDR_REGS
);
3491 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTENT
);
3492 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3493 emit_move_insn (temp
, new_rtx
);
3495 new_rtx
= gen_const_mem (Pmode
, temp
);
3496 emit_move_insn (reg
, new_rtx
);
3501 /* If the GOT offset might be >= 4k, we have to load it
3502 from the literal pool (@GOT). */
3504 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3506 gcc_assert (REGNO (temp
) >= FIRST_PSEUDO_REGISTER
3507 || REGNO_REG_CLASS (REGNO (temp
)) == ADDR_REGS
);
3509 if (reload_in_progress
|| reload_completed
)
3510 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3512 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOT
);
3513 addr
= gen_rtx_CONST (Pmode
, addr
);
3514 addr
= force_const_mem (Pmode
, addr
);
3515 emit_move_insn (temp
, addr
);
3517 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3518 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3519 emit_move_insn (reg
, new_rtx
);
3525 if (GET_CODE (addr
) == CONST
)
3527 addr
= XEXP (addr
, 0);
3528 if (GET_CODE (addr
) == UNSPEC
)
3530 gcc_assert (XVECLEN (addr
, 0) == 1);
3531 switch (XINT (addr
, 1))
3533 /* If someone moved a GOT-relative UNSPEC
3534 out of the literal pool, force them back in. */
3537 new_rtx
= force_const_mem (Pmode
, orig
);
3540 /* @GOT is OK as is if small. */
3543 new_rtx
= force_const_mem (Pmode
, orig
);
3546 /* @GOTENT is OK as is. */
3550 /* @PLT is OK as is on 64-bit, must be converted to
3551 GOT-relative @PLTOFF on 31-bit. */
3553 if (!TARGET_CPU_ZARCH
)
3555 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3557 if (reload_in_progress
|| reload_completed
)
3558 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3560 addr
= XVECEXP (addr
, 0, 0);
3561 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
),
3563 addr
= gen_rtx_CONST (Pmode
, addr
);
3564 addr
= force_const_mem (Pmode
, addr
);
3565 emit_move_insn (temp
, addr
);
3567 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3570 s390_load_address (reg
, new_rtx
);
3576 /* Everything else cannot happen. */
3582 gcc_assert (GET_CODE (addr
) == PLUS
);
3584 if (GET_CODE (addr
) == PLUS
)
3586 rtx op0
= XEXP (addr
, 0), op1
= XEXP (addr
, 1);
3588 gcc_assert (!TLS_SYMBOLIC_CONST (op0
));
3589 gcc_assert (!TLS_SYMBOLIC_CONST (op1
));
3591 /* Check first to see if this is a constant offset
3592 from a local symbol reference. */
3593 if ((GET_CODE (op0
) == LABEL_REF
3594 || (GET_CODE (op0
) == SYMBOL_REF
&& SYMBOL_REF_LOCAL_P (op0
)))
3595 && GET_CODE (op1
) == CONST_INT
)
3597 if (TARGET_CPU_ZARCH
3598 && larl_operand (op0
, VOIDmode
)
3599 && INTVAL (op1
) < (HOST_WIDE_INT
)1 << 31
3600 && INTVAL (op1
) >= -((HOST_WIDE_INT
)1 << 31))
3602 if (INTVAL (op1
) & 1)
3604 /* LARL can't handle odd offsets, so emit a
3605 pair of LARL and LA. */
3606 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3608 if (!DISP_IN_RANGE (INTVAL (op1
)))
3610 HOST_WIDE_INT even
= INTVAL (op1
) - 1;
3611 op0
= gen_rtx_PLUS (Pmode
, op0
, GEN_INT (even
));
3612 op0
= gen_rtx_CONST (Pmode
, op0
);
3616 emit_move_insn (temp
, op0
);
3617 new_rtx
= gen_rtx_PLUS (Pmode
, temp
, op1
);
3621 s390_load_address (reg
, new_rtx
);
3627 /* If the offset is even, we can just use LARL.
3628 This will happen automatically. */
3633 /* Access local symbols relative to the GOT. */
3635 rtx temp
= reg
? reg
: gen_reg_rtx (Pmode
);
3637 if (reload_in_progress
|| reload_completed
)
3638 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3640 addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op0
),
3642 addr
= gen_rtx_PLUS (Pmode
, addr
, op1
);
3643 addr
= gen_rtx_CONST (Pmode
, addr
);
3644 addr
= force_const_mem (Pmode
, addr
);
3645 emit_move_insn (temp
, addr
);
3647 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3650 s390_load_address (reg
, new_rtx
);
3656 /* Now, check whether it is a GOT relative symbol plus offset
3657 that was pulled out of the literal pool. Force it back in. */
3659 else if (GET_CODE (op0
) == UNSPEC
3660 && GET_CODE (op1
) == CONST_INT
3661 && XINT (op0
, 1) == UNSPEC_GOTOFF
)
3663 gcc_assert (XVECLEN (op0
, 0) == 1);
3665 new_rtx
= force_const_mem (Pmode
, orig
);
3668 /* Otherwise, compute the sum. */
3671 base
= legitimize_pic_address (XEXP (addr
, 0), reg
);
3672 new_rtx
= legitimize_pic_address (XEXP (addr
, 1),
3673 base
== reg
? NULL_RTX
: reg
);
3674 if (GET_CODE (new_rtx
) == CONST_INT
)
3675 new_rtx
= plus_constant (base
, INTVAL (new_rtx
));
3678 if (GET_CODE (new_rtx
) == PLUS
&& CONSTANT_P (XEXP (new_rtx
, 1)))
3680 base
= gen_rtx_PLUS (Pmode
, base
, XEXP (new_rtx
, 0));
3681 new_rtx
= XEXP (new_rtx
, 1);
3683 new_rtx
= gen_rtx_PLUS (Pmode
, base
, new_rtx
);
3686 if (GET_CODE (new_rtx
) == CONST
)
3687 new_rtx
= XEXP (new_rtx
, 0);
3688 new_rtx
= force_operand (new_rtx
, 0);
3695 /* Load the thread pointer into a register. */
3698 s390_get_thread_pointer (void)
3700 rtx tp
= gen_reg_rtx (Pmode
);
3702 emit_move_insn (tp
, gen_rtx_REG (Pmode
, TP_REGNUM
));
3703 mark_reg_pointer (tp
, BITS_PER_WORD
);
3708 /* Emit a tls call insn. The call target is the SYMBOL_REF stored
3709 in s390_tls_symbol which always refers to __tls_get_offset.
3710 The returned offset is written to RESULT_REG and an USE rtx is
3711 generated for TLS_CALL. */
3713 static GTY(()) rtx s390_tls_symbol
;
3716 s390_emit_tls_call_insn (rtx result_reg
, rtx tls_call
)
3720 gcc_assert (flag_pic
);
3722 if (!s390_tls_symbol
)
3723 s390_tls_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "__tls_get_offset");
3725 insn
= s390_emit_call (s390_tls_symbol
, tls_call
, result_reg
,
3726 gen_rtx_REG (Pmode
, RETURN_REGNUM
));
3728 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), result_reg
);
3729 RTL_CONST_CALL_P (insn
) = 1;
3732 /* ADDR contains a thread-local SYMBOL_REF. Generate code to compute
3733 this (thread-local) address. REG may be used as temporary. */
3736 legitimize_tls_address (rtx addr
, rtx reg
)
3738 rtx new_rtx
, tls_call
, temp
, base
, r2
, insn
;
3740 if (GET_CODE (addr
) == SYMBOL_REF
)
3741 switch (tls_symbolic_operand (addr
))
3743 case TLS_MODEL_GLOBAL_DYNAMIC
:
3745 r2
= gen_rtx_REG (Pmode
, 2);
3746 tls_call
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_TLSGD
);
3747 new_rtx
= gen_rtx_CONST (Pmode
, tls_call
);
3748 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3749 emit_move_insn (r2
, new_rtx
);
3750 s390_emit_tls_call_insn (r2
, tls_call
);
3751 insn
= get_insns ();
3754 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_NTPOFF
);
3755 temp
= gen_reg_rtx (Pmode
);
3756 emit_libcall_block (insn
, temp
, r2
, new_rtx
);
3758 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3761 s390_load_address (reg
, new_rtx
);
3766 case TLS_MODEL_LOCAL_DYNAMIC
:
3768 r2
= gen_rtx_REG (Pmode
, 2);
3769 tls_call
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, const0_rtx
), UNSPEC_TLSLDM
);
3770 new_rtx
= gen_rtx_CONST (Pmode
, tls_call
);
3771 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3772 emit_move_insn (r2
, new_rtx
);
3773 s390_emit_tls_call_insn (r2
, tls_call
);
3774 insn
= get_insns ();
3777 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, const0_rtx
), UNSPEC_TLSLDM_NTPOFF
);
3778 temp
= gen_reg_rtx (Pmode
);
3779 emit_libcall_block (insn
, temp
, r2
, new_rtx
);
3781 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3782 base
= gen_reg_rtx (Pmode
);
3783 s390_load_address (base
, new_rtx
);
3785 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_DTPOFF
);
3786 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3787 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3788 temp
= gen_reg_rtx (Pmode
);
3789 emit_move_insn (temp
, new_rtx
);
3791 new_rtx
= gen_rtx_PLUS (Pmode
, base
, temp
);
3794 s390_load_address (reg
, new_rtx
);
3799 case TLS_MODEL_INITIAL_EXEC
:
3802 /* Assume GOT offset < 4k. This is handled the same way
3803 in both 31- and 64-bit code. */
3805 if (reload_in_progress
|| reload_completed
)
3806 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3808 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTNTPOFF
);
3809 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3810 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, new_rtx
);
3811 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3812 temp
= gen_reg_rtx (Pmode
);
3813 emit_move_insn (temp
, new_rtx
);
3815 else if (TARGET_CPU_ZARCH
)
3817 /* If the GOT offset might be >= 4k, we determine the position
3818 of the GOT entry via a PC-relative LARL. */
3820 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_INDNTPOFF
);
3821 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3822 temp
= gen_reg_rtx (Pmode
);
3823 emit_move_insn (temp
, new_rtx
);
3825 new_rtx
= gen_const_mem (Pmode
, temp
);
3826 temp
= gen_reg_rtx (Pmode
);
3827 emit_move_insn (temp
, new_rtx
);
3831 /* If the GOT offset might be >= 4k, we have to load it
3832 from the literal pool. */
3834 if (reload_in_progress
|| reload_completed
)
3835 df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM
, true);
3837 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_GOTNTPOFF
);
3838 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3839 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3840 temp
= gen_reg_rtx (Pmode
);
3841 emit_move_insn (temp
, new_rtx
);
3843 new_rtx
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, temp
);
3844 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3846 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, new_rtx
, addr
), UNSPEC_TLS_LOAD
);
3847 temp
= gen_reg_rtx (Pmode
);
3848 emit_insn (gen_rtx_SET (Pmode
, temp
, new_rtx
));
3852 /* In position-dependent code, load the absolute address of
3853 the GOT entry from the literal pool. */
3855 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_INDNTPOFF
);
3856 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3857 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3858 temp
= gen_reg_rtx (Pmode
);
3859 emit_move_insn (temp
, new_rtx
);
3862 new_rtx
= gen_const_mem (Pmode
, new_rtx
);
3863 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, new_rtx
, addr
), UNSPEC_TLS_LOAD
);
3864 temp
= gen_reg_rtx (Pmode
);
3865 emit_insn (gen_rtx_SET (Pmode
, temp
, new_rtx
));
3868 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3871 s390_load_address (reg
, new_rtx
);
3876 case TLS_MODEL_LOCAL_EXEC
:
3877 new_rtx
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, addr
), UNSPEC_NTPOFF
);
3878 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3879 new_rtx
= force_const_mem (Pmode
, new_rtx
);
3880 temp
= gen_reg_rtx (Pmode
);
3881 emit_move_insn (temp
, new_rtx
);
3883 new_rtx
= gen_rtx_PLUS (Pmode
, s390_get_thread_pointer (), temp
);
3886 s390_load_address (reg
, new_rtx
);
3895 else if (GET_CODE (addr
) == CONST
&& GET_CODE (XEXP (addr
, 0)) == UNSPEC
)
3897 switch (XINT (XEXP (addr
, 0), 1))
3899 case UNSPEC_INDNTPOFF
:
3900 gcc_assert (TARGET_CPU_ZARCH
);
3909 else if (GET_CODE (addr
) == CONST
&& GET_CODE (XEXP (addr
, 0)) == PLUS
3910 && GET_CODE (XEXP (XEXP (addr
, 0), 1)) == CONST_INT
)
3912 new_rtx
= XEXP (XEXP (addr
, 0), 0);
3913 if (GET_CODE (new_rtx
) != SYMBOL_REF
)
3914 new_rtx
= gen_rtx_CONST (Pmode
, new_rtx
);
3916 new_rtx
= legitimize_tls_address (new_rtx
, reg
);
3917 new_rtx
= plus_constant (new_rtx
, INTVAL (XEXP (XEXP (addr
, 0), 1)));
3918 new_rtx
= force_operand (new_rtx
, 0);
3922 gcc_unreachable (); /* for now ... */
3927 /* Emit insns making the address in operands[1] valid for a standard
3928 move to operands[0]. operands[1] is replaced by an address which
3929 should be used instead of the former RTX to emit the move
3933 emit_symbolic_move (rtx
*operands
)
3935 rtx temp
= !can_create_pseudo_p () ? operands
[0] : gen_reg_rtx (Pmode
);
3937 if (GET_CODE (operands
[0]) == MEM
)
3938 operands
[1] = force_reg (Pmode
, operands
[1]);
3939 else if (TLS_SYMBOLIC_CONST (operands
[1]))
3940 operands
[1] = legitimize_tls_address (operands
[1], temp
);
3942 operands
[1] = legitimize_pic_address (operands
[1], temp
);
3945 /* Try machine-dependent ways of modifying an illegitimate address X
3946 to be legitimate. If we find one, return the new, valid address.
3948 OLDX is the address as it was before break_out_memory_refs was called.
3949 In some cases it is useful to look at this to decide what needs to be done.
3951 MODE is the mode of the operand pointed to by X.
3953 When -fpic is used, special handling is needed for symbolic references.
3954 See comments by legitimize_pic_address for details. */
3957 s390_legitimize_address (rtx x
, rtx oldx ATTRIBUTE_UNUSED
,
3958 enum machine_mode mode ATTRIBUTE_UNUSED
)
3960 rtx constant_term
= const0_rtx
;
3962 if (TLS_SYMBOLIC_CONST (x
))
3964 x
= legitimize_tls_address (x
, 0);
3966 if (s390_legitimate_address_p (mode
, x
, FALSE
))
3969 else if (GET_CODE (x
) == PLUS
3970 && (TLS_SYMBOLIC_CONST (XEXP (x
, 0))
3971 || TLS_SYMBOLIC_CONST (XEXP (x
, 1))))
3977 if (SYMBOLIC_CONST (x
)
3978 || (GET_CODE (x
) == PLUS
3979 && (SYMBOLIC_CONST (XEXP (x
, 0))
3980 || SYMBOLIC_CONST (XEXP (x
, 1)))))
3981 x
= legitimize_pic_address (x
, 0);
3983 if (s390_legitimate_address_p (mode
, x
, FALSE
))
3987 x
= eliminate_constant_term (x
, &constant_term
);
3989 /* Optimize loading of large displacements by splitting them
3990 into the multiple of 4K and the rest; this allows the
3991 former to be CSE'd if possible.
3993 Don't do this if the displacement is added to a register
3994 pointing into the stack frame, as the offsets will
3995 change later anyway. */
3997 if (GET_CODE (constant_term
) == CONST_INT
3998 && !TARGET_LONG_DISPLACEMENT
3999 && !DISP_IN_RANGE (INTVAL (constant_term
))
4000 && !(REG_P (x
) && REGNO_PTR_FRAME_P (REGNO (x
))))
4002 HOST_WIDE_INT lower
= INTVAL (constant_term
) & 0xfff;
4003 HOST_WIDE_INT upper
= INTVAL (constant_term
) ^ lower
;
4005 rtx temp
= gen_reg_rtx (Pmode
);
4006 rtx val
= force_operand (GEN_INT (upper
), temp
);
4008 emit_move_insn (temp
, val
);
4010 x
= gen_rtx_PLUS (Pmode
, x
, temp
);
4011 constant_term
= GEN_INT (lower
);
4014 if (GET_CODE (x
) == PLUS
)
4016 if (GET_CODE (XEXP (x
, 0)) == REG
)
4018 rtx temp
= gen_reg_rtx (Pmode
);
4019 rtx val
= force_operand (XEXP (x
, 1), temp
);
4021 emit_move_insn (temp
, val
);
4023 x
= gen_rtx_PLUS (Pmode
, XEXP (x
, 0), temp
);
4026 else if (GET_CODE (XEXP (x
, 1)) == REG
)
4028 rtx temp
= gen_reg_rtx (Pmode
);
4029 rtx val
= force_operand (XEXP (x
, 0), temp
);
4031 emit_move_insn (temp
, val
);
4033 x
= gen_rtx_PLUS (Pmode
, temp
, XEXP (x
, 1));
4037 if (constant_term
!= const0_rtx
)
4038 x
= gen_rtx_PLUS (Pmode
, x
, constant_term
);
4043 /* Try a machine-dependent way of reloading an illegitimate address AD
4044 operand. If we find one, push the reload and and return the new address.
4046 MODE is the mode of the enclosing MEM. OPNUM is the operand number
4047 and TYPE is the reload type of the current reload. */
4050 legitimize_reload_address (rtx ad
, enum machine_mode mode ATTRIBUTE_UNUSED
,
4051 int opnum
, int type
)
4053 if (!optimize
|| TARGET_LONG_DISPLACEMENT
)
4056 if (GET_CODE (ad
) == PLUS
)
4058 rtx tem
= simplify_binary_operation (PLUS
, Pmode
,
4059 XEXP (ad
, 0), XEXP (ad
, 1));
4064 if (GET_CODE (ad
) == PLUS
4065 && GET_CODE (XEXP (ad
, 0)) == REG
4066 && GET_CODE (XEXP (ad
, 1)) == CONST_INT
4067 && !DISP_IN_RANGE (INTVAL (XEXP (ad
, 1))))
4069 HOST_WIDE_INT lower
= INTVAL (XEXP (ad
, 1)) & 0xfff;
4070 HOST_WIDE_INT upper
= INTVAL (XEXP (ad
, 1)) ^ lower
;
4071 rtx cst
, tem
, new_rtx
;
4073 cst
= GEN_INT (upper
);
4074 if (!legitimate_reload_constant_p (cst
))
4075 cst
= force_const_mem (Pmode
, cst
);
4077 tem
= gen_rtx_PLUS (Pmode
, XEXP (ad
, 0), cst
);
4078 new_rtx
= gen_rtx_PLUS (Pmode
, tem
, GEN_INT (lower
));
4080 push_reload (XEXP (tem
, 1), 0, &XEXP (tem
, 1), 0,
4081 BASE_REG_CLASS
, Pmode
, VOIDmode
, 0, 0,
4082 opnum
, (enum reload_type
) type
);
4089 /* Emit code to move LEN bytes from DST to SRC. */
4092 s390_expand_movmem (rtx dst
, rtx src
, rtx len
)
4094 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) >= 0 && INTVAL (len
) <= 256)
4096 if (INTVAL (len
) > 0)
4097 emit_insn (gen_movmem_short (dst
, src
, GEN_INT (INTVAL (len
) - 1)));
4100 else if (TARGET_MVCLE
)
4102 emit_insn (gen_movmem_long (dst
, src
, convert_to_mode (Pmode
, len
, 1)));
4107 rtx dst_addr
, src_addr
, count
, blocks
, temp
;
4108 rtx loop_start_label
= gen_label_rtx ();
4109 rtx loop_end_label
= gen_label_rtx ();
4110 rtx end_label
= gen_label_rtx ();
4111 enum machine_mode mode
;
4113 mode
= GET_MODE (len
);
4114 if (mode
== VOIDmode
)
4117 dst_addr
= gen_reg_rtx (Pmode
);
4118 src_addr
= gen_reg_rtx (Pmode
);
4119 count
= gen_reg_rtx (mode
);
4120 blocks
= gen_reg_rtx (mode
);
4122 convert_move (count
, len
, 1);
4123 emit_cmp_and_jump_insns (count
, const0_rtx
,
4124 EQ
, NULL_RTX
, mode
, 1, end_label
);
4126 emit_move_insn (dst_addr
, force_operand (XEXP (dst
, 0), NULL_RTX
));
4127 emit_move_insn (src_addr
, force_operand (XEXP (src
, 0), NULL_RTX
));
4128 dst
= change_address (dst
, VOIDmode
, dst_addr
);
4129 src
= change_address (src
, VOIDmode
, src_addr
);
4131 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4134 emit_move_insn (count
, temp
);
4136 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4139 emit_move_insn (blocks
, temp
);
4141 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4142 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4144 emit_label (loop_start_label
);
4147 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 768))
4151 /* Issue a read prefetch for the +3 cache line. */
4152 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, src_addr
, GEN_INT (768)),
4153 const0_rtx
, const0_rtx
);
4154 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4155 emit_insn (prefetch
);
4157 /* Issue a write prefetch for the +3 cache line. */
4158 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (768)),
4159 const1_rtx
, const0_rtx
);
4160 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4161 emit_insn (prefetch
);
4164 emit_insn (gen_movmem_short (dst
, src
, GEN_INT (255)));
4165 s390_load_address (dst_addr
,
4166 gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (256)));
4167 s390_load_address (src_addr
,
4168 gen_rtx_PLUS (Pmode
, src_addr
, GEN_INT (256)));
4170 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4173 emit_move_insn (blocks
, temp
);
4175 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4176 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4178 emit_jump (loop_start_label
);
4179 emit_label (loop_end_label
);
4181 emit_insn (gen_movmem_short (dst
, src
,
4182 convert_to_mode (Pmode
, count
, 1)));
4183 emit_label (end_label
);
4187 /* Emit code to set LEN bytes at DST to VAL.
4188 Make use of clrmem if VAL is zero. */
4191 s390_expand_setmem (rtx dst
, rtx len
, rtx val
)
4193 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) == 0)
4196 gcc_assert (GET_CODE (val
) == CONST_INT
|| GET_MODE (val
) == QImode
);
4198 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) > 0 && INTVAL (len
) <= 257)
4200 if (val
== const0_rtx
&& INTVAL (len
) <= 256)
4201 emit_insn (gen_clrmem_short (dst
, GEN_INT (INTVAL (len
) - 1)));
4204 /* Initialize memory by storing the first byte. */
4205 emit_move_insn (adjust_address (dst
, QImode
, 0), val
);
4207 if (INTVAL (len
) > 1)
4209 /* Initiate 1 byte overlap move.
4210 The first byte of DST is propagated through DSTP1.
4211 Prepare a movmem for: DST+1 = DST (length = LEN - 1).
4212 DST is set to size 1 so the rest of the memory location
4213 does not count as source operand. */
4214 rtx dstp1
= adjust_address (dst
, VOIDmode
, 1);
4215 set_mem_size (dst
, const1_rtx
);
4217 emit_insn (gen_movmem_short (dstp1
, dst
,
4218 GEN_INT (INTVAL (len
) - 2)));
4223 else if (TARGET_MVCLE
)
4225 val
= force_not_mem (convert_modes (Pmode
, QImode
, val
, 1));
4226 emit_insn (gen_setmem_long (dst
, convert_to_mode (Pmode
, len
, 1), val
));
4231 rtx dst_addr
, count
, blocks
, temp
, dstp1
= NULL_RTX
;
4232 rtx loop_start_label
= gen_label_rtx ();
4233 rtx loop_end_label
= gen_label_rtx ();
4234 rtx end_label
= gen_label_rtx ();
4235 enum machine_mode mode
;
4237 mode
= GET_MODE (len
);
4238 if (mode
== VOIDmode
)
4241 dst_addr
= gen_reg_rtx (Pmode
);
4242 count
= gen_reg_rtx (mode
);
4243 blocks
= gen_reg_rtx (mode
);
4245 convert_move (count
, len
, 1);
4246 emit_cmp_and_jump_insns (count
, const0_rtx
,
4247 EQ
, NULL_RTX
, mode
, 1, end_label
);
4249 emit_move_insn (dst_addr
, force_operand (XEXP (dst
, 0), NULL_RTX
));
4250 dst
= change_address (dst
, VOIDmode
, dst_addr
);
4252 if (val
== const0_rtx
)
4253 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4257 dstp1
= adjust_address (dst
, VOIDmode
, 1);
4258 set_mem_size (dst
, const1_rtx
);
4260 /* Initialize memory by storing the first byte. */
4261 emit_move_insn (adjust_address (dst
, QImode
, 0), val
);
4263 /* If count is 1 we are done. */
4264 emit_cmp_and_jump_insns (count
, const1_rtx
,
4265 EQ
, NULL_RTX
, mode
, 1, end_label
);
4267 temp
= expand_binop (mode
, add_optab
, count
, GEN_INT (-2), count
, 1,
4271 emit_move_insn (count
, temp
);
4273 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4276 emit_move_insn (blocks
, temp
);
4278 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4279 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4281 emit_label (loop_start_label
);
4284 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 1024))
4286 /* Issue a write prefetch for the +4 cache line. */
4287 rtx prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, dst_addr
,
4289 const1_rtx
, const0_rtx
);
4290 emit_insn (prefetch
);
4291 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4294 if (val
== const0_rtx
)
4295 emit_insn (gen_clrmem_short (dst
, GEN_INT (255)));
4297 emit_insn (gen_movmem_short (dstp1
, dst
, GEN_INT (255)));
4298 s390_load_address (dst_addr
,
4299 gen_rtx_PLUS (Pmode
, dst_addr
, GEN_INT (256)));
4301 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4304 emit_move_insn (blocks
, temp
);
4306 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4307 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4309 emit_jump (loop_start_label
);
4310 emit_label (loop_end_label
);
4312 if (val
== const0_rtx
)
4313 emit_insn (gen_clrmem_short (dst
, convert_to_mode (Pmode
, count
, 1)));
4315 emit_insn (gen_movmem_short (dstp1
, dst
, convert_to_mode (Pmode
, count
, 1)));
4316 emit_label (end_label
);
4320 /* Emit code to compare LEN bytes at OP0 with those at OP1,
4321 and return the result in TARGET. */
4324 s390_expand_cmpmem (rtx target
, rtx op0
, rtx op1
, rtx len
)
4326 rtx ccreg
= gen_rtx_REG (CCUmode
, CC_REGNUM
);
4329 /* As the result of CMPINT is inverted compared to what we need,
4330 we have to swap the operands. */
4331 tmp
= op0
; op0
= op1
; op1
= tmp
;
4333 if (GET_CODE (len
) == CONST_INT
&& INTVAL (len
) >= 0 && INTVAL (len
) <= 256)
4335 if (INTVAL (len
) > 0)
4337 emit_insn (gen_cmpmem_short (op0
, op1
, GEN_INT (INTVAL (len
) - 1)));
4338 emit_insn (gen_cmpint (target
, ccreg
));
4341 emit_move_insn (target
, const0_rtx
);
4343 else if (TARGET_MVCLE
)
4345 emit_insn (gen_cmpmem_long (op0
, op1
, convert_to_mode (Pmode
, len
, 1)));
4346 emit_insn (gen_cmpint (target
, ccreg
));
4350 rtx addr0
, addr1
, count
, blocks
, temp
;
4351 rtx loop_start_label
= gen_label_rtx ();
4352 rtx loop_end_label
= gen_label_rtx ();
4353 rtx end_label
= gen_label_rtx ();
4354 enum machine_mode mode
;
4356 mode
= GET_MODE (len
);
4357 if (mode
== VOIDmode
)
4360 addr0
= gen_reg_rtx (Pmode
);
4361 addr1
= gen_reg_rtx (Pmode
);
4362 count
= gen_reg_rtx (mode
);
4363 blocks
= gen_reg_rtx (mode
);
4365 convert_move (count
, len
, 1);
4366 emit_cmp_and_jump_insns (count
, const0_rtx
,
4367 EQ
, NULL_RTX
, mode
, 1, end_label
);
4369 emit_move_insn (addr0
, force_operand (XEXP (op0
, 0), NULL_RTX
));
4370 emit_move_insn (addr1
, force_operand (XEXP (op1
, 0), NULL_RTX
));
4371 op0
= change_address (op0
, VOIDmode
, addr0
);
4372 op1
= change_address (op1
, VOIDmode
, addr1
);
4374 temp
= expand_binop (mode
, add_optab
, count
, constm1_rtx
, count
, 1,
4377 emit_move_insn (count
, temp
);
4379 temp
= expand_binop (mode
, lshr_optab
, count
, GEN_INT (8), blocks
, 1,
4382 emit_move_insn (blocks
, temp
);
4384 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4385 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4387 emit_label (loop_start_label
);
4390 && (GET_CODE (len
) != CONST_INT
|| INTVAL (len
) > 512))
4394 /* Issue a read prefetch for the +2 cache line of operand 1. */
4395 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, addr0
, GEN_INT (512)),
4396 const0_rtx
, const0_rtx
);
4397 emit_insn (prefetch
);
4398 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4400 /* Issue a read prefetch for the +2 cache line of operand 2. */
4401 prefetch
= gen_prefetch (gen_rtx_PLUS (Pmode
, addr1
, GEN_INT (512)),
4402 const0_rtx
, const0_rtx
);
4403 emit_insn (prefetch
);
4404 PREFETCH_SCHEDULE_BARRIER_P (prefetch
) = true;
4407 emit_insn (gen_cmpmem_short (op0
, op1
, GEN_INT (255)));
4408 temp
= gen_rtx_NE (VOIDmode
, ccreg
, const0_rtx
);
4409 temp
= gen_rtx_IF_THEN_ELSE (VOIDmode
, temp
,
4410 gen_rtx_LABEL_REF (VOIDmode
, end_label
), pc_rtx
);
4411 temp
= gen_rtx_SET (VOIDmode
, pc_rtx
, temp
);
4412 emit_jump_insn (temp
);
4414 s390_load_address (addr0
,
4415 gen_rtx_PLUS (Pmode
, addr0
, GEN_INT (256)));
4416 s390_load_address (addr1
,
4417 gen_rtx_PLUS (Pmode
, addr1
, GEN_INT (256)));
4419 temp
= expand_binop (mode
, add_optab
, blocks
, constm1_rtx
, blocks
, 1,
4422 emit_move_insn (blocks
, temp
);
4424 emit_cmp_and_jump_insns (blocks
, const0_rtx
,
4425 EQ
, NULL_RTX
, mode
, 1, loop_end_label
);
4427 emit_jump (loop_start_label
);
4428 emit_label (loop_end_label
);
4430 emit_insn (gen_cmpmem_short (op0
, op1
,
4431 convert_to_mode (Pmode
, count
, 1)));
4432 emit_label (end_label
);
4434 emit_insn (gen_cmpint (target
, ccreg
));
4439 /* Expand conditional increment or decrement using alc/slb instructions.
4440 Should generate code setting DST to either SRC or SRC + INCREMENT,
4441 depending on the result of the comparison CMP_OP0 CMP_CODE CMP_OP1.
4442 Returns true if successful, false otherwise.
4444 That makes it possible to implement some if-constructs without jumps e.g.:
4445 (borrow = CC0 | CC1 and carry = CC2 | CC3)
4446 unsigned int a, b, c;
4447 if (a < b) c++; -> CCU b > a -> CC2; c += carry;
4448 if (a < b) c--; -> CCL3 a - b -> borrow; c -= borrow;
4449 if (a <= b) c++; -> CCL3 b - a -> borrow; c += carry;
4450 if (a <= b) c--; -> CCU a <= b -> borrow; c -= borrow;
4452 Checks for EQ and NE with a nonzero value need an additional xor e.g.:
4453 if (a == b) c++; -> CCL3 a ^= b; 0 - a -> borrow; c += carry;
4454 if (a == b) c--; -> CCU a ^= b; a <= 0 -> CC0 | CC1; c -= borrow;
4455 if (a != b) c++; -> CCU a ^= b; a > 0 -> CC2; c += carry;
4456 if (a != b) c--; -> CCL3 a ^= b; 0 - a -> borrow; c -= borrow; */
4459 s390_expand_addcc (enum rtx_code cmp_code
, rtx cmp_op0
, rtx cmp_op1
,
4460 rtx dst
, rtx src
, rtx increment
)
4462 enum machine_mode cmp_mode
;
4463 enum machine_mode cc_mode
;
4469 if ((GET_MODE (cmp_op0
) == SImode
|| GET_MODE (cmp_op0
) == VOIDmode
)
4470 && (GET_MODE (cmp_op1
) == SImode
|| GET_MODE (cmp_op1
) == VOIDmode
))
4472 else if ((GET_MODE (cmp_op0
) == DImode
|| GET_MODE (cmp_op0
) == VOIDmode
)
4473 && (GET_MODE (cmp_op1
) == DImode
|| GET_MODE (cmp_op1
) == VOIDmode
))
4478 /* Try ADD LOGICAL WITH CARRY. */
4479 if (increment
== const1_rtx
)
4481 /* Determine CC mode to use. */
4482 if (cmp_code
== EQ
|| cmp_code
== NE
)
4484 if (cmp_op1
!= const0_rtx
)
4486 cmp_op0
= expand_simple_binop (cmp_mode
, XOR
, cmp_op0
, cmp_op1
,
4487 NULL_RTX
, 0, OPTAB_WIDEN
);
4488 cmp_op1
= const0_rtx
;
4491 cmp_code
= cmp_code
== EQ
? LEU
: GTU
;
4494 if (cmp_code
== LTU
|| cmp_code
== LEU
)
4499 cmp_code
= swap_condition (cmp_code
);
4516 /* Emit comparison instruction pattern. */
4517 if (!register_operand (cmp_op0
, cmp_mode
))
4518 cmp_op0
= force_reg (cmp_mode
, cmp_op0
);
4520 insn
= gen_rtx_SET (VOIDmode
, gen_rtx_REG (cc_mode
, CC_REGNUM
),
4521 gen_rtx_COMPARE (cc_mode
, cmp_op0
, cmp_op1
));
4522 /* We use insn_invalid_p here to add clobbers if required. */
4523 ret
= insn_invalid_p (emit_insn (insn
));
4526 /* Emit ALC instruction pattern. */
4527 op_res
= gen_rtx_fmt_ee (cmp_code
, GET_MODE (dst
),
4528 gen_rtx_REG (cc_mode
, CC_REGNUM
),
4531 if (src
!= const0_rtx
)
4533 if (!register_operand (src
, GET_MODE (dst
)))
4534 src
= force_reg (GET_MODE (dst
), src
);
4536 op_res
= gen_rtx_PLUS (GET_MODE (dst
), op_res
, src
);
4537 op_res
= gen_rtx_PLUS (GET_MODE (dst
), op_res
, const0_rtx
);
4540 p
= rtvec_alloc (2);
4542 gen_rtx_SET (VOIDmode
, dst
, op_res
);
4544 gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4545 emit_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
4550 /* Try SUBTRACT LOGICAL WITH BORROW. */
4551 if (increment
== constm1_rtx
)
4553 /* Determine CC mode to use. */
4554 if (cmp_code
== EQ
|| cmp_code
== NE
)
4556 if (cmp_op1
!= const0_rtx
)
4558 cmp_op0
= expand_simple_binop (cmp_mode
, XOR
, cmp_op0
, cmp_op1
,
4559 NULL_RTX
, 0, OPTAB_WIDEN
);
4560 cmp_op1
= const0_rtx
;
4563 cmp_code
= cmp_code
== EQ
? LEU
: GTU
;
4566 if (cmp_code
== GTU
|| cmp_code
== GEU
)
4571 cmp_code
= swap_condition (cmp_code
);
4588 /* Emit comparison instruction pattern. */
4589 if (!register_operand (cmp_op0
, cmp_mode
))
4590 cmp_op0
= force_reg (cmp_mode
, cmp_op0
);
4592 insn
= gen_rtx_SET (VOIDmode
, gen_rtx_REG (cc_mode
, CC_REGNUM
),
4593 gen_rtx_COMPARE (cc_mode
, cmp_op0
, cmp_op1
));
4594 /* We use insn_invalid_p here to add clobbers if required. */
4595 ret
= insn_invalid_p (emit_insn (insn
));
4598 /* Emit SLB instruction pattern. */
4599 if (!register_operand (src
, GET_MODE (dst
)))
4600 src
= force_reg (GET_MODE (dst
), src
);
4602 op_res
= gen_rtx_MINUS (GET_MODE (dst
),
4603 gen_rtx_MINUS (GET_MODE (dst
), src
, const0_rtx
),
4604 gen_rtx_fmt_ee (cmp_code
, GET_MODE (dst
),
4605 gen_rtx_REG (cc_mode
, CC_REGNUM
),
4607 p
= rtvec_alloc (2);
4609 gen_rtx_SET (VOIDmode
, dst
, op_res
);
4611 gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4612 emit_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
4620 /* Expand code for the insv template. Return true if successful. */
4623 s390_expand_insv (rtx dest
, rtx op1
, rtx op2
, rtx src
)
4625 int bitsize
= INTVAL (op1
);
4626 int bitpos
= INTVAL (op2
);
4628 /* On z10 we can use the risbg instruction to implement insv. */
4630 && ((GET_MODE (dest
) == DImode
&& GET_MODE (src
) == DImode
)
4631 || (GET_MODE (dest
) == SImode
&& GET_MODE (src
) == SImode
)))
4636 op
= gen_rtx_SET (GET_MODE(src
),
4637 gen_rtx_ZERO_EXTRACT (GET_MODE (dest
), dest
, op1
, op2
),
4639 clobber
= gen_rtx_CLOBBER (VOIDmode
, gen_rtx_REG (CCmode
, CC_REGNUM
));
4640 emit_insn (gen_rtx_PARALLEL (VOIDmode
, gen_rtvec (2, op
, clobber
)));
4645 /* We need byte alignment. */
4646 if (bitsize
% BITS_PER_UNIT
)
4650 && memory_operand (dest
, VOIDmode
)
4651 && (register_operand (src
, word_mode
)
4652 || const_int_operand (src
, VOIDmode
)))
4654 /* Emit standard pattern if possible. */
4655 enum machine_mode mode
= smallest_mode_for_size (bitsize
, MODE_INT
);
4656 if (GET_MODE_BITSIZE (mode
) == bitsize
)
4657 emit_move_insn (adjust_address (dest
, mode
, 0), gen_lowpart (mode
, src
));
4659 /* (set (ze (mem)) (const_int)). */
4660 else if (const_int_operand (src
, VOIDmode
))
4662 int size
= bitsize
/ BITS_PER_UNIT
;
4663 rtx src_mem
= adjust_address (force_const_mem (word_mode
, src
), BLKmode
,
4664 GET_MODE_SIZE (word_mode
) - size
);
4666 dest
= adjust_address (dest
, BLKmode
, 0);
4667 set_mem_size (dest
, GEN_INT (size
));
4668 s390_expand_movmem (dest
, src_mem
, GEN_INT (size
));
4671 /* (set (ze (mem)) (reg)). */
4672 else if (register_operand (src
, word_mode
))
4674 if (bitsize
<= GET_MODE_BITSIZE (SImode
))
4675 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
, op1
,
4679 /* Emit st,stcmh sequence. */
4680 int stcmh_width
= bitsize
- GET_MODE_BITSIZE (SImode
);
4681 int size
= stcmh_width
/ BITS_PER_UNIT
;
4683 emit_move_insn (adjust_address (dest
, SImode
, size
),
4684 gen_lowpart (SImode
, src
));
4685 set_mem_size (dest
, GEN_INT (size
));
4686 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
, GEN_INT
4687 (stcmh_width
), const0_rtx
),
4688 gen_rtx_LSHIFTRT (word_mode
, src
, GEN_INT
4689 (GET_MODE_BITSIZE (SImode
))));
4698 /* (set (ze (reg)) (const_int)). */
4700 && register_operand (dest
, word_mode
)
4701 && (bitpos
% 16) == 0
4702 && (bitsize
% 16) == 0
4703 && const_int_operand (src
, VOIDmode
))
4705 HOST_WIDE_INT val
= INTVAL (src
);
4706 int regpos
= bitpos
+ bitsize
;
4708 while (regpos
> bitpos
)
4710 enum machine_mode putmode
;
4713 if (TARGET_EXTIMM
&& (regpos
% 32 == 0) && (regpos
>= bitpos
+ 32))
4718 putsize
= GET_MODE_BITSIZE (putmode
);
4720 emit_move_insn (gen_rtx_ZERO_EXTRACT (word_mode
, dest
,
4723 gen_int_mode (val
, putmode
));
4726 gcc_assert (regpos
== bitpos
);
4733 /* A subroutine of s390_expand_cs_hqi and s390_expand_atomic which returns a
4734 register that holds VAL of mode MODE shifted by COUNT bits. */
4737 s390_expand_mask_and_shift (rtx val
, enum machine_mode mode
, rtx count
)
4739 val
= expand_simple_binop (SImode
, AND
, val
, GEN_INT (GET_MODE_MASK (mode
)),
4740 NULL_RTX
, 1, OPTAB_DIRECT
);
4741 return expand_simple_binop (SImode
, ASHIFT
, val
, count
,
4742 NULL_RTX
, 1, OPTAB_DIRECT
);
4745 /* Structure to hold the initial parameters for a compare_and_swap operation
4746 in HImode and QImode. */
4748 struct alignment_context
4750 rtx memsi
; /* SI aligned memory location. */
4751 rtx shift
; /* Bit offset with regard to lsb. */
4752 rtx modemask
; /* Mask of the HQImode shifted by SHIFT bits. */
4753 rtx modemaski
; /* ~modemask */
4754 bool aligned
; /* True if memory is aligned, false else. */
4757 /* A subroutine of s390_expand_cs_hqi and s390_expand_atomic to initialize
4758 structure AC for transparent simplifying, if the memory alignment is known
4759 to be at least 32bit. MEM is the memory location for the actual operation
4760 and MODE its mode. */
4763 init_alignment_context (struct alignment_context
*ac
, rtx mem
,
4764 enum machine_mode mode
)
4766 ac
->shift
= GEN_INT (GET_MODE_SIZE (SImode
) - GET_MODE_SIZE (mode
));
4767 ac
->aligned
= (MEM_ALIGN (mem
) >= GET_MODE_BITSIZE (SImode
));
4770 ac
->memsi
= adjust_address (mem
, SImode
, 0); /* Memory is aligned. */
4773 /* Alignment is unknown. */
4774 rtx byteoffset
, addr
, align
;
4776 /* Force the address into a register. */
4777 addr
= force_reg (Pmode
, XEXP (mem
, 0));
4779 /* Align it to SImode. */
4780 align
= expand_simple_binop (Pmode
, AND
, addr
,
4781 GEN_INT (-GET_MODE_SIZE (SImode
)),
4782 NULL_RTX
, 1, OPTAB_DIRECT
);
4784 ac
->memsi
= gen_rtx_MEM (SImode
, align
);
4785 MEM_VOLATILE_P (ac
->memsi
) = MEM_VOLATILE_P (mem
);
4786 set_mem_alias_set (ac
->memsi
, ALIAS_SET_MEMORY_BARRIER
);
4787 set_mem_align (ac
->memsi
, GET_MODE_BITSIZE (SImode
));
4789 /* Calculate shiftcount. */
4790 byteoffset
= expand_simple_binop (Pmode
, AND
, addr
,
4791 GEN_INT (GET_MODE_SIZE (SImode
) - 1),
4792 NULL_RTX
, 1, OPTAB_DIRECT
);
4793 /* As we already have some offset, evaluate the remaining distance. */
4794 ac
->shift
= expand_simple_binop (SImode
, MINUS
, ac
->shift
, byteoffset
,
4795 NULL_RTX
, 1, OPTAB_DIRECT
);
4798 /* Shift is the byte count, but we need the bitcount. */
4799 ac
->shift
= expand_simple_binop (SImode
, MULT
, ac
->shift
, GEN_INT (BITS_PER_UNIT
),
4800 NULL_RTX
, 1, OPTAB_DIRECT
);
4801 /* Calculate masks. */
4802 ac
->modemask
= expand_simple_binop (SImode
, ASHIFT
,
4803 GEN_INT (GET_MODE_MASK (mode
)), ac
->shift
,
4804 NULL_RTX
, 1, OPTAB_DIRECT
);
4805 ac
->modemaski
= expand_simple_unop (SImode
, NOT
, ac
->modemask
, NULL_RTX
, 1);
4808 /* Expand an atomic compare and swap operation for HImode and QImode. MEM is
4809 the memory location, CMP the old value to compare MEM with and NEW_RTX the value
4810 to set if CMP == MEM.
4811 CMP is never in memory for compare_and_swap_cc because
4812 expand_bool_compare_and_swap puts it into a register for later compare. */
4815 s390_expand_cs_hqi (enum machine_mode mode
, rtx target
, rtx mem
, rtx cmp
, rtx new_rtx
)
4817 struct alignment_context ac
;
4818 rtx cmpv
, newv
, val
, resv
, cc
;
4819 rtx res
= gen_reg_rtx (SImode
);
4820 rtx csloop
= gen_label_rtx ();
4821 rtx csend
= gen_label_rtx ();
4823 gcc_assert (register_operand (target
, VOIDmode
));
4824 gcc_assert (MEM_P (mem
));
4826 init_alignment_context (&ac
, mem
, mode
);
4828 /* Shift the values to the correct bit positions. */
4829 if (!(ac
.aligned
&& MEM_P (cmp
)))
4830 cmp
= s390_expand_mask_and_shift (cmp
, mode
, ac
.shift
);
4831 if (!(ac
.aligned
&& MEM_P (new_rtx
)))
4832 new_rtx
= s390_expand_mask_and_shift (new_rtx
, mode
, ac
.shift
);
4834 /* Load full word. Subsequent loads are performed by CS. */
4835 val
= expand_simple_binop (SImode
, AND
, ac
.memsi
, ac
.modemaski
,
4836 NULL_RTX
, 1, OPTAB_DIRECT
);
4838 /* Start CS loop. */
4839 emit_label (csloop
);
4840 /* val = "<mem>00..0<mem>"
4841 * cmp = "00..0<cmp>00..0"
4842 * new = "00..0<new>00..0"
4845 /* Patch cmp and new with val at correct position. */
4846 if (ac
.aligned
&& MEM_P (cmp
))
4848 cmpv
= force_reg (SImode
, val
);
4849 store_bit_field (cmpv
, GET_MODE_BITSIZE (mode
), 0, SImode
, cmp
);
4852 cmpv
= force_reg (SImode
, expand_simple_binop (SImode
, IOR
, cmp
, val
,
4853 NULL_RTX
, 1, OPTAB_DIRECT
));
4854 if (ac
.aligned
&& MEM_P (new_rtx
))
4856 newv
= force_reg (SImode
, val
);
4857 store_bit_field (newv
, GET_MODE_BITSIZE (mode
), 0, SImode
, new_rtx
);
4860 newv
= force_reg (SImode
, expand_simple_binop (SImode
, IOR
, new_rtx
, val
,
4861 NULL_RTX
, 1, OPTAB_DIRECT
));
4863 /* Jump to end if we're done (likely?). */
4864 s390_emit_jump (csend
, s390_emit_compare_and_swap (EQ
, res
, ac
.memsi
,
4867 /* Check for changes outside mode. */
4868 resv
= expand_simple_binop (SImode
, AND
, res
, ac
.modemaski
,
4869 NULL_RTX
, 1, OPTAB_DIRECT
);
4870 cc
= s390_emit_compare (NE
, resv
, val
);
4871 emit_move_insn (val
, resv
);
4872 /* Loop internal if so. */
4873 s390_emit_jump (csloop
, cc
);
4877 /* Return the correct part of the bitfield. */
4878 convert_move (target
, expand_simple_binop (SImode
, LSHIFTRT
, res
, ac
.shift
,
4879 NULL_RTX
, 1, OPTAB_DIRECT
), 1);
4882 /* Expand an atomic operation CODE of mode MODE. MEM is the memory location
4883 and VAL the value to play with. If AFTER is true then store the value
4884 MEM holds after the operation, if AFTER is false then store the value MEM
4885 holds before the operation. If TARGET is zero then discard that value, else
4886 store it to TARGET. */
4889 s390_expand_atomic (enum machine_mode mode
, enum rtx_code code
,
4890 rtx target
, rtx mem
, rtx val
, bool after
)
4892 struct alignment_context ac
;
4894 rtx new_rtx
= gen_reg_rtx (SImode
);
4895 rtx orig
= gen_reg_rtx (SImode
);
4896 rtx csloop
= gen_label_rtx ();
4898 gcc_assert (!target
|| register_operand (target
, VOIDmode
));
4899 gcc_assert (MEM_P (mem
));
4901 init_alignment_context (&ac
, mem
, mode
);
4903 /* Shift val to the correct bit positions.
4904 Preserve "icm", but prevent "ex icm". */
4905 if (!(ac
.aligned
&& code
== SET
&& MEM_P (val
)))
4906 val
= s390_expand_mask_and_shift (val
, mode
, ac
.shift
);
4908 /* Further preparation insns. */
4909 if (code
== PLUS
|| code
== MINUS
)
4910 emit_move_insn (orig
, val
);
4911 else if (code
== MULT
|| code
== AND
) /* val = "11..1<val>11..1" */
4912 val
= expand_simple_binop (SImode
, XOR
, val
, ac
.modemaski
,
4913 NULL_RTX
, 1, OPTAB_DIRECT
);
4915 /* Load full word. Subsequent loads are performed by CS. */
4916 cmp
= force_reg (SImode
, ac
.memsi
);
4918 /* Start CS loop. */
4919 emit_label (csloop
);
4920 emit_move_insn (new_rtx
, cmp
);
4922 /* Patch new with val at correct position. */
4927 val
= expand_simple_binop (SImode
, code
, new_rtx
, orig
,
4928 NULL_RTX
, 1, OPTAB_DIRECT
);
4929 val
= expand_simple_binop (SImode
, AND
, val
, ac
.modemask
,
4930 NULL_RTX
, 1, OPTAB_DIRECT
);
4933 if (ac
.aligned
&& MEM_P (val
))
4934 store_bit_field (new_rtx
, GET_MODE_BITSIZE (mode
), 0, SImode
, val
);
4937 new_rtx
= expand_simple_binop (SImode
, AND
, new_rtx
, ac
.modemaski
,
4938 NULL_RTX
, 1, OPTAB_DIRECT
);
4939 new_rtx
= expand_simple_binop (SImode
, IOR
, new_rtx
, val
,
4940 NULL_RTX
, 1, OPTAB_DIRECT
);
4946 new_rtx
= expand_simple_binop (SImode
, code
, new_rtx
, val
,
4947 NULL_RTX
, 1, OPTAB_DIRECT
);
4949 case MULT
: /* NAND */
4950 new_rtx
= expand_simple_binop (SImode
, AND
, new_rtx
, val
,
4951 NULL_RTX
, 1, OPTAB_DIRECT
);
4952 new_rtx
= expand_simple_binop (SImode
, XOR
, new_rtx
, ac
.modemask
,
4953 NULL_RTX
, 1, OPTAB_DIRECT
);
4959 s390_emit_jump (csloop
, s390_emit_compare_and_swap (NE
, cmp
,
4960 ac
.memsi
, cmp
, new_rtx
));
4962 /* Return the correct part of the bitfield. */
4964 convert_move (target
, expand_simple_binop (SImode
, LSHIFTRT
,
4965 after
? new_rtx
: cmp
, ac
.shift
,
4966 NULL_RTX
, 1, OPTAB_DIRECT
), 1);
4969 /* This is called from dwarf2out.c via TARGET_ASM_OUTPUT_DWARF_DTPREL.
4970 We need to emit DTP-relative relocations. */
4972 static void s390_output_dwarf_dtprel (FILE *, int, rtx
) ATTRIBUTE_UNUSED
;
4975 s390_output_dwarf_dtprel (FILE *file
, int size
, rtx x
)
4980 fputs ("\t.long\t", file
);
4983 fputs ("\t.quad\t", file
);
4988 output_addr_const (file
, x
);
4989 fputs ("@DTPOFF", file
);
4992 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
4993 /* Implement TARGET_MANGLE_TYPE. */
4996 s390_mangle_type (const_tree type
)
4998 if (TYPE_MAIN_VARIANT (type
) == long_double_type_node
4999 && TARGET_LONG_DOUBLE_128
)
5002 /* For all other types, use normal C++ mangling. */
5007 /* In the name of slightly smaller debug output, and to cater to
5008 general assembler lossage, recognize various UNSPEC sequences
5009 and turn them back into a direct symbol reference. */
5012 s390_delegitimize_address (rtx orig_x
)
5016 orig_x
= delegitimize_mem_from_attrs (orig_x
);
5019 /* Extract the symbol ref from:
5020 (plus:SI (reg:SI 12 %r12)
5021 (const:SI (unspec:SI [(symbol_ref/f:SI ("*.LC0"))]
5022 UNSPEC_GOTOFF/PLTOFF)))
5024 (plus:SI (reg:SI 12 %r12)
5025 (const:SI (plus:SI (unspec:SI [(symbol_ref:SI ("L"))]
5026 UNSPEC_GOTOFF/PLTOFF)
5027 (const_int 4 [0x4])))) */
5028 if (GET_CODE (x
) == PLUS
5029 && REG_P (XEXP (x
, 0))
5030 && REGNO (XEXP (x
, 0)) == PIC_OFFSET_TABLE_REGNUM
5031 && GET_CODE (XEXP (x
, 1)) == CONST
)
5033 /* The const operand. */
5034 y
= XEXP (XEXP (x
, 1), 0);
5036 if (GET_CODE (y
) == PLUS
5037 && GET_CODE (XEXP (y
, 1)) == CONST_INT
)
5040 if (GET_CODE (y
) == UNSPEC
5041 && (XINT (y
, 1) == UNSPEC_GOTOFF
5042 || XINT (y
, 1) == UNSPEC_PLTOFF
))
5043 return XVECEXP (y
, 0, 0);
5046 if (GET_CODE (x
) != MEM
)
5050 if (GET_CODE (x
) == PLUS
5051 && GET_CODE (XEXP (x
, 1)) == CONST
5052 && GET_CODE (XEXP (x
, 0)) == REG
5053 && REGNO (XEXP (x
, 0)) == PIC_OFFSET_TABLE_REGNUM
)
5055 y
= XEXP (XEXP (x
, 1), 0);
5056 if (GET_CODE (y
) == UNSPEC
5057 && XINT (y
, 1) == UNSPEC_GOT
)
5058 y
= XVECEXP (y
, 0, 0);
5062 else if (GET_CODE (x
) == CONST
)
5064 /* Extract the symbol ref from:
5065 (mem:QI (const:DI (unspec:DI [(symbol_ref:DI ("foo"))]
5066 UNSPEC_PLT/GOTENT))) */
5069 if (GET_CODE (y
) == UNSPEC
5070 && (XINT (y
, 1) == UNSPEC_GOTENT
5071 || XINT (y
, 1) == UNSPEC_PLT
))
5072 y
= XVECEXP (y
, 0, 0);
5079 if (GET_MODE (orig_x
) != Pmode
)
5081 if (GET_MODE (orig_x
) == BLKmode
)
5083 y
= lowpart_subreg (GET_MODE (orig_x
), y
, Pmode
);
5090 /* Output operand OP to stdio stream FILE.
5091 OP is an address (register + offset) which is not used to address data;
5092 instead the rightmost bits are interpreted as the value. */
5095 print_shift_count_operand (FILE *file
, rtx op
)
5097 HOST_WIDE_INT offset
;
5100 /* Extract base register and offset. */
5101 if (!s390_decompose_shift_count (op
, &base
, &offset
))
5107 gcc_assert (GET_CODE (base
) == REG
);
5108 gcc_assert (REGNO (base
) < FIRST_PSEUDO_REGISTER
);
5109 gcc_assert (REGNO_REG_CLASS (REGNO (base
)) == ADDR_REGS
);
5112 /* Offsets are constricted to twelve bits. */
5113 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, offset
& ((1 << 12) - 1));
5115 fprintf (file
, "(%s)", reg_names
[REGNO (base
)]);
5118 /* See 'get_some_local_dynamic_name'. */
5121 get_some_local_dynamic_name_1 (rtx
*px
, void *data ATTRIBUTE_UNUSED
)
5125 if (GET_CODE (x
) == SYMBOL_REF
&& CONSTANT_POOL_ADDRESS_P (x
))
5127 x
= get_pool_constant (x
);
5128 return for_each_rtx (&x
, get_some_local_dynamic_name_1
, 0);
5131 if (GET_CODE (x
) == SYMBOL_REF
5132 && tls_symbolic_operand (x
) == TLS_MODEL_LOCAL_DYNAMIC
)
5134 cfun
->machine
->some_ld_name
= XSTR (x
, 0);
5141 /* Locate some local-dynamic symbol still in use by this function
5142 so that we can print its name in local-dynamic base patterns. */
5145 get_some_local_dynamic_name (void)
5149 if (cfun
->machine
->some_ld_name
)
5150 return cfun
->machine
->some_ld_name
;
5152 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
5154 && for_each_rtx (&PATTERN (insn
), get_some_local_dynamic_name_1
, 0))
5155 return cfun
->machine
->some_ld_name
;
5160 /* Output machine-dependent UNSPECs occurring in address constant X
5161 in assembler syntax to stdio stream FILE. Returns true if the
5162 constant X could be recognized, false otherwise. */
5165 s390_output_addr_const_extra (FILE *file
, rtx x
)
5167 if (GET_CODE (x
) == UNSPEC
&& XVECLEN (x
, 0) == 1)
5168 switch (XINT (x
, 1))
5171 output_addr_const (file
, XVECEXP (x
, 0, 0));
5172 fprintf (file
, "@GOTENT");
5175 output_addr_const (file
, XVECEXP (x
, 0, 0));
5176 fprintf (file
, "@GOT");
5179 output_addr_const (file
, XVECEXP (x
, 0, 0));
5180 fprintf (file
, "@GOTOFF");
5183 output_addr_const (file
, XVECEXP (x
, 0, 0));
5184 fprintf (file
, "@PLT");
5187 output_addr_const (file
, XVECEXP (x
, 0, 0));
5188 fprintf (file
, "@PLTOFF");
5191 output_addr_const (file
, XVECEXP (x
, 0, 0));
5192 fprintf (file
, "@TLSGD");
5195 assemble_name (file
, get_some_local_dynamic_name ());
5196 fprintf (file
, "@TLSLDM");
5199 output_addr_const (file
, XVECEXP (x
, 0, 0));
5200 fprintf (file
, "@DTPOFF");
5203 output_addr_const (file
, XVECEXP (x
, 0, 0));
5204 fprintf (file
, "@NTPOFF");
5206 case UNSPEC_GOTNTPOFF
:
5207 output_addr_const (file
, XVECEXP (x
, 0, 0));
5208 fprintf (file
, "@GOTNTPOFF");
5210 case UNSPEC_INDNTPOFF
:
5211 output_addr_const (file
, XVECEXP (x
, 0, 0));
5212 fprintf (file
, "@INDNTPOFF");
5216 if (GET_CODE (x
) == UNSPEC
&& XVECLEN (x
, 0) == 2)
5217 switch (XINT (x
, 1))
5219 case UNSPEC_POOL_OFFSET
:
5220 x
= gen_rtx_MINUS (GET_MODE (x
), XVECEXP (x
, 0, 0), XVECEXP (x
, 0, 1));
5221 output_addr_const (file
, x
);
5227 /* Output address operand ADDR in assembler syntax to
5228 stdio stream FILE. */
5231 print_operand_address (FILE *file
, rtx addr
)
5233 struct s390_address ad
;
5235 if (s390_symref_operand_p (addr
, NULL
, NULL
))
5239 output_operand_lossage ("symbolic memory references are "
5240 "only supported on z10 or later");
5243 output_addr_const (file
, addr
);
5247 if (!s390_decompose_address (addr
, &ad
)
5248 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5249 || (ad
.indx
&& !REGNO_OK_FOR_INDEX_P (REGNO (ad
.indx
))))
5250 output_operand_lossage ("cannot decompose address");
5253 output_addr_const (file
, ad
.disp
);
5255 fprintf (file
, "0");
5257 if (ad
.base
&& ad
.indx
)
5258 fprintf (file
, "(%s,%s)", reg_names
[REGNO (ad
.indx
)],
5259 reg_names
[REGNO (ad
.base
)]);
5261 fprintf (file
, "(%s)", reg_names
[REGNO (ad
.base
)]);
5264 /* Output operand X in assembler syntax to stdio stream FILE.
5265 CODE specified the format flag. The following format flags
5268 'C': print opcode suffix for branch condition.
5269 'D': print opcode suffix for inverse branch condition.
5270 'E': print opcode suffix for branch on index instruction.
5271 'J': print tls_load/tls_gdcall/tls_ldcall suffix
5272 'G': print the size of the operand in bytes.
5273 'O': print only the displacement of a memory reference.
5274 'R': print only the base register of a memory reference.
5275 'S': print S-type memory reference (base+displacement).
5276 'N': print the second word of a DImode operand.
5277 'M': print the second word of a TImode operand.
5278 'Y': print shift count operand.
5280 'b': print integer X as if it's an unsigned byte.
5281 'c': print integer X as if it's an signed byte.
5282 'x': print integer X as if it's an unsigned halfword.
5283 'h': print integer X as if it's a signed halfword.
5284 'i': print the first nonzero HImode part of X.
5285 'j': print the first HImode part unequal to -1 of X.
5286 'k': print the first nonzero SImode part of X.
5287 'm': print the first SImode part unequal to -1 of X.
5288 'o': print integer X as if it's an unsigned 32bit word. */
5291 print_operand (FILE *file
, rtx x
, int code
)
5296 fprintf (file
, s390_branch_condition_mnemonic (x
, FALSE
));
5300 fprintf (file
, s390_branch_condition_mnemonic (x
, TRUE
));
5304 if (GET_CODE (x
) == LE
)
5305 fprintf (file
, "l");
5306 else if (GET_CODE (x
) == GT
)
5307 fprintf (file
, "h");
5309 output_operand_lossage ("invalid comparison operator "
5310 "for 'E' output modifier");
5314 if (GET_CODE (x
) == SYMBOL_REF
)
5316 fprintf (file
, "%s", ":tls_load:");
5317 output_addr_const (file
, x
);
5319 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLSGD
)
5321 fprintf (file
, "%s", ":tls_gdcall:");
5322 output_addr_const (file
, XVECEXP (x
, 0, 0));
5324 else if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_TLSLDM
)
5326 fprintf (file
, "%s", ":tls_ldcall:");
5327 assemble_name (file
, get_some_local_dynamic_name ());
5330 output_operand_lossage ("invalid reference for 'J' output modifier");
5334 fprintf (file
, "%u", GET_MODE_SIZE (GET_MODE (x
)));
5339 struct s390_address ad
;
5344 output_operand_lossage ("memory reference expected for "
5345 "'O' output modifier");
5349 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5352 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5355 output_operand_lossage ("invalid address for 'O' output modifier");
5360 output_addr_const (file
, ad
.disp
);
5362 fprintf (file
, "0");
5368 struct s390_address ad
;
5373 output_operand_lossage ("memory reference expected for "
5374 "'R' output modifier");
5378 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5381 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5384 output_operand_lossage ("invalid address for 'R' output modifier");
5389 fprintf (file
, "%s", reg_names
[REGNO (ad
.base
)]);
5391 fprintf (file
, "0");
5397 struct s390_address ad
;
5402 output_operand_lossage ("memory reference expected for "
5403 "'S' output modifier");
5406 ret
= s390_decompose_address (XEXP (x
, 0), &ad
);
5409 || (ad
.base
&& !REGNO_OK_FOR_BASE_P (REGNO (ad
.base
)))
5412 output_operand_lossage ("invalid address for 'S' output modifier");
5417 output_addr_const (file
, ad
.disp
);
5419 fprintf (file
, "0");
5422 fprintf (file
, "(%s)", reg_names
[REGNO (ad
.base
)]);
5427 if (GET_CODE (x
) == REG
)
5428 x
= gen_rtx_REG (GET_MODE (x
), REGNO (x
) + 1);
5429 else if (GET_CODE (x
) == MEM
)
5430 x
= change_address (x
, VOIDmode
, plus_constant (XEXP (x
, 0), 4));
5432 output_operand_lossage ("register or memory expression expected "
5433 "for 'N' output modifier");
5437 if (GET_CODE (x
) == REG
)
5438 x
= gen_rtx_REG (GET_MODE (x
), REGNO (x
) + 1);
5439 else if (GET_CODE (x
) == MEM
)
5440 x
= change_address (x
, VOIDmode
, plus_constant (XEXP (x
, 0), 8));
5442 output_operand_lossage ("register or memory expression expected "
5443 "for 'M' output modifier");
5447 print_shift_count_operand (file
, x
);
5451 switch (GET_CODE (x
))
5454 fprintf (file
, "%s", reg_names
[REGNO (x
)]);
5458 output_address (XEXP (x
, 0));
5465 output_addr_const (file
, x
);
5470 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xff);
5471 else if (code
== 'c')
5472 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ((INTVAL (x
) & 0xff) ^ 0x80) - 0x80);
5473 else if (code
== 'x')
5474 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xffff);
5475 else if (code
== 'h')
5476 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ((INTVAL (x
) & 0xffff) ^ 0x8000) - 0x8000);
5477 else if (code
== 'i')
5478 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5479 s390_extract_part (x
, HImode
, 0));
5480 else if (code
== 'j')
5481 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5482 s390_extract_part (x
, HImode
, -1));
5483 else if (code
== 'k')
5484 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5485 s390_extract_part (x
, SImode
, 0));
5486 else if (code
== 'm')
5487 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5488 s390_extract_part (x
, SImode
, -1));
5489 else if (code
== 'o')
5490 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
) & 0xffffffff);
5492 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, INTVAL (x
));
5496 gcc_assert (GET_MODE (x
) == VOIDmode
);
5498 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, CONST_DOUBLE_LOW (x
) & 0xff);
5499 else if (code
== 'x')
5500 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, CONST_DOUBLE_LOW (x
) & 0xffff);
5501 else if (code
== 'h')
5502 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
,
5503 ((CONST_DOUBLE_LOW (x
) & 0xffff) ^ 0x8000) - 0x8000);
5507 output_operand_lossage ("invalid constant - try using "
5508 "an output modifier");
5510 output_operand_lossage ("invalid constant for output modifier '%c'",
5517 output_operand_lossage ("invalid expression - try using "
5518 "an output modifier");
5520 output_operand_lossage ("invalid expression for output "
5521 "modifier '%c'", code
);
5526 /* Target hook for assembling integer objects. We need to define it
5527 here to work a round a bug in some versions of GAS, which couldn't
5528 handle values smaller than INT_MIN when printed in decimal. */
5531 s390_assemble_integer (rtx x
, unsigned int size
, int aligned_p
)
5533 if (size
== 8 && aligned_p
5534 && GET_CODE (x
) == CONST_INT
&& INTVAL (x
) < INT_MIN
)
5536 fprintf (asm_out_file
, "\t.quad\t" HOST_WIDE_INT_PRINT_HEX
"\n",
5540 return default_assemble_integer (x
, size
, aligned_p
);
5543 /* Returns true if register REGNO is used for forming
5544 a memory address in expression X. */
5547 reg_used_in_mem_p (int regno
, rtx x
)
5549 enum rtx_code code
= GET_CODE (x
);
5555 if (refers_to_regno_p (regno
, regno
+1,
5559 else if (code
== SET
5560 && GET_CODE (SET_DEST (x
)) == PC
)
5562 if (refers_to_regno_p (regno
, regno
+1,
5567 fmt
= GET_RTX_FORMAT (code
);
5568 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
5571 && reg_used_in_mem_p (regno
, XEXP (x
, i
)))
5574 else if (fmt
[i
] == 'E')
5575 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5576 if (reg_used_in_mem_p (regno
, XVECEXP (x
, i
, j
)))
5582 /* Returns true if expression DEP_RTX sets an address register
5583 used by instruction INSN to address memory. */
5586 addr_generation_dependency_p (rtx dep_rtx
, rtx insn
)
5590 if (GET_CODE (dep_rtx
) == INSN
)
5591 dep_rtx
= PATTERN (dep_rtx
);
5593 if (GET_CODE (dep_rtx
) == SET
)
5595 target
= SET_DEST (dep_rtx
);
5596 if (GET_CODE (target
) == STRICT_LOW_PART
)
5597 target
= XEXP (target
, 0);
5598 while (GET_CODE (target
) == SUBREG
)
5599 target
= SUBREG_REG (target
);
5601 if (GET_CODE (target
) == REG
)
5603 int regno
= REGNO (target
);
5605 if (s390_safe_attr_type (insn
) == TYPE_LA
)
5607 pat
= PATTERN (insn
);
5608 if (GET_CODE (pat
) == PARALLEL
)
5610 gcc_assert (XVECLEN (pat
, 0) == 2);
5611 pat
= XVECEXP (pat
, 0, 0);
5613 gcc_assert (GET_CODE (pat
) == SET
);
5614 return refers_to_regno_p (regno
, regno
+1, SET_SRC (pat
), 0);
5616 else if (get_attr_atype (insn
) == ATYPE_AGEN
)
5617 return reg_used_in_mem_p (regno
, PATTERN (insn
));
5623 /* Return 1, if dep_insn sets register used in insn in the agen unit. */
5626 s390_agen_dep_p (rtx dep_insn
, rtx insn
)
5628 rtx dep_rtx
= PATTERN (dep_insn
);
5631 if (GET_CODE (dep_rtx
) == SET
5632 && addr_generation_dependency_p (dep_rtx
, insn
))
5634 else if (GET_CODE (dep_rtx
) == PARALLEL
)
5636 for (i
= 0; i
< XVECLEN (dep_rtx
, 0); i
++)
5638 if (addr_generation_dependency_p (XVECEXP (dep_rtx
, 0, i
), insn
))
5646 /* A C statement (sans semicolon) to update the integer scheduling priority
5647 INSN_PRIORITY (INSN). Increase the priority to execute the INSN earlier,
5648 reduce the priority to execute INSN later. Do not define this macro if
5649 you do not need to adjust the scheduling priorities of insns.
5651 A STD instruction should be scheduled earlier,
5652 in order to use the bypass. */
5654 s390_adjust_priority (rtx insn ATTRIBUTE_UNUSED
, int priority
)
5656 if (! INSN_P (insn
))
5659 if (s390_tune
!= PROCESSOR_2084_Z990
5660 && s390_tune
!= PROCESSOR_2094_Z9_109
5661 && s390_tune
!= PROCESSOR_2097_Z10
5662 && s390_tune
!= PROCESSOR_2817_Z196
)
5665 switch (s390_safe_attr_type (insn
))
5669 priority
= priority
<< 3;
5673 priority
= priority
<< 1;
5682 /* The number of instructions that can be issued per cycle. */
5685 s390_issue_rate (void)
5689 case PROCESSOR_2084_Z990
:
5690 case PROCESSOR_2094_Z9_109
:
5691 case PROCESSOR_2817_Z196
:
5693 case PROCESSOR_2097_Z10
:
5701 s390_first_cycle_multipass_dfa_lookahead (void)
5706 /* Annotate every literal pool reference in X by an UNSPEC_LTREF expression.
5707 Fix up MEMs as required. */
5710 annotate_constant_pool_refs (rtx
*x
)
5715 gcc_assert (GET_CODE (*x
) != SYMBOL_REF
5716 || !CONSTANT_POOL_ADDRESS_P (*x
));
5718 /* Literal pool references can only occur inside a MEM ... */
5719 if (GET_CODE (*x
) == MEM
)
5721 rtx memref
= XEXP (*x
, 0);
5723 if (GET_CODE (memref
) == SYMBOL_REF
5724 && CONSTANT_POOL_ADDRESS_P (memref
))
5726 rtx base
= cfun
->machine
->base_reg
;
5727 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, memref
, base
),
5730 *x
= replace_equiv_address (*x
, addr
);
5734 if (GET_CODE (memref
) == CONST
5735 && GET_CODE (XEXP (memref
, 0)) == PLUS
5736 && GET_CODE (XEXP (XEXP (memref
, 0), 1)) == CONST_INT
5737 && GET_CODE (XEXP (XEXP (memref
, 0), 0)) == SYMBOL_REF
5738 && CONSTANT_POOL_ADDRESS_P (XEXP (XEXP (memref
, 0), 0)))
5740 HOST_WIDE_INT off
= INTVAL (XEXP (XEXP (memref
, 0), 1));
5741 rtx sym
= XEXP (XEXP (memref
, 0), 0);
5742 rtx base
= cfun
->machine
->base_reg
;
5743 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, sym
, base
),
5746 *x
= replace_equiv_address (*x
, plus_constant (addr
, off
));
5751 /* ... or a load-address type pattern. */
5752 if (GET_CODE (*x
) == SET
)
5754 rtx addrref
= SET_SRC (*x
);
5756 if (GET_CODE (addrref
) == SYMBOL_REF
5757 && CONSTANT_POOL_ADDRESS_P (addrref
))
5759 rtx base
= cfun
->machine
->base_reg
;
5760 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, addrref
, base
),
5763 SET_SRC (*x
) = addr
;
5767 if (GET_CODE (addrref
) == CONST
5768 && GET_CODE (XEXP (addrref
, 0)) == PLUS
5769 && GET_CODE (XEXP (XEXP (addrref
, 0), 1)) == CONST_INT
5770 && GET_CODE (XEXP (XEXP (addrref
, 0), 0)) == SYMBOL_REF
5771 && CONSTANT_POOL_ADDRESS_P (XEXP (XEXP (addrref
, 0), 0)))
5773 HOST_WIDE_INT off
= INTVAL (XEXP (XEXP (addrref
, 0), 1));
5774 rtx sym
= XEXP (XEXP (addrref
, 0), 0);
5775 rtx base
= cfun
->machine
->base_reg
;
5776 rtx addr
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, sym
, base
),
5779 SET_SRC (*x
) = plus_constant (addr
, off
);
5784 /* Annotate LTREL_BASE as well. */
5785 if (GET_CODE (*x
) == UNSPEC
5786 && XINT (*x
, 1) == UNSPEC_LTREL_BASE
)
5788 rtx base
= cfun
->machine
->base_reg
;
5789 *x
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, XVECEXP (*x
, 0, 0), base
),
5794 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5795 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5799 annotate_constant_pool_refs (&XEXP (*x
, i
));
5801 else if (fmt
[i
] == 'E')
5803 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5804 annotate_constant_pool_refs (&XVECEXP (*x
, i
, j
));
5809 /* Split all branches that exceed the maximum distance.
5810 Returns true if this created a new literal pool entry. */
5813 s390_split_branches (void)
5815 rtx temp_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
5816 int new_literal
= 0, ret
;
5817 rtx insn
, pat
, tmp
, target
;
5820 /* We need correct insn addresses. */
5822 shorten_branches (get_insns ());
5824 /* Find all branches that exceed 64KB, and split them. */
5826 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
5828 if (GET_CODE (insn
) != JUMP_INSN
)
5831 pat
= PATTERN (insn
);
5832 if (GET_CODE (pat
) == PARALLEL
&& XVECLEN (pat
, 0) > 2)
5833 pat
= XVECEXP (pat
, 0, 0);
5834 if (GET_CODE (pat
) != SET
|| SET_DEST (pat
) != pc_rtx
)
5837 if (GET_CODE (SET_SRC (pat
)) == LABEL_REF
)
5839 label
= &SET_SRC (pat
);
5841 else if (GET_CODE (SET_SRC (pat
)) == IF_THEN_ELSE
)
5843 if (GET_CODE (XEXP (SET_SRC (pat
), 1)) == LABEL_REF
)
5844 label
= &XEXP (SET_SRC (pat
), 1);
5845 else if (GET_CODE (XEXP (SET_SRC (pat
), 2)) == LABEL_REF
)
5846 label
= &XEXP (SET_SRC (pat
), 2);
5853 if (get_attr_length (insn
) <= 4)
5856 /* We are going to use the return register as scratch register,
5857 make sure it will be saved/restored by the prologue/epilogue. */
5858 cfun_frame_layout
.save_return_addr_p
= 1;
5863 tmp
= force_const_mem (Pmode
, *label
);
5864 tmp
= emit_insn_before (gen_rtx_SET (Pmode
, temp_reg
, tmp
), insn
);
5865 INSN_ADDRESSES_NEW (tmp
, -1);
5866 annotate_constant_pool_refs (&PATTERN (tmp
));
5873 target
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, *label
),
5874 UNSPEC_LTREL_OFFSET
);
5875 target
= gen_rtx_CONST (Pmode
, target
);
5876 target
= force_const_mem (Pmode
, target
);
5877 tmp
= emit_insn_before (gen_rtx_SET (Pmode
, temp_reg
, target
), insn
);
5878 INSN_ADDRESSES_NEW (tmp
, -1);
5879 annotate_constant_pool_refs (&PATTERN (tmp
));
5881 target
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (2, XEXP (target
, 0),
5882 cfun
->machine
->base_reg
),
5884 target
= gen_rtx_PLUS (Pmode
, temp_reg
, target
);
5887 ret
= validate_change (insn
, label
, target
, 0);
5895 /* Find an annotated literal pool symbol referenced in RTX X,
5896 and store it at REF. Will abort if X contains references to
5897 more than one such pool symbol; multiple references to the same
5898 symbol are allowed, however.
5900 The rtx pointed to by REF must be initialized to NULL_RTX
5901 by the caller before calling this routine. */
5904 find_constant_pool_ref (rtx x
, rtx
*ref
)
5909 /* Ignore LTREL_BASE references. */
5910 if (GET_CODE (x
) == UNSPEC
5911 && XINT (x
, 1) == UNSPEC_LTREL_BASE
)
5913 /* Likewise POOL_ENTRY insns. */
5914 if (GET_CODE (x
) == UNSPEC_VOLATILE
5915 && XINT (x
, 1) == UNSPECV_POOL_ENTRY
)
5918 gcc_assert (GET_CODE (x
) != SYMBOL_REF
5919 || !CONSTANT_POOL_ADDRESS_P (x
));
5921 if (GET_CODE (x
) == UNSPEC
&& XINT (x
, 1) == UNSPEC_LTREF
)
5923 rtx sym
= XVECEXP (x
, 0, 0);
5924 gcc_assert (GET_CODE (sym
) == SYMBOL_REF
5925 && CONSTANT_POOL_ADDRESS_P (sym
));
5927 if (*ref
== NULL_RTX
)
5930 gcc_assert (*ref
== sym
);
5935 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
5936 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
5940 find_constant_pool_ref (XEXP (x
, i
), ref
);
5942 else if (fmt
[i
] == 'E')
5944 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5945 find_constant_pool_ref (XVECEXP (x
, i
, j
), ref
);
5950 /* Replace every reference to the annotated literal pool
5951 symbol REF in X by its base plus OFFSET. */
5954 replace_constant_pool_ref (rtx
*x
, rtx ref
, rtx offset
)
5959 gcc_assert (*x
!= ref
);
5961 if (GET_CODE (*x
) == UNSPEC
5962 && XINT (*x
, 1) == UNSPEC_LTREF
5963 && XVECEXP (*x
, 0, 0) == ref
)
5965 *x
= gen_rtx_PLUS (Pmode
, XVECEXP (*x
, 0, 1), offset
);
5969 if (GET_CODE (*x
) == PLUS
5970 && GET_CODE (XEXP (*x
, 1)) == CONST_INT
5971 && GET_CODE (XEXP (*x
, 0)) == UNSPEC
5972 && XINT (XEXP (*x
, 0), 1) == UNSPEC_LTREF
5973 && XVECEXP (XEXP (*x
, 0), 0, 0) == ref
)
5975 rtx addr
= gen_rtx_PLUS (Pmode
, XVECEXP (XEXP (*x
, 0), 0, 1), offset
);
5976 *x
= plus_constant (addr
, INTVAL (XEXP (*x
, 1)));
5980 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
5981 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
5985 replace_constant_pool_ref (&XEXP (*x
, i
), ref
, offset
);
5987 else if (fmt
[i
] == 'E')
5989 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
5990 replace_constant_pool_ref (&XVECEXP (*x
, i
, j
), ref
, offset
);
5995 /* Check whether X contains an UNSPEC_LTREL_BASE.
5996 Return its constant pool symbol if found, NULL_RTX otherwise. */
5999 find_ltrel_base (rtx x
)
6004 if (GET_CODE (x
) == UNSPEC
6005 && XINT (x
, 1) == UNSPEC_LTREL_BASE
)
6006 return XVECEXP (x
, 0, 0);
6008 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
6009 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
6013 rtx fnd
= find_ltrel_base (XEXP (x
, i
));
6017 else if (fmt
[i
] == 'E')
6019 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
6021 rtx fnd
= find_ltrel_base (XVECEXP (x
, i
, j
));
6031 /* Replace any occurrence of UNSPEC_LTREL_BASE in X with its base. */
6034 replace_ltrel_base (rtx
*x
)
6039 if (GET_CODE (*x
) == UNSPEC
6040 && XINT (*x
, 1) == UNSPEC_LTREL_BASE
)
6042 *x
= XVECEXP (*x
, 0, 1);
6046 fmt
= GET_RTX_FORMAT (GET_CODE (*x
));
6047 for (i
= GET_RTX_LENGTH (GET_CODE (*x
)) - 1; i
>= 0; i
--)
6051 replace_ltrel_base (&XEXP (*x
, i
));
6053 else if (fmt
[i
] == 'E')
6055 for (j
= 0; j
< XVECLEN (*x
, i
); j
++)
6056 replace_ltrel_base (&XVECEXP (*x
, i
, j
));
6062 /* We keep a list of constants which we have to add to internal
6063 constant tables in the middle of large functions. */
6065 #define NR_C_MODES 11
6066 enum machine_mode constant_modes
[NR_C_MODES
] =
6068 TFmode
, TImode
, TDmode
,
6069 DFmode
, DImode
, DDmode
,
6070 SFmode
, SImode
, SDmode
,
6077 struct constant
*next
;
6082 struct constant_pool
6084 struct constant_pool
*next
;
6088 rtx emit_pool_after
;
6090 struct constant
*constants
[NR_C_MODES
];
6091 struct constant
*execute
;
6096 /* Allocate new constant_pool structure. */
6098 static struct constant_pool
*
6099 s390_alloc_pool (void)
6101 struct constant_pool
*pool
;
6104 pool
= (struct constant_pool
*) xmalloc (sizeof *pool
);
6106 for (i
= 0; i
< NR_C_MODES
; i
++)
6107 pool
->constants
[i
] = NULL
;
6109 pool
->execute
= NULL
;
6110 pool
->label
= gen_label_rtx ();
6111 pool
->first_insn
= NULL_RTX
;
6112 pool
->pool_insn
= NULL_RTX
;
6113 pool
->insns
= BITMAP_ALLOC (NULL
);
6115 pool
->emit_pool_after
= NULL_RTX
;
6120 /* Create new constant pool covering instructions starting at INSN
6121 and chain it to the end of POOL_LIST. */
6123 static struct constant_pool
*
6124 s390_start_pool (struct constant_pool
**pool_list
, rtx insn
)
6126 struct constant_pool
*pool
, **prev
;
6128 pool
= s390_alloc_pool ();
6129 pool
->first_insn
= insn
;
6131 for (prev
= pool_list
; *prev
; prev
= &(*prev
)->next
)
6138 /* End range of instructions covered by POOL at INSN and emit
6139 placeholder insn representing the pool. */
6142 s390_end_pool (struct constant_pool
*pool
, rtx insn
)
6144 rtx pool_size
= GEN_INT (pool
->size
+ 8 /* alignment slop */);
6147 insn
= get_last_insn ();
6149 pool
->pool_insn
= emit_insn_after (gen_pool (pool_size
), insn
);
6150 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6153 /* Add INSN to the list of insns covered by POOL. */
6156 s390_add_pool_insn (struct constant_pool
*pool
, rtx insn
)
6158 bitmap_set_bit (pool
->insns
, INSN_UID (insn
));
6161 /* Return pool out of POOL_LIST that covers INSN. */
6163 static struct constant_pool
*
6164 s390_find_pool (struct constant_pool
*pool_list
, rtx insn
)
6166 struct constant_pool
*pool
;
6168 for (pool
= pool_list
; pool
; pool
= pool
->next
)
6169 if (bitmap_bit_p (pool
->insns
, INSN_UID (insn
)))
6175 /* Add constant VAL of mode MODE to the constant pool POOL. */
6178 s390_add_constant (struct constant_pool
*pool
, rtx val
, enum machine_mode mode
)
6183 for (i
= 0; i
< NR_C_MODES
; i
++)
6184 if (constant_modes
[i
] == mode
)
6186 gcc_assert (i
!= NR_C_MODES
);
6188 for (c
= pool
->constants
[i
]; c
!= NULL
; c
= c
->next
)
6189 if (rtx_equal_p (val
, c
->value
))
6194 c
= (struct constant
*) xmalloc (sizeof *c
);
6196 c
->label
= gen_label_rtx ();
6197 c
->next
= pool
->constants
[i
];
6198 pool
->constants
[i
] = c
;
6199 pool
->size
+= GET_MODE_SIZE (mode
);
6203 /* Return an rtx that represents the offset of X from the start of
6207 s390_pool_offset (struct constant_pool
*pool
, rtx x
)
6211 label
= gen_rtx_LABEL_REF (GET_MODE (x
), pool
->label
);
6212 x
= gen_rtx_UNSPEC (GET_MODE (x
), gen_rtvec (2, x
, label
),
6213 UNSPEC_POOL_OFFSET
);
6214 return gen_rtx_CONST (GET_MODE (x
), x
);
6217 /* Find constant VAL of mode MODE in the constant pool POOL.
6218 Return an RTX describing the distance from the start of
6219 the pool to the location of the new constant. */
6222 s390_find_constant (struct constant_pool
*pool
, rtx val
,
6223 enum machine_mode mode
)
6228 for (i
= 0; i
< NR_C_MODES
; i
++)
6229 if (constant_modes
[i
] == mode
)
6231 gcc_assert (i
!= NR_C_MODES
);
6233 for (c
= pool
->constants
[i
]; c
!= NULL
; c
= c
->next
)
6234 if (rtx_equal_p (val
, c
->value
))
6239 return s390_pool_offset (pool
, gen_rtx_LABEL_REF (Pmode
, c
->label
));
6242 /* Check whether INSN is an execute. Return the label_ref to its
6243 execute target template if so, NULL_RTX otherwise. */
6246 s390_execute_label (rtx insn
)
6248 if (GET_CODE (insn
) == INSN
6249 && GET_CODE (PATTERN (insn
)) == PARALLEL
6250 && GET_CODE (XVECEXP (PATTERN (insn
), 0, 0)) == UNSPEC
6251 && XINT (XVECEXP (PATTERN (insn
), 0, 0), 1) == UNSPEC_EXECUTE
)
6252 return XVECEXP (XVECEXP (PATTERN (insn
), 0, 0), 0, 2);
6257 /* Add execute target for INSN to the constant pool POOL. */
6260 s390_add_execute (struct constant_pool
*pool
, rtx insn
)
6264 for (c
= pool
->execute
; c
!= NULL
; c
= c
->next
)
6265 if (INSN_UID (insn
) == INSN_UID (c
->value
))
6270 c
= (struct constant
*) xmalloc (sizeof *c
);
6272 c
->label
= gen_label_rtx ();
6273 c
->next
= pool
->execute
;
6279 /* Find execute target for INSN in the constant pool POOL.
6280 Return an RTX describing the distance from the start of
6281 the pool to the location of the execute target. */
6284 s390_find_execute (struct constant_pool
*pool
, rtx insn
)
6288 for (c
= pool
->execute
; c
!= NULL
; c
= c
->next
)
6289 if (INSN_UID (insn
) == INSN_UID (c
->value
))
6294 return s390_pool_offset (pool
, gen_rtx_LABEL_REF (Pmode
, c
->label
));
6297 /* For an execute INSN, extract the execute target template. */
6300 s390_execute_target (rtx insn
)
6302 rtx pattern
= PATTERN (insn
);
6303 gcc_assert (s390_execute_label (insn
));
6305 if (XVECLEN (pattern
, 0) == 2)
6307 pattern
= copy_rtx (XVECEXP (pattern
, 0, 1));
6311 rtvec vec
= rtvec_alloc (XVECLEN (pattern
, 0) - 1);
6314 for (i
= 0; i
< XVECLEN (pattern
, 0) - 1; i
++)
6315 RTVEC_ELT (vec
, i
) = copy_rtx (XVECEXP (pattern
, 0, i
+ 1));
6317 pattern
= gen_rtx_PARALLEL (VOIDmode
, vec
);
6323 /* Indicate that INSN cannot be duplicated. This is the case for
6324 execute insns that carry a unique label. */
6327 s390_cannot_copy_insn_p (rtx insn
)
6329 rtx label
= s390_execute_label (insn
);
6330 return label
&& label
!= const0_rtx
;
6333 /* Dump out the constants in POOL. If REMOTE_LABEL is true,
6334 do not emit the pool base label. */
6337 s390_dump_pool (struct constant_pool
*pool
, bool remote_label
)
6340 rtx insn
= pool
->pool_insn
;
6343 /* Switch to rodata section. */
6344 if (TARGET_CPU_ZARCH
)
6346 insn
= emit_insn_after (gen_pool_section_start (), insn
);
6347 INSN_ADDRESSES_NEW (insn
, -1);
6350 /* Ensure minimum pool alignment. */
6351 if (TARGET_CPU_ZARCH
)
6352 insn
= emit_insn_after (gen_pool_align (GEN_INT (8)), insn
);
6354 insn
= emit_insn_after (gen_pool_align (GEN_INT (4)), insn
);
6355 INSN_ADDRESSES_NEW (insn
, -1);
6357 /* Emit pool base label. */
6360 insn
= emit_label_after (pool
->label
, insn
);
6361 INSN_ADDRESSES_NEW (insn
, -1);
6364 /* Dump constants in descending alignment requirement order,
6365 ensuring proper alignment for every constant. */
6366 for (i
= 0; i
< NR_C_MODES
; i
++)
6367 for (c
= pool
->constants
[i
]; c
; c
= c
->next
)
6369 /* Convert UNSPEC_LTREL_OFFSET unspecs to pool-relative references. */
6370 rtx value
= copy_rtx (c
->value
);
6371 if (GET_CODE (value
) == CONST
6372 && GET_CODE (XEXP (value
, 0)) == UNSPEC
6373 && XINT (XEXP (value
, 0), 1) == UNSPEC_LTREL_OFFSET
6374 && XVECLEN (XEXP (value
, 0), 0) == 1)
6375 value
= s390_pool_offset (pool
, XVECEXP (XEXP (value
, 0), 0, 0));
6377 insn
= emit_label_after (c
->label
, insn
);
6378 INSN_ADDRESSES_NEW (insn
, -1);
6380 value
= gen_rtx_UNSPEC_VOLATILE (constant_modes
[i
],
6381 gen_rtvec (1, value
),
6382 UNSPECV_POOL_ENTRY
);
6383 insn
= emit_insn_after (value
, insn
);
6384 INSN_ADDRESSES_NEW (insn
, -1);
6387 /* Ensure minimum alignment for instructions. */
6388 insn
= emit_insn_after (gen_pool_align (GEN_INT (2)), insn
);
6389 INSN_ADDRESSES_NEW (insn
, -1);
6391 /* Output in-pool execute template insns. */
6392 for (c
= pool
->execute
; c
; c
= c
->next
)
6394 insn
= emit_label_after (c
->label
, insn
);
6395 INSN_ADDRESSES_NEW (insn
, -1);
6397 insn
= emit_insn_after (s390_execute_target (c
->value
), insn
);
6398 INSN_ADDRESSES_NEW (insn
, -1);
6401 /* Switch back to previous section. */
6402 if (TARGET_CPU_ZARCH
)
6404 insn
= emit_insn_after (gen_pool_section_end (), insn
);
6405 INSN_ADDRESSES_NEW (insn
, -1);
6408 insn
= emit_barrier_after (insn
);
6409 INSN_ADDRESSES_NEW (insn
, -1);
6411 /* Remove placeholder insn. */
6412 remove_insn (pool
->pool_insn
);
6415 /* Free all memory used by POOL. */
6418 s390_free_pool (struct constant_pool
*pool
)
6420 struct constant
*c
, *next
;
6423 for (i
= 0; i
< NR_C_MODES
; i
++)
6424 for (c
= pool
->constants
[i
]; c
; c
= next
)
6430 for (c
= pool
->execute
; c
; c
= next
)
6436 BITMAP_FREE (pool
->insns
);
6441 /* Collect main literal pool. Return NULL on overflow. */
6443 static struct constant_pool
*
6444 s390_mainpool_start (void)
6446 struct constant_pool
*pool
;
6449 pool
= s390_alloc_pool ();
6451 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6453 if (GET_CODE (insn
) == INSN
6454 && GET_CODE (PATTERN (insn
)) == SET
6455 && GET_CODE (SET_SRC (PATTERN (insn
))) == UNSPEC_VOLATILE
6456 && XINT (SET_SRC (PATTERN (insn
)), 1) == UNSPECV_MAIN_POOL
)
6458 gcc_assert (!pool
->pool_insn
);
6459 pool
->pool_insn
= insn
;
6462 if (!TARGET_CPU_ZARCH
&& s390_execute_label (insn
))
6464 s390_add_execute (pool
, insn
);
6466 else if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6468 rtx pool_ref
= NULL_RTX
;
6469 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6472 rtx constant
= get_pool_constant (pool_ref
);
6473 enum machine_mode mode
= get_pool_mode (pool_ref
);
6474 s390_add_constant (pool
, constant
, mode
);
6478 /* If hot/cold partitioning is enabled we have to make sure that
6479 the literal pool is emitted in the same section where the
6480 initialization of the literal pool base pointer takes place.
6481 emit_pool_after is only used in the non-overflow case on non
6482 Z cpus where we can emit the literal pool at the end of the
6483 function body within the text section. */
6485 && NOTE_KIND (insn
) == NOTE_INSN_SWITCH_TEXT_SECTIONS
6486 && !pool
->emit_pool_after
)
6487 pool
->emit_pool_after
= PREV_INSN (insn
);
6490 gcc_assert (pool
->pool_insn
|| pool
->size
== 0);
6492 if (pool
->size
>= 4096)
6494 /* We're going to chunkify the pool, so remove the main
6495 pool placeholder insn. */
6496 remove_insn (pool
->pool_insn
);
6498 s390_free_pool (pool
);
6502 /* If the functions ends with the section where the literal pool
6503 should be emitted set the marker to its end. */
6504 if (pool
&& !pool
->emit_pool_after
)
6505 pool
->emit_pool_after
= get_last_insn ();
6510 /* POOL holds the main literal pool as collected by s390_mainpool_start.
6511 Modify the current function to output the pool constants as well as
6512 the pool register setup instruction. */
6515 s390_mainpool_finish (struct constant_pool
*pool
)
6517 rtx base_reg
= cfun
->machine
->base_reg
;
6520 /* If the pool is empty, we're done. */
6521 if (pool
->size
== 0)
6523 /* We don't actually need a base register after all. */
6524 cfun
->machine
->base_reg
= NULL_RTX
;
6526 if (pool
->pool_insn
)
6527 remove_insn (pool
->pool_insn
);
6528 s390_free_pool (pool
);
6532 /* We need correct insn addresses. */
6533 shorten_branches (get_insns ());
6535 /* On zSeries, we use a LARL to load the pool register. The pool is
6536 located in the .rodata section, so we emit it after the function. */
6537 if (TARGET_CPU_ZARCH
)
6539 insn
= gen_main_base_64 (base_reg
, pool
->label
);
6540 insn
= emit_insn_after (insn
, pool
->pool_insn
);
6541 INSN_ADDRESSES_NEW (insn
, -1);
6542 remove_insn (pool
->pool_insn
);
6544 insn
= get_last_insn ();
6545 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6546 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6548 s390_dump_pool (pool
, 0);
6551 /* On S/390, if the total size of the function's code plus literal pool
6552 does not exceed 4096 bytes, we use BASR to set up a function base
6553 pointer, and emit the literal pool at the end of the function. */
6554 else if (INSN_ADDRESSES (INSN_UID (pool
->emit_pool_after
))
6555 + pool
->size
+ 8 /* alignment slop */ < 4096)
6557 insn
= gen_main_base_31_small (base_reg
, pool
->label
);
6558 insn
= emit_insn_after (insn
, pool
->pool_insn
);
6559 INSN_ADDRESSES_NEW (insn
, -1);
6560 remove_insn (pool
->pool_insn
);
6562 insn
= emit_label_after (pool
->label
, insn
);
6563 INSN_ADDRESSES_NEW (insn
, -1);
6565 /* emit_pool_after will be set by s390_mainpool_start to the
6566 last insn of the section where the literal pool should be
6568 insn
= pool
->emit_pool_after
;
6570 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6571 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6573 s390_dump_pool (pool
, 1);
6576 /* Otherwise, we emit an inline literal pool and use BASR to branch
6577 over it, setting up the pool register at the same time. */
6580 rtx pool_end
= gen_label_rtx ();
6582 insn
= gen_main_base_31_large (base_reg
, pool
->label
, pool_end
);
6583 insn
= emit_insn_after (insn
, pool
->pool_insn
);
6584 INSN_ADDRESSES_NEW (insn
, -1);
6585 remove_insn (pool
->pool_insn
);
6587 insn
= emit_label_after (pool
->label
, insn
);
6588 INSN_ADDRESSES_NEW (insn
, -1);
6590 pool
->pool_insn
= emit_insn_after (gen_pool (const0_rtx
), insn
);
6591 INSN_ADDRESSES_NEW (pool
->pool_insn
, -1);
6593 insn
= emit_label_after (pool_end
, pool
->pool_insn
);
6594 INSN_ADDRESSES_NEW (insn
, -1);
6596 s390_dump_pool (pool
, 1);
6600 /* Replace all literal pool references. */
6602 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6605 replace_ltrel_base (&PATTERN (insn
));
6607 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6609 rtx addr
, pool_ref
= NULL_RTX
;
6610 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6613 if (s390_execute_label (insn
))
6614 addr
= s390_find_execute (pool
, insn
);
6616 addr
= s390_find_constant (pool
, get_pool_constant (pool_ref
),
6617 get_pool_mode (pool_ref
));
6619 replace_constant_pool_ref (&PATTERN (insn
), pool_ref
, addr
);
6620 INSN_CODE (insn
) = -1;
6626 /* Free the pool. */
6627 s390_free_pool (pool
);
6630 /* POOL holds the main literal pool as collected by s390_mainpool_start.
6631 We have decided we cannot use this pool, so revert all changes
6632 to the current function that were done by s390_mainpool_start. */
6634 s390_mainpool_cancel (struct constant_pool
*pool
)
6636 /* We didn't actually change the instruction stream, so simply
6637 free the pool memory. */
6638 s390_free_pool (pool
);
6642 /* Chunkify the literal pool. */
6644 #define S390_POOL_CHUNK_MIN 0xc00
6645 #define S390_POOL_CHUNK_MAX 0xe00
6647 static struct constant_pool
*
6648 s390_chunkify_start (void)
6650 struct constant_pool
*curr_pool
= NULL
, *pool_list
= NULL
;
6653 rtx pending_ltrel
= NULL_RTX
;
6656 rtx (*gen_reload_base
) (rtx
, rtx
) =
6657 TARGET_CPU_ZARCH
? gen_reload_base_64
: gen_reload_base_31
;
6660 /* We need correct insn addresses. */
6662 shorten_branches (get_insns ());
6664 /* Scan all insns and move literals to pool chunks. */
6666 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6668 bool section_switch_p
= false;
6670 /* Check for pending LTREL_BASE. */
6673 rtx ltrel_base
= find_ltrel_base (PATTERN (insn
));
6676 gcc_assert (ltrel_base
== pending_ltrel
);
6677 pending_ltrel
= NULL_RTX
;
6681 if (!TARGET_CPU_ZARCH
&& s390_execute_label (insn
))
6684 curr_pool
= s390_start_pool (&pool_list
, insn
);
6686 s390_add_execute (curr_pool
, insn
);
6687 s390_add_pool_insn (curr_pool
, insn
);
6689 else if (GET_CODE (insn
) == INSN
|| CALL_P (insn
))
6691 rtx pool_ref
= NULL_RTX
;
6692 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6695 rtx constant
= get_pool_constant (pool_ref
);
6696 enum machine_mode mode
= get_pool_mode (pool_ref
);
6699 curr_pool
= s390_start_pool (&pool_list
, insn
);
6701 s390_add_constant (curr_pool
, constant
, mode
);
6702 s390_add_pool_insn (curr_pool
, insn
);
6704 /* Don't split the pool chunk between a LTREL_OFFSET load
6705 and the corresponding LTREL_BASE. */
6706 if (GET_CODE (constant
) == CONST
6707 && GET_CODE (XEXP (constant
, 0)) == UNSPEC
6708 && XINT (XEXP (constant
, 0), 1) == UNSPEC_LTREL_OFFSET
)
6710 gcc_assert (!pending_ltrel
);
6711 pending_ltrel
= pool_ref
;
6714 /* Make sure we do not split between a call and its
6715 corresponding CALL_ARG_LOCATION note. */
6718 rtx next
= NEXT_INSN (insn
);
6719 if (next
&& NOTE_P (next
)
6720 && NOTE_KIND (next
) == NOTE_INSN_CALL_ARG_LOCATION
)
6725 if (GET_CODE (insn
) == JUMP_INSN
|| GET_CODE (insn
) == CODE_LABEL
)
6728 s390_add_pool_insn (curr_pool
, insn
);
6729 /* An LTREL_BASE must follow within the same basic block. */
6730 gcc_assert (!pending_ltrel
);
6733 if (NOTE_P (insn
) && NOTE_KIND (insn
) == NOTE_INSN_SWITCH_TEXT_SECTIONS
)
6734 section_switch_p
= true;
6737 || INSN_ADDRESSES_SIZE () <= (size_t) INSN_UID (insn
)
6738 || INSN_ADDRESSES (INSN_UID (insn
)) == -1)
6741 if (TARGET_CPU_ZARCH
)
6743 if (curr_pool
->size
< S390_POOL_CHUNK_MAX
)
6746 s390_end_pool (curr_pool
, NULL_RTX
);
6751 int chunk_size
= INSN_ADDRESSES (INSN_UID (insn
))
6752 - INSN_ADDRESSES (INSN_UID (curr_pool
->first_insn
))
6755 /* We will later have to insert base register reload insns.
6756 Those will have an effect on code size, which we need to
6757 consider here. This calculation makes rather pessimistic
6758 worst-case assumptions. */
6759 if (GET_CODE (insn
) == CODE_LABEL
)
6762 if (chunk_size
< S390_POOL_CHUNK_MIN
6763 && curr_pool
->size
< S390_POOL_CHUNK_MIN
6764 && !section_switch_p
)
6767 /* Pool chunks can only be inserted after BARRIERs ... */
6768 if (GET_CODE (insn
) == BARRIER
)
6770 s390_end_pool (curr_pool
, insn
);
6775 /* ... so if we don't find one in time, create one. */
6776 else if (chunk_size
> S390_POOL_CHUNK_MAX
6777 || curr_pool
->size
> S390_POOL_CHUNK_MAX
6778 || section_switch_p
)
6780 rtx label
, jump
, barrier
;
6782 if (!section_switch_p
)
6784 /* We can insert the barrier only after a 'real' insn. */
6785 if (GET_CODE (insn
) != INSN
&& GET_CODE (insn
) != CALL_INSN
)
6787 if (get_attr_length (insn
) == 0)
6789 /* Don't separate LTREL_BASE from the corresponding
6790 LTREL_OFFSET load. */
6796 gcc_assert (!pending_ltrel
);
6798 /* The old pool has to end before the section switch
6799 note in order to make it part of the current
6801 insn
= PREV_INSN (insn
);
6804 label
= gen_label_rtx ();
6805 jump
= emit_jump_insn_after (gen_jump (label
), insn
);
6806 barrier
= emit_barrier_after (jump
);
6807 insn
= emit_label_after (label
, barrier
);
6808 JUMP_LABEL (jump
) = label
;
6809 LABEL_NUSES (label
) = 1;
6811 INSN_ADDRESSES_NEW (jump
, -1);
6812 INSN_ADDRESSES_NEW (barrier
, -1);
6813 INSN_ADDRESSES_NEW (insn
, -1);
6815 s390_end_pool (curr_pool
, barrier
);
6823 s390_end_pool (curr_pool
, NULL_RTX
);
6824 gcc_assert (!pending_ltrel
);
6826 /* Find all labels that are branched into
6827 from an insn belonging to a different chunk. */
6829 far_labels
= BITMAP_ALLOC (NULL
);
6831 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6833 /* Labels marked with LABEL_PRESERVE_P can be target
6834 of non-local jumps, so we have to mark them.
6835 The same holds for named labels.
6837 Don't do that, however, if it is the label before
6840 if (GET_CODE (insn
) == CODE_LABEL
6841 && (LABEL_PRESERVE_P (insn
) || LABEL_NAME (insn
)))
6843 rtx vec_insn
= next_real_insn (insn
);
6844 rtx vec_pat
= vec_insn
&& GET_CODE (vec_insn
) == JUMP_INSN
?
6845 PATTERN (vec_insn
) : NULL_RTX
;
6847 || !(GET_CODE (vec_pat
) == ADDR_VEC
6848 || GET_CODE (vec_pat
) == ADDR_DIFF_VEC
))
6849 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (insn
));
6852 /* If we have a direct jump (conditional or unconditional)
6853 or a casesi jump, check all potential targets. */
6854 else if (GET_CODE (insn
) == JUMP_INSN
)
6856 rtx pat
= PATTERN (insn
);
6857 if (GET_CODE (pat
) == PARALLEL
&& XVECLEN (pat
, 0) > 2)
6858 pat
= XVECEXP (pat
, 0, 0);
6860 if (GET_CODE (pat
) == SET
)
6862 rtx label
= JUMP_LABEL (insn
);
6865 if (s390_find_pool (pool_list
, label
)
6866 != s390_find_pool (pool_list
, insn
))
6867 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (label
));
6870 else if (GET_CODE (pat
) == PARALLEL
6871 && XVECLEN (pat
, 0) == 2
6872 && GET_CODE (XVECEXP (pat
, 0, 0)) == SET
6873 && GET_CODE (XVECEXP (pat
, 0, 1)) == USE
6874 && GET_CODE (XEXP (XVECEXP (pat
, 0, 1), 0)) == LABEL_REF
)
6876 /* Find the jump table used by this casesi jump. */
6877 rtx vec_label
= XEXP (XEXP (XVECEXP (pat
, 0, 1), 0), 0);
6878 rtx vec_insn
= next_real_insn (vec_label
);
6879 rtx vec_pat
= vec_insn
&& GET_CODE (vec_insn
) == JUMP_INSN
?
6880 PATTERN (vec_insn
) : NULL_RTX
;
6882 && (GET_CODE (vec_pat
) == ADDR_VEC
6883 || GET_CODE (vec_pat
) == ADDR_DIFF_VEC
))
6885 int i
, diff_p
= GET_CODE (vec_pat
) == ADDR_DIFF_VEC
;
6887 for (i
= 0; i
< XVECLEN (vec_pat
, diff_p
); i
++)
6889 rtx label
= XEXP (XVECEXP (vec_pat
, diff_p
, i
), 0);
6891 if (s390_find_pool (pool_list
, label
)
6892 != s390_find_pool (pool_list
, insn
))
6893 bitmap_set_bit (far_labels
, CODE_LABEL_NUMBER (label
));
6900 /* Insert base register reload insns before every pool. */
6902 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6904 rtx new_insn
= gen_reload_base (cfun
->machine
->base_reg
,
6906 rtx insn
= curr_pool
->first_insn
;
6907 INSN_ADDRESSES_NEW (emit_insn_before (new_insn
, insn
), -1);
6910 /* Insert base register reload insns at every far label. */
6912 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6913 if (GET_CODE (insn
) == CODE_LABEL
6914 && bitmap_bit_p (far_labels
, CODE_LABEL_NUMBER (insn
)))
6916 struct constant_pool
*pool
= s390_find_pool (pool_list
, insn
);
6919 rtx new_insn
= gen_reload_base (cfun
->machine
->base_reg
,
6921 INSN_ADDRESSES_NEW (emit_insn_after (new_insn
, insn
), -1);
6926 BITMAP_FREE (far_labels
);
6929 /* Recompute insn addresses. */
6931 init_insn_lengths ();
6932 shorten_branches (get_insns ());
6937 /* POOL_LIST is a chunk list as prepared by s390_chunkify_start.
6938 After we have decided to use this list, finish implementing
6939 all changes to the current function as required. */
6942 s390_chunkify_finish (struct constant_pool
*pool_list
)
6944 struct constant_pool
*curr_pool
= NULL
;
6948 /* Replace all literal pool references. */
6950 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
6953 replace_ltrel_base (&PATTERN (insn
));
6955 curr_pool
= s390_find_pool (pool_list
, insn
);
6959 if (GET_CODE (insn
) == INSN
|| GET_CODE (insn
) == CALL_INSN
)
6961 rtx addr
, pool_ref
= NULL_RTX
;
6962 find_constant_pool_ref (PATTERN (insn
), &pool_ref
);
6965 if (s390_execute_label (insn
))
6966 addr
= s390_find_execute (curr_pool
, insn
);
6968 addr
= s390_find_constant (curr_pool
,
6969 get_pool_constant (pool_ref
),
6970 get_pool_mode (pool_ref
));
6972 replace_constant_pool_ref (&PATTERN (insn
), pool_ref
, addr
);
6973 INSN_CODE (insn
) = -1;
6978 /* Dump out all literal pools. */
6980 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
6981 s390_dump_pool (curr_pool
, 0);
6983 /* Free pool list. */
6987 struct constant_pool
*next
= pool_list
->next
;
6988 s390_free_pool (pool_list
);
6993 /* POOL_LIST is a chunk list as prepared by s390_chunkify_start.
6994 We have decided we cannot use this list, so revert all changes
6995 to the current function that were done by s390_chunkify_start. */
6998 s390_chunkify_cancel (struct constant_pool
*pool_list
)
7000 struct constant_pool
*curr_pool
= NULL
;
7003 /* Remove all pool placeholder insns. */
7005 for (curr_pool
= pool_list
; curr_pool
; curr_pool
= curr_pool
->next
)
7007 /* Did we insert an extra barrier? Remove it. */
7008 rtx barrier
= PREV_INSN (curr_pool
->pool_insn
);
7009 rtx jump
= barrier
? PREV_INSN (barrier
) : NULL_RTX
;
7010 rtx label
= NEXT_INSN (curr_pool
->pool_insn
);
7012 if (jump
&& GET_CODE (jump
) == JUMP_INSN
7013 && barrier
&& GET_CODE (barrier
) == BARRIER
7014 && label
&& GET_CODE (label
) == CODE_LABEL
7015 && GET_CODE (PATTERN (jump
)) == SET
7016 && SET_DEST (PATTERN (jump
)) == pc_rtx
7017 && GET_CODE (SET_SRC (PATTERN (jump
))) == LABEL_REF
7018 && XEXP (SET_SRC (PATTERN (jump
)), 0) == label
)
7021 remove_insn (barrier
);
7022 remove_insn (label
);
7025 remove_insn (curr_pool
->pool_insn
);
7028 /* Remove all base register reload insns. */
7030 for (insn
= get_insns (); insn
; )
7032 rtx next_insn
= NEXT_INSN (insn
);
7034 if (GET_CODE (insn
) == INSN
7035 && GET_CODE (PATTERN (insn
)) == SET
7036 && GET_CODE (SET_SRC (PATTERN (insn
))) == UNSPEC
7037 && XINT (SET_SRC (PATTERN (insn
)), 1) == UNSPEC_RELOAD_BASE
)
7043 /* Free pool list. */
7047 struct constant_pool
*next
= pool_list
->next
;
7048 s390_free_pool (pool_list
);
7053 /* Output the constant pool entry EXP in mode MODE with alignment ALIGN. */
7056 s390_output_pool_entry (rtx exp
, enum machine_mode mode
, unsigned int align
)
7060 switch (GET_MODE_CLASS (mode
))
7063 case MODE_DECIMAL_FLOAT
:
7064 gcc_assert (GET_CODE (exp
) == CONST_DOUBLE
);
7066 REAL_VALUE_FROM_CONST_DOUBLE (r
, exp
);
7067 assemble_real (r
, mode
, align
);
7071 assemble_integer (exp
, GET_MODE_SIZE (mode
), align
, 1);
7072 mark_symbol_refs_as_used (exp
);
7081 /* Return an RTL expression representing the value of the return address
7082 for the frame COUNT steps up from the current frame. FRAME is the
7083 frame pointer of that frame. */
7086 s390_return_addr_rtx (int count
, rtx frame ATTRIBUTE_UNUSED
)
7091 /* Without backchain, we fail for all but the current frame. */
7093 if (!TARGET_BACKCHAIN
&& count
> 0)
7096 /* For the current frame, we need to make sure the initial
7097 value of RETURN_REGNUM is actually saved. */
7101 /* On non-z architectures branch splitting could overwrite r14. */
7102 if (TARGET_CPU_ZARCH
)
7103 return get_hard_reg_initial_val (Pmode
, RETURN_REGNUM
);
7106 cfun_frame_layout
.save_return_addr_p
= true;
7107 return gen_rtx_MEM (Pmode
, return_address_pointer_rtx
);
7111 if (TARGET_PACKED_STACK
)
7112 offset
= -2 * UNITS_PER_LONG
;
7114 offset
= RETURN_REGNUM
* UNITS_PER_LONG
;
7116 addr
= plus_constant (frame
, offset
);
7117 addr
= memory_address (Pmode
, addr
);
7118 return gen_rtx_MEM (Pmode
, addr
);
7121 /* Return an RTL expression representing the back chain stored in
7122 the current stack frame. */
7125 s390_back_chain_rtx (void)
7129 gcc_assert (TARGET_BACKCHAIN
);
7131 if (TARGET_PACKED_STACK
)
7132 chain
= plus_constant (stack_pointer_rtx
,
7133 STACK_POINTER_OFFSET
- UNITS_PER_LONG
);
7135 chain
= stack_pointer_rtx
;
7137 chain
= gen_rtx_MEM (Pmode
, chain
);
7141 /* Find first call clobbered register unused in a function.
7142 This could be used as base register in a leaf function
7143 or for holding the return address before epilogue. */
7146 find_unused_clobbered_reg (void)
7149 for (i
= 0; i
< 6; i
++)
7150 if (!df_regs_ever_live_p (i
))
7156 /* Helper function for s390_regs_ever_clobbered. Sets the fields in DATA for all
7157 clobbered hard regs in SETREG. */
7160 s390_reg_clobbered_rtx (rtx setreg
, const_rtx set_insn ATTRIBUTE_UNUSED
, void *data
)
7162 int *regs_ever_clobbered
= (int *)data
;
7163 unsigned int i
, regno
;
7164 enum machine_mode mode
= GET_MODE (setreg
);
7166 if (GET_CODE (setreg
) == SUBREG
)
7168 rtx inner
= SUBREG_REG (setreg
);
7169 if (!GENERAL_REG_P (inner
))
7171 regno
= subreg_regno (setreg
);
7173 else if (GENERAL_REG_P (setreg
))
7174 regno
= REGNO (setreg
);
7179 i
< regno
+ HARD_REGNO_NREGS (regno
, mode
);
7181 regs_ever_clobbered
[i
] = 1;
7184 /* Walks through all basic blocks of the current function looking
7185 for clobbered hard regs using s390_reg_clobbered_rtx. The fields
7186 of the passed integer array REGS_EVER_CLOBBERED are set to one for
7187 each of those regs. */
7190 s390_regs_ever_clobbered (int *regs_ever_clobbered
)
7196 memset (regs_ever_clobbered
, 0, 16 * sizeof (int));
7198 /* For non-leaf functions we have to consider all call clobbered regs to be
7200 if (!current_function_is_leaf
)
7202 for (i
= 0; i
< 16; i
++)
7203 regs_ever_clobbered
[i
] = call_really_used_regs
[i
];
7206 /* Make the "magic" eh_return registers live if necessary. For regs_ever_live
7207 this work is done by liveness analysis (mark_regs_live_at_end).
7208 Special care is needed for functions containing landing pads. Landing pads
7209 may use the eh registers, but the code which sets these registers is not
7210 contained in that function. Hence s390_regs_ever_clobbered is not able to
7211 deal with this automatically. */
7212 if (crtl
->calls_eh_return
|| cfun
->machine
->has_landing_pad_p
)
7213 for (i
= 0; EH_RETURN_DATA_REGNO (i
) != INVALID_REGNUM
; i
++)
7214 if (crtl
->calls_eh_return
7215 || (cfun
->machine
->has_landing_pad_p
7216 && df_regs_ever_live_p (EH_RETURN_DATA_REGNO (i
))))
7217 regs_ever_clobbered
[EH_RETURN_DATA_REGNO (i
)] = 1;
7219 /* For nonlocal gotos all call-saved registers have to be saved.
7220 This flag is also set for the unwinding code in libgcc.
7221 See expand_builtin_unwind_init. For regs_ever_live this is done by
7223 if (cfun
->has_nonlocal_label
)
7224 for (i
= 0; i
< 16; i
++)
7225 if (!call_really_used_regs
[i
])
7226 regs_ever_clobbered
[i
] = 1;
7228 FOR_EACH_BB (cur_bb
)
7230 FOR_BB_INSNS (cur_bb
, cur_insn
)
7232 if (INSN_P (cur_insn
))
7233 note_stores (PATTERN (cur_insn
),
7234 s390_reg_clobbered_rtx
,
7235 regs_ever_clobbered
);
7240 /* Determine the frame area which actually has to be accessed
7241 in the function epilogue. The values are stored at the
7242 given pointers AREA_BOTTOM (address of the lowest used stack
7243 address) and AREA_TOP (address of the first item which does
7244 not belong to the stack frame). */
7247 s390_frame_area (int *area_bottom
, int *area_top
)
7255 if (cfun_frame_layout
.first_restore_gpr
!= -1)
7257 b
= (cfun_frame_layout
.gprs_offset
7258 + cfun_frame_layout
.first_restore_gpr
* UNITS_PER_LONG
);
7259 t
= b
+ (cfun_frame_layout
.last_restore_gpr
7260 - cfun_frame_layout
.first_restore_gpr
+ 1) * UNITS_PER_LONG
;
7263 if (TARGET_64BIT
&& cfun_save_high_fprs_p
)
7265 b
= MIN (b
, cfun_frame_layout
.f8_offset
);
7266 t
= MAX (t
, (cfun_frame_layout
.f8_offset
7267 + cfun_frame_layout
.high_fprs
* 8));
7271 for (i
= 2; i
< 4; i
++)
7272 if (cfun_fpr_bit_p (i
))
7274 b
= MIN (b
, cfun_frame_layout
.f4_offset
+ (i
- 2) * 8);
7275 t
= MAX (t
, cfun_frame_layout
.f4_offset
+ (i
- 1) * 8);
7282 /* Fill cfun->machine with info about register usage of current function.
7283 Return in CLOBBERED_REGS which GPRs are currently considered set. */
7286 s390_register_info (int clobbered_regs
[])
7290 /* fprs 8 - 15 are call saved for 64 Bit ABI. */
7291 cfun_frame_layout
.fpr_bitmap
= 0;
7292 cfun_frame_layout
.high_fprs
= 0;
7294 for (i
= 24; i
< 32; i
++)
7295 if (df_regs_ever_live_p (i
) && !global_regs
[i
])
7297 cfun_set_fpr_bit (i
- 16);
7298 cfun_frame_layout
.high_fprs
++;
7301 /* Find first and last gpr to be saved. We trust regs_ever_live
7302 data, except that we don't save and restore global registers.
7304 Also, all registers with special meaning to the compiler need
7305 to be handled extra. */
7307 s390_regs_ever_clobbered (clobbered_regs
);
7309 for (i
= 0; i
< 16; i
++)
7310 clobbered_regs
[i
] = clobbered_regs
[i
] && !global_regs
[i
] && !fixed_regs
[i
];
7312 if (frame_pointer_needed
)
7313 clobbered_regs
[HARD_FRAME_POINTER_REGNUM
] = 1;
7316 clobbered_regs
[PIC_OFFSET_TABLE_REGNUM
]
7317 |= df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM
);
7319 clobbered_regs
[BASE_REGNUM
]
7320 |= (cfun
->machine
->base_reg
7321 && REGNO (cfun
->machine
->base_reg
) == BASE_REGNUM
);
7323 clobbered_regs
[RETURN_REGNUM
]
7324 |= (!current_function_is_leaf
7325 || TARGET_TPF_PROFILING
7326 || cfun
->machine
->split_branches_pending_p
7327 || cfun_frame_layout
.save_return_addr_p
7328 || crtl
->calls_eh_return
7331 clobbered_regs
[STACK_POINTER_REGNUM
]
7332 |= (!current_function_is_leaf
7333 || TARGET_TPF_PROFILING
7334 || cfun_save_high_fprs_p
7335 || get_frame_size () > 0
7336 || cfun
->calls_alloca
7339 for (i
= 6; i
< 16; i
++)
7340 if (df_regs_ever_live_p (i
) || clobbered_regs
[i
])
7342 for (j
= 15; j
> i
; j
--)
7343 if (df_regs_ever_live_p (j
) || clobbered_regs
[j
])
7348 /* Nothing to save/restore. */
7349 cfun_frame_layout
.first_save_gpr_slot
= -1;
7350 cfun_frame_layout
.last_save_gpr_slot
= -1;
7351 cfun_frame_layout
.first_save_gpr
= -1;
7352 cfun_frame_layout
.first_restore_gpr
= -1;
7353 cfun_frame_layout
.last_save_gpr
= -1;
7354 cfun_frame_layout
.last_restore_gpr
= -1;
7358 /* Save slots for gprs from i to j. */
7359 cfun_frame_layout
.first_save_gpr_slot
= i
;
7360 cfun_frame_layout
.last_save_gpr_slot
= j
;
7362 for (i
= cfun_frame_layout
.first_save_gpr_slot
;
7363 i
< cfun_frame_layout
.last_save_gpr_slot
+ 1;
7365 if (clobbered_regs
[i
])
7368 for (j
= cfun_frame_layout
.last_save_gpr_slot
; j
> i
; j
--)
7369 if (clobbered_regs
[j
])
7372 if (i
== cfun_frame_layout
.last_save_gpr_slot
+ 1)
7374 /* Nothing to save/restore. */
7375 cfun_frame_layout
.first_save_gpr
= -1;
7376 cfun_frame_layout
.first_restore_gpr
= -1;
7377 cfun_frame_layout
.last_save_gpr
= -1;
7378 cfun_frame_layout
.last_restore_gpr
= -1;
7382 /* Save / Restore from gpr i to j. */
7383 cfun_frame_layout
.first_save_gpr
= i
;
7384 cfun_frame_layout
.first_restore_gpr
= i
;
7385 cfun_frame_layout
.last_save_gpr
= j
;
7386 cfun_frame_layout
.last_restore_gpr
= j
;
7392 /* Varargs functions need to save gprs 2 to 6. */
7393 if (cfun
->va_list_gpr_size
7394 && crtl
->args
.info
.gprs
< GP_ARG_NUM_REG
)
7396 int min_gpr
= crtl
->args
.info
.gprs
;
7397 int max_gpr
= min_gpr
+ cfun
->va_list_gpr_size
;
7398 if (max_gpr
> GP_ARG_NUM_REG
)
7399 max_gpr
= GP_ARG_NUM_REG
;
7401 if (cfun_frame_layout
.first_save_gpr
== -1
7402 || cfun_frame_layout
.first_save_gpr
> 2 + min_gpr
)
7404 cfun_frame_layout
.first_save_gpr
= 2 + min_gpr
;
7405 cfun_frame_layout
.first_save_gpr_slot
= 2 + min_gpr
;
7408 if (cfun_frame_layout
.last_save_gpr
== -1
7409 || cfun_frame_layout
.last_save_gpr
< 2 + max_gpr
- 1)
7411 cfun_frame_layout
.last_save_gpr
= 2 + max_gpr
- 1;
7412 cfun_frame_layout
.last_save_gpr_slot
= 2 + max_gpr
- 1;
7416 /* Mark f0, f2 for 31 bit and f0-f4 for 64 bit to be saved. */
7417 if (TARGET_HARD_FLOAT
&& cfun
->va_list_fpr_size
7418 && crtl
->args
.info
.fprs
< FP_ARG_NUM_REG
)
7420 int min_fpr
= crtl
->args
.info
.fprs
;
7421 int max_fpr
= min_fpr
+ cfun
->va_list_fpr_size
;
7422 if (max_fpr
> FP_ARG_NUM_REG
)
7423 max_fpr
= FP_ARG_NUM_REG
;
7425 /* ??? This is currently required to ensure proper location
7426 of the fpr save slots within the va_list save area. */
7427 if (TARGET_PACKED_STACK
)
7430 for (i
= min_fpr
; i
< max_fpr
; i
++)
7431 cfun_set_fpr_bit (i
);
7436 for (i
= 2; i
< 4; i
++)
7437 if (df_regs_ever_live_p (i
+ 16) && !global_regs
[i
+ 16])
7438 cfun_set_fpr_bit (i
);
7441 /* Fill cfun->machine with info about frame of current function. */
7444 s390_frame_info (void)
7448 cfun_frame_layout
.frame_size
= get_frame_size ();
7449 if (!TARGET_64BIT
&& cfun_frame_layout
.frame_size
> 0x7fff0000)
7450 fatal_error ("total size of local variables exceeds architecture limit");
7452 if (!TARGET_PACKED_STACK
)
7454 cfun_frame_layout
.backchain_offset
= 0;
7455 cfun_frame_layout
.f0_offset
= 16 * UNITS_PER_LONG
;
7456 cfun_frame_layout
.f4_offset
= cfun_frame_layout
.f0_offset
+ 2 * 8;
7457 cfun_frame_layout
.f8_offset
= -cfun_frame_layout
.high_fprs
* 8;
7458 cfun_frame_layout
.gprs_offset
= (cfun_frame_layout
.first_save_gpr_slot
7461 else if (TARGET_BACKCHAIN
) /* kernel stack layout */
7463 cfun_frame_layout
.backchain_offset
= (STACK_POINTER_OFFSET
7465 cfun_frame_layout
.gprs_offset
7466 = (cfun_frame_layout
.backchain_offset
7467 - (STACK_POINTER_REGNUM
- cfun_frame_layout
.first_save_gpr_slot
+ 1)
7472 cfun_frame_layout
.f4_offset
7473 = (cfun_frame_layout
.gprs_offset
7474 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7476 cfun_frame_layout
.f0_offset
7477 = (cfun_frame_layout
.f4_offset
7478 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7482 /* On 31 bit we have to care about alignment of the
7483 floating point regs to provide fastest access. */
7484 cfun_frame_layout
.f0_offset
7485 = ((cfun_frame_layout
.gprs_offset
7486 & ~(STACK_BOUNDARY
/ BITS_PER_UNIT
- 1))
7487 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7489 cfun_frame_layout
.f4_offset
7490 = (cfun_frame_layout
.f0_offset
7491 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7494 else /* no backchain */
7496 cfun_frame_layout
.f4_offset
7497 = (STACK_POINTER_OFFSET
7498 - 8 * (cfun_fpr_bit_p (2) + cfun_fpr_bit_p (3)));
7500 cfun_frame_layout
.f0_offset
7501 = (cfun_frame_layout
.f4_offset
7502 - 8 * (cfun_fpr_bit_p (0) + cfun_fpr_bit_p (1)));
7504 cfun_frame_layout
.gprs_offset
7505 = cfun_frame_layout
.f0_offset
- cfun_gprs_save_area_size
;
7508 if (current_function_is_leaf
7509 && !TARGET_TPF_PROFILING
7510 && cfun_frame_layout
.frame_size
== 0
7511 && !cfun_save_high_fprs_p
7512 && !cfun
->calls_alloca
7516 if (!TARGET_PACKED_STACK
)
7517 cfun_frame_layout
.frame_size
+= (STACK_POINTER_OFFSET
7518 + crtl
->outgoing_args_size
7519 + cfun_frame_layout
.high_fprs
* 8);
7522 if (TARGET_BACKCHAIN
)
7523 cfun_frame_layout
.frame_size
+= UNITS_PER_LONG
;
7525 /* No alignment trouble here because f8-f15 are only saved under
7527 cfun_frame_layout
.f8_offset
= (MIN (MIN (cfun_frame_layout
.f0_offset
,
7528 cfun_frame_layout
.f4_offset
),
7529 cfun_frame_layout
.gprs_offset
)
7530 - cfun_frame_layout
.high_fprs
* 8);
7532 cfun_frame_layout
.frame_size
+= cfun_frame_layout
.high_fprs
* 8;
7534 for (i
= 0; i
< 8; i
++)
7535 if (cfun_fpr_bit_p (i
))
7536 cfun_frame_layout
.frame_size
+= 8;
7538 cfun_frame_layout
.frame_size
+= cfun_gprs_save_area_size
;
7540 /* If under 31 bit an odd number of gprs has to be saved we have to adjust
7541 the frame size to sustain 8 byte alignment of stack frames. */
7542 cfun_frame_layout
.frame_size
= ((cfun_frame_layout
.frame_size
+
7543 STACK_BOUNDARY
/ BITS_PER_UNIT
- 1)
7544 & ~(STACK_BOUNDARY
/ BITS_PER_UNIT
- 1));
7546 cfun_frame_layout
.frame_size
+= crtl
->outgoing_args_size
;
7550 /* Generate frame layout. Fills in register and frame data for the current
7551 function in cfun->machine. This routine can be called multiple times;
7552 it will re-do the complete frame layout every time. */
7555 s390_init_frame_layout (void)
7557 HOST_WIDE_INT frame_size
;
7559 int clobbered_regs
[16];
7561 /* On S/390 machines, we may need to perform branch splitting, which
7562 will require both base and return address register. We have no
7563 choice but to assume we're going to need them until right at the
7564 end of the machine dependent reorg phase. */
7565 if (!TARGET_CPU_ZARCH
)
7566 cfun
->machine
->split_branches_pending_p
= true;
7570 frame_size
= cfun_frame_layout
.frame_size
;
7572 /* Try to predict whether we'll need the base register. */
7573 base_used
= cfun
->machine
->split_branches_pending_p
7574 || crtl
->uses_const_pool
7575 || (!DISP_IN_RANGE (frame_size
)
7576 && !CONST_OK_FOR_K (frame_size
));
7578 /* Decide which register to use as literal pool base. In small
7579 leaf functions, try to use an unused call-clobbered register
7580 as base register to avoid save/restore overhead. */
7582 cfun
->machine
->base_reg
= NULL_RTX
;
7583 else if (current_function_is_leaf
&& !df_regs_ever_live_p (5))
7584 cfun
->machine
->base_reg
= gen_rtx_REG (Pmode
, 5);
7586 cfun
->machine
->base_reg
= gen_rtx_REG (Pmode
, BASE_REGNUM
);
7588 s390_register_info (clobbered_regs
);
7591 while (frame_size
!= cfun_frame_layout
.frame_size
);
7594 /* Update frame layout. Recompute actual register save data based on
7595 current info and update regs_ever_live for the special registers.
7596 May be called multiple times, but may never cause *more* registers
7597 to be saved than s390_init_frame_layout allocated room for. */
7600 s390_update_frame_layout (void)
7602 int clobbered_regs
[16];
7604 s390_register_info (clobbered_regs
);
7606 df_set_regs_ever_live (BASE_REGNUM
,
7607 clobbered_regs
[BASE_REGNUM
] ? true : false);
7608 df_set_regs_ever_live (RETURN_REGNUM
,
7609 clobbered_regs
[RETURN_REGNUM
] ? true : false);
7610 df_set_regs_ever_live (STACK_POINTER_REGNUM
,
7611 clobbered_regs
[STACK_POINTER_REGNUM
] ? true : false);
7613 if (cfun
->machine
->base_reg
)
7614 df_set_regs_ever_live (REGNO (cfun
->machine
->base_reg
), true);
7617 /* Return true if it is legal to put a value with MODE into REGNO. */
7620 s390_hard_regno_mode_ok (unsigned int regno
, enum machine_mode mode
)
7622 switch (REGNO_REG_CLASS (regno
))
7625 if (REGNO_PAIR_OK (regno
, mode
))
7627 if (mode
== SImode
|| mode
== DImode
)
7630 if (FLOAT_MODE_P (mode
) && GET_MODE_CLASS (mode
) != MODE_VECTOR_FLOAT
)
7635 if (FRAME_REGNO_P (regno
) && mode
== Pmode
)
7640 if (REGNO_PAIR_OK (regno
, mode
))
7643 || (mode
!= TFmode
&& mode
!= TCmode
&& mode
!= TDmode
))
7648 if (GET_MODE_CLASS (mode
) == MODE_CC
)
7652 if (REGNO_PAIR_OK (regno
, mode
))
7654 if (mode
== SImode
|| mode
== Pmode
)
7665 /* Return nonzero if register OLD_REG can be renamed to register NEW_REG. */
7668 s390_hard_regno_rename_ok (unsigned int old_reg
, unsigned int new_reg
)
7670 /* Once we've decided upon a register to use as base register, it must
7671 no longer be used for any other purpose. */
7672 if (cfun
->machine
->base_reg
)
7673 if (REGNO (cfun
->machine
->base_reg
) == old_reg
7674 || REGNO (cfun
->machine
->base_reg
) == new_reg
)
7680 /* Maximum number of registers to represent a value of mode MODE
7681 in a register of class RCLASS. */
7684 s390_class_max_nregs (enum reg_class rclass
, enum machine_mode mode
)
7689 if (GET_MODE_CLASS (mode
) == MODE_COMPLEX_FLOAT
)
7690 return 2 * ((GET_MODE_SIZE (mode
) / 2 + 8 - 1) / 8);
7692 return (GET_MODE_SIZE (mode
) + 8 - 1) / 8;
7694 return (GET_MODE_SIZE (mode
) + 4 - 1) / 4;
7698 return (GET_MODE_SIZE (mode
) + UNITS_PER_WORD
- 1) / UNITS_PER_WORD
;
7701 /* Return true if register FROM can be eliminated via register TO. */
7704 s390_can_eliminate (const int from
, const int to
)
7706 /* On zSeries machines, we have not marked the base register as fixed.
7707 Instead, we have an elimination rule BASE_REGNUM -> BASE_REGNUM.
7708 If a function requires the base register, we say here that this
7709 elimination cannot be performed. This will cause reload to free
7710 up the base register (as if it were fixed). On the other hand,
7711 if the current function does *not* require the base register, we
7712 say here the elimination succeeds, which in turn allows reload
7713 to allocate the base register for any other purpose. */
7714 if (from
== BASE_REGNUM
&& to
== BASE_REGNUM
)
7716 if (TARGET_CPU_ZARCH
)
7718 s390_init_frame_layout ();
7719 return cfun
->machine
->base_reg
== NULL_RTX
;
7725 /* Everything else must point into the stack frame. */
7726 gcc_assert (to
== STACK_POINTER_REGNUM
7727 || to
== HARD_FRAME_POINTER_REGNUM
);
7729 gcc_assert (from
== FRAME_POINTER_REGNUM
7730 || from
== ARG_POINTER_REGNUM
7731 || from
== RETURN_ADDRESS_POINTER_REGNUM
);
7733 /* Make sure we actually saved the return address. */
7734 if (from
== RETURN_ADDRESS_POINTER_REGNUM
)
7735 if (!crtl
->calls_eh_return
7737 && !cfun_frame_layout
.save_return_addr_p
)
7743 /* Return offset between register FROM and TO initially after prolog. */
7746 s390_initial_elimination_offset (int from
, int to
)
7748 HOST_WIDE_INT offset
;
7751 /* ??? Why are we called for non-eliminable pairs? */
7752 if (!s390_can_eliminate (from
, to
))
7757 case FRAME_POINTER_REGNUM
:
7758 offset
= (get_frame_size()
7759 + STACK_POINTER_OFFSET
7760 + crtl
->outgoing_args_size
);
7763 case ARG_POINTER_REGNUM
:
7764 s390_init_frame_layout ();
7765 offset
= cfun_frame_layout
.frame_size
+ STACK_POINTER_OFFSET
;
7768 case RETURN_ADDRESS_POINTER_REGNUM
:
7769 s390_init_frame_layout ();
7770 index
= RETURN_REGNUM
- cfun_frame_layout
.first_save_gpr_slot
;
7771 gcc_assert (index
>= 0);
7772 offset
= cfun_frame_layout
.frame_size
+ cfun_frame_layout
.gprs_offset
;
7773 offset
+= index
* UNITS_PER_LONG
;
7787 /* Emit insn to save fpr REGNUM at offset OFFSET relative
7788 to register BASE. Return generated insn. */
7791 save_fpr (rtx base
, int offset
, int regnum
)
7794 addr
= gen_rtx_MEM (DFmode
, plus_constant (base
, offset
));
7796 if (regnum
>= 16 && regnum
<= (16 + FP_ARG_NUM_REG
))
7797 set_mem_alias_set (addr
, get_varargs_alias_set ());
7799 set_mem_alias_set (addr
, get_frame_alias_set ());
7801 return emit_move_insn (addr
, gen_rtx_REG (DFmode
, regnum
));
7804 /* Emit insn to restore fpr REGNUM from offset OFFSET relative
7805 to register BASE. Return generated insn. */
7808 restore_fpr (rtx base
, int offset
, int regnum
)
7811 addr
= gen_rtx_MEM (DFmode
, plus_constant (base
, offset
));
7812 set_mem_alias_set (addr
, get_frame_alias_set ());
7814 return emit_move_insn (gen_rtx_REG (DFmode
, regnum
), addr
);
7817 /* Return true if REGNO is a global register, but not one
7818 of the special ones that need to be saved/restored in anyway. */
7821 global_not_special_regno_p (int regno
)
7823 return (global_regs
[regno
]
7824 /* These registers are special and need to be
7825 restored in any case. */
7826 && !(regno
== STACK_POINTER_REGNUM
7827 || regno
== RETURN_REGNUM
7828 || regno
== BASE_REGNUM
7829 || (flag_pic
&& regno
== (int)PIC_OFFSET_TABLE_REGNUM
)));
7832 /* Generate insn to save registers FIRST to LAST into
7833 the register save area located at offset OFFSET
7834 relative to register BASE. */
7837 save_gprs (rtx base
, int offset
, int first
, int last
)
7839 rtx addr
, insn
, note
;
7842 addr
= plus_constant (base
, offset
);
7843 addr
= gen_rtx_MEM (Pmode
, addr
);
7845 set_mem_alias_set (addr
, get_frame_alias_set ());
7847 /* Special-case single register. */
7851 insn
= gen_movdi (addr
, gen_rtx_REG (Pmode
, first
));
7853 insn
= gen_movsi (addr
, gen_rtx_REG (Pmode
, first
));
7855 if (!global_not_special_regno_p (first
))
7856 RTX_FRAME_RELATED_P (insn
) = 1;
7861 insn
= gen_store_multiple (addr
,
7862 gen_rtx_REG (Pmode
, first
),
7863 GEN_INT (last
- first
+ 1));
7865 if (first
<= 6 && cfun
->stdarg
)
7866 for (i
= 0; i
< XVECLEN (PATTERN (insn
), 0); i
++)
7868 rtx mem
= XEXP (XVECEXP (PATTERN (insn
), 0, i
), 0);
7871 set_mem_alias_set (mem
, get_varargs_alias_set ());
7874 /* We need to set the FRAME_RELATED flag on all SETs
7875 inside the store-multiple pattern.
7877 However, we must not emit DWARF records for registers 2..5
7878 if they are stored for use by variable arguments ...
7880 ??? Unfortunately, it is not enough to simply not the
7881 FRAME_RELATED flags for those SETs, because the first SET
7882 of the PARALLEL is always treated as if it had the flag
7883 set, even if it does not. Therefore we emit a new pattern
7884 without those registers as REG_FRAME_RELATED_EXPR note. */
7886 if (first
>= 6 && !global_not_special_regno_p (first
))
7888 rtx pat
= PATTERN (insn
);
7890 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
7891 if (GET_CODE (XVECEXP (pat
, 0, i
)) == SET
7892 && !global_not_special_regno_p (REGNO (SET_SRC (XVECEXP (pat
,
7894 RTX_FRAME_RELATED_P (XVECEXP (pat
, 0, i
)) = 1;
7896 RTX_FRAME_RELATED_P (insn
) = 1;
7902 for (start
= first
>= 6 ? first
: 6; start
<= last
; start
++)
7903 if (!global_not_special_regno_p (start
))
7909 addr
= plus_constant (base
, offset
+ (start
- first
) * UNITS_PER_LONG
);
7910 note
= gen_store_multiple (gen_rtx_MEM (Pmode
, addr
),
7911 gen_rtx_REG (Pmode
, start
),
7912 GEN_INT (last
- start
+ 1));
7913 note
= PATTERN (note
);
7915 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
, note
);
7917 for (i
= 0; i
< XVECLEN (note
, 0); i
++)
7918 if (GET_CODE (XVECEXP (note
, 0, i
)) == SET
7919 && !global_not_special_regno_p (REGNO (SET_SRC (XVECEXP (note
,
7921 RTX_FRAME_RELATED_P (XVECEXP (note
, 0, i
)) = 1;
7923 RTX_FRAME_RELATED_P (insn
) = 1;
7929 /* Generate insn to restore registers FIRST to LAST from
7930 the register save area located at offset OFFSET
7931 relative to register BASE. */
7934 restore_gprs (rtx base
, int offset
, int first
, int last
)
7938 addr
= plus_constant (base
, offset
);
7939 addr
= gen_rtx_MEM (Pmode
, addr
);
7940 set_mem_alias_set (addr
, get_frame_alias_set ());
7942 /* Special-case single register. */
7946 insn
= gen_movdi (gen_rtx_REG (Pmode
, first
), addr
);
7948 insn
= gen_movsi (gen_rtx_REG (Pmode
, first
), addr
);
7953 insn
= gen_load_multiple (gen_rtx_REG (Pmode
, first
),
7955 GEN_INT (last
- first
+ 1));
7959 /* Return insn sequence to load the GOT register. */
7961 static GTY(()) rtx got_symbol
;
7963 s390_load_got (void)
7969 got_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "_GLOBAL_OFFSET_TABLE_");
7970 SYMBOL_REF_FLAGS (got_symbol
) = SYMBOL_FLAG_LOCAL
;
7975 if (TARGET_CPU_ZARCH
)
7977 emit_move_insn (pic_offset_table_rtx
, got_symbol
);
7983 offset
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, got_symbol
),
7984 UNSPEC_LTREL_OFFSET
);
7985 offset
= gen_rtx_CONST (Pmode
, offset
);
7986 offset
= force_const_mem (Pmode
, offset
);
7988 emit_move_insn (pic_offset_table_rtx
, offset
);
7990 offset
= gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, XEXP (offset
, 0)),
7992 offset
= gen_rtx_PLUS (Pmode
, pic_offset_table_rtx
, offset
);
7994 emit_move_insn (pic_offset_table_rtx
, offset
);
7997 insns
= get_insns ();
8002 /* This ties together stack memory (MEM with an alias set of frame_alias_set)
8003 and the change to the stack pointer. */
8006 s390_emit_stack_tie (void)
8008 rtx mem
= gen_frame_mem (BLKmode
,
8009 gen_rtx_REG (Pmode
, STACK_POINTER_REGNUM
));
8011 emit_insn (gen_stack_tie (mem
));
8014 /* Expand the prologue into a bunch of separate insns. */
8017 s390_emit_prologue (void)
8025 /* Complete frame layout. */
8027 s390_update_frame_layout ();
8029 /* Annotate all constant pool references to let the scheduler know
8030 they implicitly use the base register. */
8032 push_topmost_sequence ();
8034 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
8037 annotate_constant_pool_refs (&PATTERN (insn
));
8038 df_insn_rescan (insn
);
8041 pop_topmost_sequence ();
8043 /* Choose best register to use for temp use within prologue.
8044 See below for why TPF must use the register 1. */
8046 if (!has_hard_reg_initial_val (Pmode
, RETURN_REGNUM
)
8047 && !current_function_is_leaf
8048 && !TARGET_TPF_PROFILING
)
8049 temp_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
8051 temp_reg
= gen_rtx_REG (Pmode
, 1);
8053 /* Save call saved gprs. */
8054 if (cfun_frame_layout
.first_save_gpr
!= -1)
8056 insn
= save_gprs (stack_pointer_rtx
,
8057 cfun_frame_layout
.gprs_offset
+
8058 UNITS_PER_LONG
* (cfun_frame_layout
.first_save_gpr
8059 - cfun_frame_layout
.first_save_gpr_slot
),
8060 cfun_frame_layout
.first_save_gpr
,
8061 cfun_frame_layout
.last_save_gpr
);
8065 /* Dummy insn to mark literal pool slot. */
8067 if (cfun
->machine
->base_reg
)
8068 emit_insn (gen_main_pool (cfun
->machine
->base_reg
));
8070 offset
= cfun_frame_layout
.f0_offset
;
8072 /* Save f0 and f2. */
8073 for (i
= 0; i
< 2; i
++)
8075 if (cfun_fpr_bit_p (i
))
8077 save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8080 else if (!TARGET_PACKED_STACK
)
8084 /* Save f4 and f6. */
8085 offset
= cfun_frame_layout
.f4_offset
;
8086 for (i
= 2; i
< 4; i
++)
8088 if (cfun_fpr_bit_p (i
))
8090 insn
= save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8093 /* If f4 and f6 are call clobbered they are saved due to stdargs and
8094 therefore are not frame related. */
8095 if (!call_really_used_regs
[i
+ 16])
8096 RTX_FRAME_RELATED_P (insn
) = 1;
8098 else if (!TARGET_PACKED_STACK
)
8102 if (TARGET_PACKED_STACK
8103 && cfun_save_high_fprs_p
8104 && cfun_frame_layout
.f8_offset
+ cfun_frame_layout
.high_fprs
* 8 > 0)
8106 offset
= (cfun_frame_layout
.f8_offset
8107 + (cfun_frame_layout
.high_fprs
- 1) * 8);
8109 for (i
= 15; i
> 7 && offset
>= 0; i
--)
8110 if (cfun_fpr_bit_p (i
))
8112 insn
= save_fpr (stack_pointer_rtx
, offset
, i
+ 16);
8114 RTX_FRAME_RELATED_P (insn
) = 1;
8117 if (offset
>= cfun_frame_layout
.f8_offset
)
8121 if (!TARGET_PACKED_STACK
)
8122 next_fpr
= cfun_save_high_fprs_p
? 31 : 0;
8124 if (flag_stack_usage
)
8125 current_function_static_stack_size
= cfun_frame_layout
.frame_size
;
8127 /* Decrement stack pointer. */
8129 if (cfun_frame_layout
.frame_size
> 0)
8131 rtx frame_off
= GEN_INT (-cfun_frame_layout
.frame_size
);
8134 if (s390_stack_size
)
8136 HOST_WIDE_INT stack_guard
;
8138 if (s390_stack_guard
)
8139 stack_guard
= s390_stack_guard
;
8142 /* If no value for stack guard is provided the smallest power of 2
8143 larger than the current frame size is chosen. */
8145 while (stack_guard
< cfun_frame_layout
.frame_size
)
8149 if (cfun_frame_layout
.frame_size
>= s390_stack_size
)
8151 warning (0, "frame size of function %qs is "
8152 HOST_WIDE_INT_PRINT_DEC
8153 " bytes exceeding user provided stack limit of "
8154 HOST_WIDE_INT_PRINT_DEC
" bytes. "
8155 "An unconditional trap is added.",
8156 current_function_name(), cfun_frame_layout
.frame_size
,
8158 emit_insn (gen_trap ());
8162 /* stack_guard has to be smaller than s390_stack_size.
8163 Otherwise we would emit an AND with zero which would
8164 not match the test under mask pattern. */
8165 if (stack_guard
>= s390_stack_size
)
8167 warning (0, "frame size of function %qs is "
8168 HOST_WIDE_INT_PRINT_DEC
8169 " bytes which is more than half the stack size. "
8170 "The dynamic check would not be reliable. "
8171 "No check emitted for this function.",
8172 current_function_name(),
8173 cfun_frame_layout
.frame_size
);
8177 HOST_WIDE_INT stack_check_mask
= ((s390_stack_size
- 1)
8178 & ~(stack_guard
- 1));
8180 rtx t
= gen_rtx_AND (Pmode
, stack_pointer_rtx
,
8181 GEN_INT (stack_check_mask
));
8183 emit_insn (gen_ctrapdi4 (gen_rtx_EQ (VOIDmode
,
8185 t
, const0_rtx
, const0_rtx
));
8187 emit_insn (gen_ctrapsi4 (gen_rtx_EQ (VOIDmode
,
8189 t
, const0_rtx
, const0_rtx
));
8194 if (s390_warn_framesize
> 0
8195 && cfun_frame_layout
.frame_size
>= s390_warn_framesize
)
8196 warning (0, "frame size of %qs is " HOST_WIDE_INT_PRINT_DEC
" bytes",
8197 current_function_name (), cfun_frame_layout
.frame_size
);
8199 if (s390_warn_dynamicstack_p
&& cfun
->calls_alloca
)
8200 warning (0, "%qs uses dynamic stack allocation", current_function_name ());
8202 /* Save incoming stack pointer into temp reg. */
8203 if (TARGET_BACKCHAIN
|| next_fpr
)
8204 insn
= emit_insn (gen_move_insn (temp_reg
, stack_pointer_rtx
));
8206 /* Subtract frame size from stack pointer. */
8208 if (DISP_IN_RANGE (INTVAL (frame_off
)))
8210 insn
= gen_rtx_SET (VOIDmode
, stack_pointer_rtx
,
8211 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
8213 insn
= emit_insn (insn
);
8217 if (!CONST_OK_FOR_K (INTVAL (frame_off
)))
8218 frame_off
= force_const_mem (Pmode
, frame_off
);
8220 insn
= emit_insn (gen_add2_insn (stack_pointer_rtx
, frame_off
));
8221 annotate_constant_pool_refs (&PATTERN (insn
));
8224 RTX_FRAME_RELATED_P (insn
) = 1;
8225 real_frame_off
= GEN_INT (-cfun_frame_layout
.frame_size
);
8226 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
8227 gen_rtx_SET (VOIDmode
, stack_pointer_rtx
,
8228 gen_rtx_PLUS (Pmode
, stack_pointer_rtx
,
8231 /* Set backchain. */
8233 if (TARGET_BACKCHAIN
)
8235 if (cfun_frame_layout
.backchain_offset
)
8236 addr
= gen_rtx_MEM (Pmode
,
8237 plus_constant (stack_pointer_rtx
,
8238 cfun_frame_layout
.backchain_offset
));
8240 addr
= gen_rtx_MEM (Pmode
, stack_pointer_rtx
);
8241 set_mem_alias_set (addr
, get_frame_alias_set ());
8242 insn
= emit_insn (gen_move_insn (addr
, temp_reg
));
8245 /* If we support non-call exceptions (e.g. for Java),
8246 we need to make sure the backchain pointer is set up
8247 before any possibly trapping memory access. */
8248 if (TARGET_BACKCHAIN
&& cfun
->can_throw_non_call_exceptions
)
8250 addr
= gen_rtx_MEM (BLKmode
, gen_rtx_SCRATCH (VOIDmode
));
8251 emit_clobber (addr
);
8255 /* Save fprs 8 - 15 (64 bit ABI). */
8257 if (cfun_save_high_fprs_p
&& next_fpr
)
8259 /* If the stack might be accessed through a different register
8260 we have to make sure that the stack pointer decrement is not
8261 moved below the use of the stack slots. */
8262 s390_emit_stack_tie ();
8264 insn
= emit_insn (gen_add2_insn (temp_reg
,
8265 GEN_INT (cfun_frame_layout
.f8_offset
)));
8269 for (i
= 24; i
<= next_fpr
; i
++)
8270 if (cfun_fpr_bit_p (i
- 16))
8272 rtx addr
= plus_constant (stack_pointer_rtx
,
8273 cfun_frame_layout
.frame_size
8274 + cfun_frame_layout
.f8_offset
8277 insn
= save_fpr (temp_reg
, offset
, i
);
8279 RTX_FRAME_RELATED_P (insn
) = 1;
8280 add_reg_note (insn
, REG_FRAME_RELATED_EXPR
,
8281 gen_rtx_SET (VOIDmode
,
8282 gen_rtx_MEM (DFmode
, addr
),
8283 gen_rtx_REG (DFmode
, i
)));
8287 /* Set frame pointer, if needed. */
8289 if (frame_pointer_needed
)
8291 insn
= emit_move_insn (hard_frame_pointer_rtx
, stack_pointer_rtx
);
8292 RTX_FRAME_RELATED_P (insn
) = 1;
8295 /* Set up got pointer, if needed. */
8297 if (flag_pic
&& df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM
))
8299 rtx insns
= s390_load_got ();
8301 for (insn
= insns
; insn
; insn
= NEXT_INSN (insn
))
8302 annotate_constant_pool_refs (&PATTERN (insn
));
8307 if (TARGET_TPF_PROFILING
)
8309 /* Generate a BAS instruction to serve as a function
8310 entry intercept to facilitate the use of tracing
8311 algorithms located at the branch target. */
8312 emit_insn (gen_prologue_tpf ());
8314 /* Emit a blockage here so that all code
8315 lies between the profiling mechanisms. */
8316 emit_insn (gen_blockage ());
8320 /* Expand the epilogue into a bunch of separate insns. */
8323 s390_emit_epilogue (bool sibcall
)
8325 rtx frame_pointer
, return_reg
, cfa_restores
= NULL_RTX
;
8326 int area_bottom
, area_top
, offset
= 0;
8331 if (TARGET_TPF_PROFILING
)
8334 /* Generate a BAS instruction to serve as a function
8335 entry intercept to facilitate the use of tracing
8336 algorithms located at the branch target. */
8338 /* Emit a blockage here so that all code
8339 lies between the profiling mechanisms. */
8340 emit_insn (gen_blockage ());
8342 emit_insn (gen_epilogue_tpf ());
8345 /* Check whether to use frame or stack pointer for restore. */
8347 frame_pointer
= (frame_pointer_needed
8348 ? hard_frame_pointer_rtx
: stack_pointer_rtx
);
8350 s390_frame_area (&area_bottom
, &area_top
);
8352 /* Check whether we can access the register save area.
8353 If not, increment the frame pointer as required. */
8355 if (area_top
<= area_bottom
)
8357 /* Nothing to restore. */
8359 else if (DISP_IN_RANGE (cfun_frame_layout
.frame_size
+ area_bottom
)
8360 && DISP_IN_RANGE (cfun_frame_layout
.frame_size
+ area_top
- 1))
8362 /* Area is in range. */
8363 offset
= cfun_frame_layout
.frame_size
;
8367 rtx insn
, frame_off
, cfa
;
8369 offset
= area_bottom
< 0 ? -area_bottom
: 0;
8370 frame_off
= GEN_INT (cfun_frame_layout
.frame_size
- offset
);
8372 cfa
= gen_rtx_SET (VOIDmode
, frame_pointer
,
8373 gen_rtx_PLUS (Pmode
, frame_pointer
, frame_off
));
8374 if (DISP_IN_RANGE (INTVAL (frame_off
)))
8376 insn
= gen_rtx_SET (VOIDmode
, frame_pointer
,
8377 gen_rtx_PLUS (Pmode
, frame_pointer
, frame_off
));
8378 insn
= emit_insn (insn
);
8382 if (!CONST_OK_FOR_K (INTVAL (frame_off
)))
8383 frame_off
= force_const_mem (Pmode
, frame_off
);
8385 insn
= emit_insn (gen_add2_insn (frame_pointer
, frame_off
));
8386 annotate_constant_pool_refs (&PATTERN (insn
));
8388 add_reg_note (insn
, REG_CFA_ADJUST_CFA
, cfa
);
8389 RTX_FRAME_RELATED_P (insn
) = 1;
8392 /* Restore call saved fprs. */
8396 if (cfun_save_high_fprs_p
)
8398 next_offset
= cfun_frame_layout
.f8_offset
;
8399 for (i
= 24; i
< 32; i
++)
8401 if (cfun_fpr_bit_p (i
- 16))
8403 restore_fpr (frame_pointer
,
8404 offset
+ next_offset
, i
);
8406 = alloc_reg_note (REG_CFA_RESTORE
,
8407 gen_rtx_REG (DFmode
, i
), cfa_restores
);
8416 next_offset
= cfun_frame_layout
.f4_offset
;
8417 for (i
= 18; i
< 20; i
++)
8419 if (cfun_fpr_bit_p (i
- 16))
8421 restore_fpr (frame_pointer
,
8422 offset
+ next_offset
, i
);
8424 = alloc_reg_note (REG_CFA_RESTORE
,
8425 gen_rtx_REG (DFmode
, i
), cfa_restores
);
8428 else if (!TARGET_PACKED_STACK
)
8434 /* Return register. */
8436 return_reg
= gen_rtx_REG (Pmode
, RETURN_REGNUM
);
8438 /* Restore call saved gprs. */
8440 if (cfun_frame_layout
.first_restore_gpr
!= -1)
8445 /* Check for global register and save them
8446 to stack location from where they get restored. */
8448 for (i
= cfun_frame_layout
.first_restore_gpr
;
8449 i
<= cfun_frame_layout
.last_restore_gpr
;
8452 if (global_not_special_regno_p (i
))
8454 addr
= plus_constant (frame_pointer
,
8455 offset
+ cfun_frame_layout
.gprs_offset
8456 + (i
- cfun_frame_layout
.first_save_gpr_slot
)
8458 addr
= gen_rtx_MEM (Pmode
, addr
);
8459 set_mem_alias_set (addr
, get_frame_alias_set ());
8460 emit_move_insn (addr
, gen_rtx_REG (Pmode
, i
));
8464 = alloc_reg_note (REG_CFA_RESTORE
,
8465 gen_rtx_REG (Pmode
, i
), cfa_restores
);
8470 /* Fetch return address from stack before load multiple,
8471 this will do good for scheduling. */
8473 if (cfun_frame_layout
.save_return_addr_p
8474 || (cfun_frame_layout
.first_restore_gpr
< BASE_REGNUM
8475 && cfun_frame_layout
.last_restore_gpr
> RETURN_REGNUM
))
8477 int return_regnum
= find_unused_clobbered_reg();
8480 return_reg
= gen_rtx_REG (Pmode
, return_regnum
);
8482 addr
= plus_constant (frame_pointer
,
8483 offset
+ cfun_frame_layout
.gprs_offset
8485 - cfun_frame_layout
.first_save_gpr_slot
)
8487 addr
= gen_rtx_MEM (Pmode
, addr
);
8488 set_mem_alias_set (addr
, get_frame_alias_set ());
8489 emit_move_insn (return_reg
, addr
);
8493 insn
= restore_gprs (frame_pointer
,
8494 offset
+ cfun_frame_layout
.gprs_offset
8495 + (cfun_frame_layout
.first_restore_gpr
8496 - cfun_frame_layout
.first_save_gpr_slot
)
8498 cfun_frame_layout
.first_restore_gpr
,
8499 cfun_frame_layout
.last_restore_gpr
);
8500 insn
= emit_insn (insn
);
8501 REG_NOTES (insn
) = cfa_restores
;
8502 add_reg_note (insn
, REG_CFA_DEF_CFA
,
8503 plus_constant (stack_pointer_rtx
, STACK_POINTER_OFFSET
));
8504 RTX_FRAME_RELATED_P (insn
) = 1;
8510 /* Return to caller. */
8512 p
= rtvec_alloc (2);
8514 RTVEC_ELT (p
, 0) = gen_rtx_RETURN (VOIDmode
);
8515 RTVEC_ELT (p
, 1) = gen_rtx_USE (VOIDmode
, return_reg
);
8516 emit_jump_insn (gen_rtx_PARALLEL (VOIDmode
, p
));
8521 /* Return the size in bytes of a function argument of
8522 type TYPE and/or mode MODE. At least one of TYPE or
8523 MODE must be specified. */
8526 s390_function_arg_size (enum machine_mode mode
, const_tree type
)
8529 return int_size_in_bytes (type
);
8531 /* No type info available for some library calls ... */
8532 if (mode
!= BLKmode
)
8533 return GET_MODE_SIZE (mode
);
8535 /* If we have neither type nor mode, abort */
8539 /* Return true if a function argument of type TYPE and mode MODE
8540 is to be passed in a floating-point register, if available. */
8543 s390_function_arg_float (enum machine_mode mode
, const_tree type
)
8545 int size
= s390_function_arg_size (mode
, type
);
8549 /* Soft-float changes the ABI: no floating-point registers are used. */
8550 if (TARGET_SOFT_FLOAT
)
8553 /* No type info available for some library calls ... */
8555 return mode
== SFmode
|| mode
== DFmode
|| mode
== SDmode
|| mode
== DDmode
;
8557 /* The ABI says that record types with a single member are treated
8558 just like that member would be. */
8559 while (TREE_CODE (type
) == RECORD_TYPE
)
8561 tree field
, single
= NULL_TREE
;
8563 for (field
= TYPE_FIELDS (type
); field
; field
= DECL_CHAIN (field
))
8565 if (TREE_CODE (field
) != FIELD_DECL
)
8568 if (single
== NULL_TREE
)
8569 single
= TREE_TYPE (field
);
8574 if (single
== NULL_TREE
)
8580 return TREE_CODE (type
) == REAL_TYPE
;
8583 /* Return true if a function argument of type TYPE and mode MODE
8584 is to be passed in an integer register, or a pair of integer
8585 registers, if available. */
8588 s390_function_arg_integer (enum machine_mode mode
, const_tree type
)
8590 int size
= s390_function_arg_size (mode
, type
);
8594 /* No type info available for some library calls ... */
8596 return GET_MODE_CLASS (mode
) == MODE_INT
8597 || (TARGET_SOFT_FLOAT
&& SCALAR_FLOAT_MODE_P (mode
));
8599 /* We accept small integral (and similar) types. */
8600 if (INTEGRAL_TYPE_P (type
)
8601 || POINTER_TYPE_P (type
)
8602 || TREE_CODE (type
) == NULLPTR_TYPE
8603 || TREE_CODE (type
) == OFFSET_TYPE
8604 || (TARGET_SOFT_FLOAT
&& TREE_CODE (type
) == REAL_TYPE
))
8607 /* We also accept structs of size 1, 2, 4, 8 that are not
8608 passed in floating-point registers. */
8609 if (AGGREGATE_TYPE_P (type
)
8610 && exact_log2 (size
) >= 0
8611 && !s390_function_arg_float (mode
, type
))
8617 /* Return 1 if a function argument of type TYPE and mode MODE
8618 is to be passed by reference. The ABI specifies that only
8619 structures of size 1, 2, 4, or 8 bytes are passed by value,
8620 all other structures (and complex numbers) are passed by
8624 s390_pass_by_reference (CUMULATIVE_ARGS
*ca ATTRIBUTE_UNUSED
,
8625 enum machine_mode mode
, const_tree type
,
8626 bool named ATTRIBUTE_UNUSED
)
8628 int size
= s390_function_arg_size (mode
, type
);
8634 if (AGGREGATE_TYPE_P (type
) && exact_log2 (size
) < 0)
8637 if (TREE_CODE (type
) == COMPLEX_TYPE
8638 || TREE_CODE (type
) == VECTOR_TYPE
)
8645 /* Update the data in CUM to advance over an argument of mode MODE and
8646 data type TYPE. (TYPE is null for libcalls where that information
8647 may not be available.). The boolean NAMED specifies whether the
8648 argument is a named argument (as opposed to an unnamed argument
8649 matching an ellipsis). */
8652 s390_function_arg_advance (CUMULATIVE_ARGS
*cum
, enum machine_mode mode
,
8653 const_tree type
, bool named ATTRIBUTE_UNUSED
)
8655 if (s390_function_arg_float (mode
, type
))
8659 else if (s390_function_arg_integer (mode
, type
))
8661 int size
= s390_function_arg_size (mode
, type
);
8662 cum
->gprs
+= ((size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
);
8668 /* Define where to put the arguments to a function.
8669 Value is zero to push the argument on the stack,
8670 or a hard register in which to store the argument.
8672 MODE is the argument's machine mode.
8673 TYPE is the data type of the argument (as a tree).
8674 This is null for libcalls where that information may
8676 CUM is a variable of type CUMULATIVE_ARGS which gives info about
8677 the preceding args and about the function being called.
8678 NAMED is nonzero if this argument is a named parameter
8679 (otherwise it is an extra parameter matching an ellipsis).
8681 On S/390, we use general purpose registers 2 through 6 to
8682 pass integer, pointer, and certain structure arguments, and
8683 floating point registers 0 and 2 (0, 2, 4, and 6 on 64-bit)
8684 to pass floating point arguments. All remaining arguments
8685 are pushed to the stack. */
8688 s390_function_arg (CUMULATIVE_ARGS
*cum
, enum machine_mode mode
,
8689 const_tree type
, bool named ATTRIBUTE_UNUSED
)
8691 if (s390_function_arg_float (mode
, type
))
8693 if (cum
->fprs
+ 1 > FP_ARG_NUM_REG
)
8696 return gen_rtx_REG (mode
, cum
->fprs
+ 16);
8698 else if (s390_function_arg_integer (mode
, type
))
8700 int size
= s390_function_arg_size (mode
, type
);
8701 int n_gprs
= (size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
;
8703 if (cum
->gprs
+ n_gprs
> GP_ARG_NUM_REG
)
8705 else if (n_gprs
== 1 || UNITS_PER_WORD
== UNITS_PER_LONG
)
8706 return gen_rtx_REG (mode
, cum
->gprs
+ 2);
8707 else if (n_gprs
== 2)
8709 rtvec p
= rtvec_alloc (2);
8712 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, cum
->gprs
+ 2),
8715 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, cum
->gprs
+ 3),
8718 return gen_rtx_PARALLEL (mode
, p
);
8722 /* After the real arguments, expand_call calls us once again
8723 with a void_type_node type. Whatever we return here is
8724 passed as operand 2 to the call expanders.
8726 We don't need this feature ... */
8727 else if (type
== void_type_node
)
8733 /* Return true if return values of type TYPE should be returned
8734 in a memory buffer whose address is passed by the caller as
8735 hidden first argument. */
8738 s390_return_in_memory (const_tree type
, const_tree fundecl ATTRIBUTE_UNUSED
)
8740 /* We accept small integral (and similar) types. */
8741 if (INTEGRAL_TYPE_P (type
)
8742 || POINTER_TYPE_P (type
)
8743 || TREE_CODE (type
) == OFFSET_TYPE
8744 || TREE_CODE (type
) == REAL_TYPE
)
8745 return int_size_in_bytes (type
) > 8;
8747 /* Aggregates and similar constructs are always returned
8749 if (AGGREGATE_TYPE_P (type
)
8750 || TREE_CODE (type
) == COMPLEX_TYPE
8751 || TREE_CODE (type
) == VECTOR_TYPE
)
8754 /* ??? We get called on all sorts of random stuff from
8755 aggregate_value_p. We can't abort, but it's not clear
8756 what's safe to return. Pretend it's a struct I guess. */
8760 /* Function arguments and return values are promoted to word size. */
8762 static enum machine_mode
8763 s390_promote_function_mode (const_tree type
, enum machine_mode mode
,
8765 const_tree fntype ATTRIBUTE_UNUSED
,
8766 int for_return ATTRIBUTE_UNUSED
)
8768 if (INTEGRAL_MODE_P (mode
)
8769 && GET_MODE_SIZE (mode
) < UNITS_PER_LONG
)
8771 if (POINTER_TYPE_P (type
))
8772 *punsignedp
= POINTERS_EXTEND_UNSIGNED
;
8779 /* Define where to return a (scalar) value of type RET_TYPE.
8780 If RET_TYPE is null, define where to return a (scalar)
8781 value of mode MODE from a libcall. */
8784 s390_function_and_libcall_value (enum machine_mode mode
,
8785 const_tree ret_type
,
8786 const_tree fntype_or_decl
,
8787 bool outgoing ATTRIBUTE_UNUSED
)
8789 /* For normal functions perform the promotion as
8790 promote_function_mode would do. */
8793 int unsignedp
= TYPE_UNSIGNED (ret_type
);
8794 mode
= promote_function_mode (ret_type
, mode
, &unsignedp
,
8798 gcc_assert (GET_MODE_CLASS (mode
) == MODE_INT
|| SCALAR_FLOAT_MODE_P (mode
));
8799 gcc_assert (GET_MODE_SIZE (mode
) <= 8);
8801 if (TARGET_HARD_FLOAT
&& SCALAR_FLOAT_MODE_P (mode
))
8802 return gen_rtx_REG (mode
, 16);
8803 else if (GET_MODE_SIZE (mode
) <= UNITS_PER_LONG
8804 || UNITS_PER_LONG
== UNITS_PER_WORD
)
8805 return gen_rtx_REG (mode
, 2);
8806 else if (GET_MODE_SIZE (mode
) == 2 * UNITS_PER_LONG
)
8808 /* This case is triggered when returning a 64 bit value with
8809 -m31 -mzarch. Although the value would fit into a single
8810 register it has to be forced into a 32 bit register pair in
8811 order to match the ABI. */
8812 rtvec p
= rtvec_alloc (2);
8815 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, 2), const0_rtx
);
8817 = gen_rtx_EXPR_LIST (SImode
, gen_rtx_REG (SImode
, 3), GEN_INT (4));
8819 return gen_rtx_PARALLEL (mode
, p
);
8825 /* Define where to return a scalar return value of type RET_TYPE. */
8828 s390_function_value (const_tree ret_type
, const_tree fn_decl_or_type
,
8831 return s390_function_and_libcall_value (TYPE_MODE (ret_type
), ret_type
,
8832 fn_decl_or_type
, outgoing
);
8835 /* Define where to return a scalar libcall return value of mode
8839 s390_libcall_value (enum machine_mode mode
, const_rtx fun ATTRIBUTE_UNUSED
)
8841 return s390_function_and_libcall_value (mode
, NULL_TREE
,
8846 /* Create and return the va_list datatype.
8848 On S/390, va_list is an array type equivalent to
8850 typedef struct __va_list_tag
8854 void *__overflow_arg_area;
8855 void *__reg_save_area;
8858 where __gpr and __fpr hold the number of general purpose
8859 or floating point arguments used up to now, respectively,
8860 __overflow_arg_area points to the stack location of the
8861 next argument passed on the stack, and __reg_save_area
8862 always points to the start of the register area in the
8863 call frame of the current function. The function prologue
8864 saves all registers used for argument passing into this
8865 area if the function uses variable arguments. */
8868 s390_build_builtin_va_list (void)
8870 tree f_gpr
, f_fpr
, f_ovf
, f_sav
, record
, type_decl
;
8872 record
= lang_hooks
.types
.make_type (RECORD_TYPE
);
8875 build_decl (BUILTINS_LOCATION
,
8876 TYPE_DECL
, get_identifier ("__va_list_tag"), record
);
8878 f_gpr
= build_decl (BUILTINS_LOCATION
,
8879 FIELD_DECL
, get_identifier ("__gpr"),
8880 long_integer_type_node
);
8881 f_fpr
= build_decl (BUILTINS_LOCATION
,
8882 FIELD_DECL
, get_identifier ("__fpr"),
8883 long_integer_type_node
);
8884 f_ovf
= build_decl (BUILTINS_LOCATION
,
8885 FIELD_DECL
, get_identifier ("__overflow_arg_area"),
8887 f_sav
= build_decl (BUILTINS_LOCATION
,
8888 FIELD_DECL
, get_identifier ("__reg_save_area"),
8891 va_list_gpr_counter_field
= f_gpr
;
8892 va_list_fpr_counter_field
= f_fpr
;
8894 DECL_FIELD_CONTEXT (f_gpr
) = record
;
8895 DECL_FIELD_CONTEXT (f_fpr
) = record
;
8896 DECL_FIELD_CONTEXT (f_ovf
) = record
;
8897 DECL_FIELD_CONTEXT (f_sav
) = record
;
8899 TYPE_STUB_DECL (record
) = type_decl
;
8900 TYPE_NAME (record
) = type_decl
;
8901 TYPE_FIELDS (record
) = f_gpr
;
8902 DECL_CHAIN (f_gpr
) = f_fpr
;
8903 DECL_CHAIN (f_fpr
) = f_ovf
;
8904 DECL_CHAIN (f_ovf
) = f_sav
;
8906 layout_type (record
);
8908 /* The correct type is an array type of one element. */
8909 return build_array_type (record
, build_index_type (size_zero_node
));
8912 /* Implement va_start by filling the va_list structure VALIST.
8913 STDARG_P is always true, and ignored.
8914 NEXTARG points to the first anonymous stack argument.
8916 The following global variables are used to initialize
8917 the va_list structure:
8920 holds number of gprs and fprs used for named arguments.
8921 crtl->args.arg_offset_rtx:
8922 holds the offset of the first anonymous stack argument
8923 (relative to the virtual arg pointer). */
8926 s390_va_start (tree valist
, rtx nextarg ATTRIBUTE_UNUSED
)
8928 HOST_WIDE_INT n_gpr
, n_fpr
;
8930 tree f_gpr
, f_fpr
, f_ovf
, f_sav
;
8931 tree gpr
, fpr
, ovf
, sav
, t
;
8933 f_gpr
= TYPE_FIELDS (TREE_TYPE (va_list_type_node
));
8934 f_fpr
= DECL_CHAIN (f_gpr
);
8935 f_ovf
= DECL_CHAIN (f_fpr
);
8936 f_sav
= DECL_CHAIN (f_ovf
);
8938 valist
= build_simple_mem_ref (valist
);
8939 gpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_gpr
), valist
, f_gpr
, NULL_TREE
);
8940 fpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_fpr
), valist
, f_fpr
, NULL_TREE
);
8941 ovf
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovf
), valist
, f_ovf
, NULL_TREE
);
8942 sav
= build3 (COMPONENT_REF
, TREE_TYPE (f_sav
), valist
, f_sav
, NULL_TREE
);
8944 /* Count number of gp and fp argument registers used. */
8946 n_gpr
= crtl
->args
.info
.gprs
;
8947 n_fpr
= crtl
->args
.info
.fprs
;
8949 if (cfun
->va_list_gpr_size
)
8951 t
= build2 (MODIFY_EXPR
, TREE_TYPE (gpr
), gpr
,
8952 build_int_cst (NULL_TREE
, n_gpr
));
8953 TREE_SIDE_EFFECTS (t
) = 1;
8954 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8957 if (cfun
->va_list_fpr_size
)
8959 t
= build2 (MODIFY_EXPR
, TREE_TYPE (fpr
), fpr
,
8960 build_int_cst (NULL_TREE
, n_fpr
));
8961 TREE_SIDE_EFFECTS (t
) = 1;
8962 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8965 /* Find the overflow area. */
8966 if (n_gpr
+ cfun
->va_list_gpr_size
> GP_ARG_NUM_REG
8967 || n_fpr
+ cfun
->va_list_fpr_size
> FP_ARG_NUM_REG
)
8969 t
= make_tree (TREE_TYPE (ovf
), virtual_incoming_args_rtx
);
8971 off
= INTVAL (crtl
->args
.arg_offset_rtx
);
8972 off
= off
< 0 ? 0 : off
;
8973 if (TARGET_DEBUG_ARG
)
8974 fprintf (stderr
, "va_start: n_gpr = %d, n_fpr = %d off %d\n",
8975 (int)n_gpr
, (int)n_fpr
, off
);
8977 t
= build2 (POINTER_PLUS_EXPR
, TREE_TYPE (ovf
), t
, size_int (off
));
8979 t
= build2 (MODIFY_EXPR
, TREE_TYPE (ovf
), ovf
, t
);
8980 TREE_SIDE_EFFECTS (t
) = 1;
8981 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8984 /* Find the register save area. */
8985 if ((cfun
->va_list_gpr_size
&& n_gpr
< GP_ARG_NUM_REG
)
8986 || (cfun
->va_list_fpr_size
&& n_fpr
< FP_ARG_NUM_REG
))
8988 t
= make_tree (TREE_TYPE (sav
), return_address_pointer_rtx
);
8989 t
= build2 (POINTER_PLUS_EXPR
, TREE_TYPE (sav
), t
,
8990 size_int (-RETURN_REGNUM
* UNITS_PER_LONG
));
8992 t
= build2 (MODIFY_EXPR
, TREE_TYPE (sav
), sav
, t
);
8993 TREE_SIDE_EFFECTS (t
) = 1;
8994 expand_expr (t
, const0_rtx
, VOIDmode
, EXPAND_NORMAL
);
8998 /* Implement va_arg by updating the va_list structure
8999 VALIST as required to retrieve an argument of type
9000 TYPE, and returning that argument.
9002 Generates code equivalent to:
9004 if (integral value) {
9005 if (size <= 4 && args.gpr < 5 ||
9006 size > 4 && args.gpr < 4 )
9007 ret = args.reg_save_area[args.gpr+8]
9009 ret = *args.overflow_arg_area++;
9010 } else if (float value) {
9012 ret = args.reg_save_area[args.fpr+64]
9014 ret = *args.overflow_arg_area++;
9015 } else if (aggregate value) {
9017 ret = *args.reg_save_area[args.gpr]
9019 ret = **args.overflow_arg_area++;
9023 s390_gimplify_va_arg (tree valist
, tree type
, gimple_seq
*pre_p
,
9024 gimple_seq
*post_p ATTRIBUTE_UNUSED
)
9026 tree f_gpr
, f_fpr
, f_ovf
, f_sav
;
9027 tree gpr
, fpr
, ovf
, sav
, reg
, t
, u
;
9028 int indirect_p
, size
, n_reg
, sav_ofs
, sav_scale
, max_reg
;
9029 tree lab_false
, lab_over
, addr
;
9031 f_gpr
= TYPE_FIELDS (TREE_TYPE (va_list_type_node
));
9032 f_fpr
= DECL_CHAIN (f_gpr
);
9033 f_ovf
= DECL_CHAIN (f_fpr
);
9034 f_sav
= DECL_CHAIN (f_ovf
);
9036 valist
= build_va_arg_indirect_ref (valist
);
9037 gpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_gpr
), valist
, f_gpr
, NULL_TREE
);
9038 fpr
= build3 (COMPONENT_REF
, TREE_TYPE (f_fpr
), valist
, f_fpr
, NULL_TREE
);
9039 sav
= build3 (COMPONENT_REF
, TREE_TYPE (f_sav
), valist
, f_sav
, NULL_TREE
);
9041 /* The tree for args* cannot be shared between gpr/fpr and ovf since
9042 both appear on a lhs. */
9043 valist
= unshare_expr (valist
);
9044 ovf
= build3 (COMPONENT_REF
, TREE_TYPE (f_ovf
), valist
, f_ovf
, NULL_TREE
);
9046 size
= int_size_in_bytes (type
);
9048 if (pass_by_reference (NULL
, TYPE_MODE (type
), type
, false))
9050 if (TARGET_DEBUG_ARG
)
9052 fprintf (stderr
, "va_arg: aggregate type");
9056 /* Aggregates are passed by reference. */
9061 /* kernel stack layout on 31 bit: It is assumed here that no padding
9062 will be added by s390_frame_info because for va_args always an even
9063 number of gprs has to be saved r15-r2 = 14 regs. */
9064 sav_ofs
= 2 * UNITS_PER_LONG
;
9065 sav_scale
= UNITS_PER_LONG
;
9066 size
= UNITS_PER_LONG
;
9067 max_reg
= GP_ARG_NUM_REG
- n_reg
;
9069 else if (s390_function_arg_float (TYPE_MODE (type
), type
))
9071 if (TARGET_DEBUG_ARG
)
9073 fprintf (stderr
, "va_arg: float type");
9077 /* FP args go in FP registers, if present. */
9081 sav_ofs
= 16 * UNITS_PER_LONG
;
9083 max_reg
= FP_ARG_NUM_REG
- n_reg
;
9087 if (TARGET_DEBUG_ARG
)
9089 fprintf (stderr
, "va_arg: other type");
9093 /* Otherwise into GP registers. */
9096 n_reg
= (size
+ UNITS_PER_LONG
- 1) / UNITS_PER_LONG
;
9098 /* kernel stack layout on 31 bit: It is assumed here that no padding
9099 will be added by s390_frame_info because for va_args always an even
9100 number of gprs has to be saved r15-r2 = 14 regs. */
9101 sav_ofs
= 2 * UNITS_PER_LONG
;
9103 if (size
< UNITS_PER_LONG
)
9104 sav_ofs
+= UNITS_PER_LONG
- size
;
9106 sav_scale
= UNITS_PER_LONG
;
9107 max_reg
= GP_ARG_NUM_REG
- n_reg
;
9110 /* Pull the value out of the saved registers ... */
9112 lab_false
= create_artificial_label (UNKNOWN_LOCATION
);
9113 lab_over
= create_artificial_label (UNKNOWN_LOCATION
);
9114 addr
= create_tmp_var (ptr_type_node
, "addr");
9116 t
= fold_convert (TREE_TYPE (reg
), size_int (max_reg
));
9117 t
= build2 (GT_EXPR
, boolean_type_node
, reg
, t
);
9118 u
= build1 (GOTO_EXPR
, void_type_node
, lab_false
);
9119 t
= build3 (COND_EXPR
, void_type_node
, t
, u
, NULL_TREE
);
9120 gimplify_and_add (t
, pre_p
);
9122 t
= build2 (POINTER_PLUS_EXPR
, ptr_type_node
, sav
,
9123 size_int (sav_ofs
));
9124 u
= build2 (MULT_EXPR
, TREE_TYPE (reg
), reg
,
9125 fold_convert (TREE_TYPE (reg
), size_int (sav_scale
)));
9126 t
= build2 (POINTER_PLUS_EXPR
, ptr_type_node
, t
, fold_convert (sizetype
, u
));
9128 gimplify_assign (addr
, t
, pre_p
);
9130 gimple_seq_add_stmt (pre_p
, gimple_build_goto (lab_over
));
9132 gimple_seq_add_stmt (pre_p
, gimple_build_label (lab_false
));
9135 /* ... Otherwise out of the overflow area. */
9138 if (size
< UNITS_PER_LONG
)
9139 t
= build2 (POINTER_PLUS_EXPR
, ptr_type_node
, t
,
9140 size_int (UNITS_PER_LONG
- size
));
9142 gimplify_expr (&t
, pre_p
, NULL
, is_gimple_val
, fb_rvalue
);
9144 gimplify_assign (addr
, t
, pre_p
);
9146 t
= build2 (POINTER_PLUS_EXPR
, ptr_type_node
, t
,
9148 gimplify_assign (ovf
, t
, pre_p
);
9150 gimple_seq_add_stmt (pre_p
, gimple_build_label (lab_over
));
9153 /* Increment register save count. */
9155 u
= build2 (PREINCREMENT_EXPR
, TREE_TYPE (reg
), reg
,
9156 fold_convert (TREE_TYPE (reg
), size_int (n_reg
)));
9157 gimplify_and_add (u
, pre_p
);
9161 t
= build_pointer_type_for_mode (build_pointer_type (type
),
9163 addr
= fold_convert (t
, addr
);
9164 addr
= build_va_arg_indirect_ref (addr
);
9168 t
= build_pointer_type_for_mode (type
, ptr_mode
, true);
9169 addr
= fold_convert (t
, addr
);
9172 return build_va_arg_indirect_ref (addr
);
9180 S390_BUILTIN_THREAD_POINTER
,
9181 S390_BUILTIN_SET_THREAD_POINTER
,
9186 static enum insn_code
const code_for_builtin_64
[S390_BUILTIN_max
] = {
9191 static enum insn_code
const code_for_builtin_31
[S390_BUILTIN_max
] = {
9197 s390_init_builtins (void)
9201 ftype
= build_function_type (ptr_type_node
, void_list_node
);
9202 add_builtin_function ("__builtin_thread_pointer", ftype
,
9203 S390_BUILTIN_THREAD_POINTER
, BUILT_IN_MD
,
9206 ftype
= build_function_type_list (void_type_node
, ptr_type_node
, NULL_TREE
);
9207 add_builtin_function ("__builtin_set_thread_pointer", ftype
,
9208 S390_BUILTIN_SET_THREAD_POINTER
, BUILT_IN_MD
,
9212 /* Expand an expression EXP that calls a built-in function,
9213 with result going to TARGET if that's convenient
9214 (and in mode MODE if that's convenient).
9215 SUBTARGET may be used as the target for computing one of EXP's operands.
9216 IGNORE is nonzero if the value is to be ignored. */
9219 s390_expand_builtin (tree exp
, rtx target
, rtx subtarget ATTRIBUTE_UNUSED
,
9220 enum machine_mode mode ATTRIBUTE_UNUSED
,
9221 int ignore ATTRIBUTE_UNUSED
)
9225 enum insn_code
const *code_for_builtin
=
9226 TARGET_64BIT
? code_for_builtin_64
: code_for_builtin_31
;
9228 tree fndecl
= TREE_OPERAND (CALL_EXPR_FN (exp
), 0);
9229 unsigned int fcode
= DECL_FUNCTION_CODE (fndecl
);
9230 enum insn_code icode
;
9231 rtx op
[MAX_ARGS
], pat
;
9235 call_expr_arg_iterator iter
;
9237 if (fcode
>= S390_BUILTIN_max
)
9238 internal_error ("bad builtin fcode");
9239 icode
= code_for_builtin
[fcode
];
9241 internal_error ("bad builtin fcode");
9243 nonvoid
= TREE_TYPE (TREE_TYPE (fndecl
)) != void_type_node
;
9246 FOR_EACH_CALL_EXPR_ARG (arg
, iter
, exp
)
9248 const struct insn_operand_data
*insn_op
;
9250 if (arg
== error_mark_node
)
9252 if (arity
> MAX_ARGS
)
9255 insn_op
= &insn_data
[icode
].operand
[arity
+ nonvoid
];
9257 op
[arity
] = expand_expr (arg
, NULL_RTX
, insn_op
->mode
, EXPAND_NORMAL
);
9259 if (!(*insn_op
->predicate
) (op
[arity
], insn_op
->mode
))
9260 op
[arity
] = copy_to_mode_reg (insn_op
->mode
, op
[arity
]);
9266 enum machine_mode tmode
= insn_data
[icode
].operand
[0].mode
;
9268 || GET_MODE (target
) != tmode
9269 || !(*insn_data
[icode
].operand
[0].predicate
) (target
, tmode
))
9270 target
= gen_reg_rtx (tmode
);
9276 pat
= GEN_FCN (icode
) (target
);
9280 pat
= GEN_FCN (icode
) (target
, op
[0]);
9282 pat
= GEN_FCN (icode
) (op
[0]);
9285 pat
= GEN_FCN (icode
) (target
, op
[0], op
[1]);
9301 /* Output assembly code for the trampoline template to
9304 On S/390, we use gpr 1 internally in the trampoline code;
9305 gpr 0 is used to hold the static chain. */
9308 s390_asm_trampoline_template (FILE *file
)
9311 op
[0] = gen_rtx_REG (Pmode
, 0);
9312 op
[1] = gen_rtx_REG (Pmode
, 1);
9316 output_asm_insn ("basr\t%1,0", op
);
9317 output_asm_insn ("lmg\t%0,%1,14(%1)", op
);
9318 output_asm_insn ("br\t%1", op
);
9319 ASM_OUTPUT_SKIP (file
, (HOST_WIDE_INT
)(TRAMPOLINE_SIZE
- 10));
9323 output_asm_insn ("basr\t%1,0", op
);
9324 output_asm_insn ("lm\t%0,%1,6(%1)", op
);
9325 output_asm_insn ("br\t%1", op
);
9326 ASM_OUTPUT_SKIP (file
, (HOST_WIDE_INT
)(TRAMPOLINE_SIZE
- 8));
9330 /* Emit RTL insns to initialize the variable parts of a trampoline.
9331 FNADDR is an RTX for the address of the function's pure code.
9332 CXT is an RTX for the static chain value for the function. */
9335 s390_trampoline_init (rtx m_tramp
, tree fndecl
, rtx cxt
)
9337 rtx fnaddr
= XEXP (DECL_RTL (fndecl
), 0);
9340 emit_block_move (m_tramp
, assemble_trampoline_template (),
9341 GEN_INT (2*UNITS_PER_WORD
), BLOCK_OP_NORMAL
);
9343 mem
= adjust_address (m_tramp
, Pmode
, 2*UNITS_PER_WORD
);
9344 emit_move_insn (mem
, cxt
);
9345 mem
= adjust_address (m_tramp
, Pmode
, 3*UNITS_PER_WORD
);
9346 emit_move_insn (mem
, fnaddr
);
9349 /* Output assembler code to FILE to increment profiler label # LABELNO
9350 for profiling a function entry. */
9353 s390_function_profiler (FILE *file
, int labelno
)
9358 ASM_GENERATE_INTERNAL_LABEL (label
, "LP", labelno
);
9360 fprintf (file
, "# function profiler \n");
9362 op
[0] = gen_rtx_REG (Pmode
, RETURN_REGNUM
);
9363 op
[1] = gen_rtx_REG (Pmode
, STACK_POINTER_REGNUM
);
9364 op
[1] = gen_rtx_MEM (Pmode
, plus_constant (op
[1], UNITS_PER_LONG
));
9366 op
[2] = gen_rtx_REG (Pmode
, 1);
9367 op
[3] = gen_rtx_SYMBOL_REF (Pmode
, label
);
9368 SYMBOL_REF_FLAGS (op
[3]) = SYMBOL_FLAG_LOCAL
;
9370 op
[4] = gen_rtx_SYMBOL_REF (Pmode
, "_mcount");
9373 op
[4] = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op
[4]), UNSPEC_PLT
);
9374 op
[4] = gen_rtx_CONST (Pmode
, op
[4]);
9379 output_asm_insn ("stg\t%0,%1", op
);
9380 output_asm_insn ("larl\t%2,%3", op
);
9381 output_asm_insn ("brasl\t%0,%4", op
);
9382 output_asm_insn ("lg\t%0,%1", op
);
9386 op
[6] = gen_label_rtx ();
9388 output_asm_insn ("st\t%0,%1", op
);
9389 output_asm_insn ("bras\t%2,%l6", op
);
9390 output_asm_insn (".long\t%4", op
);
9391 output_asm_insn (".long\t%3", op
);
9392 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[6]));
9393 output_asm_insn ("l\t%0,0(%2)", op
);
9394 output_asm_insn ("l\t%2,4(%2)", op
);
9395 output_asm_insn ("basr\t%0,%0", op
);
9396 output_asm_insn ("l\t%0,%1", op
);
9400 op
[5] = gen_label_rtx ();
9401 op
[6] = gen_label_rtx ();
9403 output_asm_insn ("st\t%0,%1", op
);
9404 output_asm_insn ("bras\t%2,%l6", op
);
9405 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[5]));
9406 output_asm_insn (".long\t%4-%l5", op
);
9407 output_asm_insn (".long\t%3-%l5", op
);
9408 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[6]));
9409 output_asm_insn ("lr\t%0,%2", op
);
9410 output_asm_insn ("a\t%0,0(%2)", op
);
9411 output_asm_insn ("a\t%2,4(%2)", op
);
9412 output_asm_insn ("basr\t%0,%0", op
);
9413 output_asm_insn ("l\t%0,%1", op
);
9417 /* Encode symbol attributes (local vs. global, tls model) of a SYMBOL_REF
9418 into its SYMBOL_REF_FLAGS. */
9421 s390_encode_section_info (tree decl
, rtx rtl
, int first
)
9423 default_encode_section_info (decl
, rtl
, first
);
9425 if (TREE_CODE (decl
) == VAR_DECL
)
9427 /* If a variable has a forced alignment to < 2 bytes, mark it
9428 with SYMBOL_FLAG_ALIGN1 to prevent it from being used as LARL
9430 if (DECL_USER_ALIGN (decl
) && DECL_ALIGN (decl
) < 16)
9431 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_ALIGN1
;
9432 if (!DECL_SIZE (decl
)
9433 || !DECL_ALIGN (decl
)
9434 || !host_integerp (DECL_SIZE (decl
), 0)
9435 || (DECL_ALIGN (decl
) <= 64
9436 && DECL_ALIGN (decl
) != tree_low_cst (DECL_SIZE (decl
), 0)))
9437 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_NOT_NATURALLY_ALIGNED
;
9440 /* Literal pool references don't have a decl so they are handled
9441 differently here. We rely on the information in the MEM_ALIGN
9442 entry to decide upon natural alignment. */
9444 && GET_CODE (XEXP (rtl
, 0)) == SYMBOL_REF
9445 && TREE_CONSTANT_POOL_ADDRESS_P (XEXP (rtl
, 0))
9446 && (MEM_ALIGN (rtl
) == 0
9447 || GET_MODE_BITSIZE (GET_MODE (rtl
)) == 0
9448 || MEM_ALIGN (rtl
) < GET_MODE_BITSIZE (GET_MODE (rtl
))))
9449 SYMBOL_REF_FLAGS (XEXP (rtl
, 0)) |= SYMBOL_FLAG_NOT_NATURALLY_ALIGNED
;
9452 /* Output thunk to FILE that implements a C++ virtual function call (with
9453 multiple inheritance) to FUNCTION. The thunk adjusts the this pointer
9454 by DELTA, and unless VCALL_OFFSET is zero, applies an additional adjustment
9455 stored at VCALL_OFFSET in the vtable whose address is located at offset 0
9456 relative to the resulting this pointer. */
9459 s390_output_mi_thunk (FILE *file
, tree thunk ATTRIBUTE_UNUSED
,
9460 HOST_WIDE_INT delta
, HOST_WIDE_INT vcall_offset
,
9466 /* Make sure unwind info is emitted for the thunk if needed. */
9467 final_start_function (emit_barrier (), file
, 1);
9469 /* Operand 0 is the target function. */
9470 op
[0] = XEXP (DECL_RTL (function
), 0);
9471 if (flag_pic
&& !SYMBOL_REF_LOCAL_P (op
[0]))
9474 op
[0] = gen_rtx_UNSPEC (Pmode
, gen_rtvec (1, op
[0]),
9475 TARGET_64BIT
? UNSPEC_PLT
: UNSPEC_GOT
);
9476 op
[0] = gen_rtx_CONST (Pmode
, op
[0]);
9479 /* Operand 1 is the 'this' pointer. */
9480 if (aggregate_value_p (TREE_TYPE (TREE_TYPE (function
)), function
))
9481 op
[1] = gen_rtx_REG (Pmode
, 3);
9483 op
[1] = gen_rtx_REG (Pmode
, 2);
9485 /* Operand 2 is the delta. */
9486 op
[2] = GEN_INT (delta
);
9488 /* Operand 3 is the vcall_offset. */
9489 op
[3] = GEN_INT (vcall_offset
);
9491 /* Operand 4 is the temporary register. */
9492 op
[4] = gen_rtx_REG (Pmode
, 1);
9494 /* Operands 5 to 8 can be used as labels. */
9500 /* Operand 9 can be used for temporary register. */
9503 /* Generate code. */
9506 /* Setup literal pool pointer if required. */
9507 if ((!DISP_IN_RANGE (delta
)
9508 && !CONST_OK_FOR_K (delta
)
9509 && !CONST_OK_FOR_Os (delta
))
9510 || (!DISP_IN_RANGE (vcall_offset
)
9511 && !CONST_OK_FOR_K (vcall_offset
)
9512 && !CONST_OK_FOR_Os (vcall_offset
)))
9514 op
[5] = gen_label_rtx ();
9515 output_asm_insn ("larl\t%4,%5", op
);
9518 /* Add DELTA to this pointer. */
9521 if (CONST_OK_FOR_J (delta
))
9522 output_asm_insn ("la\t%1,%2(%1)", op
);
9523 else if (DISP_IN_RANGE (delta
))
9524 output_asm_insn ("lay\t%1,%2(%1)", op
);
9525 else if (CONST_OK_FOR_K (delta
))
9526 output_asm_insn ("aghi\t%1,%2", op
);
9527 else if (CONST_OK_FOR_Os (delta
))
9528 output_asm_insn ("agfi\t%1,%2", op
);
9531 op
[6] = gen_label_rtx ();
9532 output_asm_insn ("agf\t%1,%6-%5(%4)", op
);
9536 /* Perform vcall adjustment. */
9539 if (DISP_IN_RANGE (vcall_offset
))
9541 output_asm_insn ("lg\t%4,0(%1)", op
);
9542 output_asm_insn ("ag\t%1,%3(%4)", op
);
9544 else if (CONST_OK_FOR_K (vcall_offset
))
9546 output_asm_insn ("lghi\t%4,%3", op
);
9547 output_asm_insn ("ag\t%4,0(%1)", op
);
9548 output_asm_insn ("ag\t%1,0(%4)", op
);
9550 else if (CONST_OK_FOR_Os (vcall_offset
))
9552 output_asm_insn ("lgfi\t%4,%3", op
);
9553 output_asm_insn ("ag\t%4,0(%1)", op
);
9554 output_asm_insn ("ag\t%1,0(%4)", op
);
9558 op
[7] = gen_label_rtx ();
9559 output_asm_insn ("llgf\t%4,%7-%5(%4)", op
);
9560 output_asm_insn ("ag\t%4,0(%1)", op
);
9561 output_asm_insn ("ag\t%1,0(%4)", op
);
9565 /* Jump to target. */
9566 output_asm_insn ("jg\t%0", op
);
9568 /* Output literal pool if required. */
9571 output_asm_insn (".align\t4", op
);
9572 targetm
.asm_out
.internal_label (file
, "L",
9573 CODE_LABEL_NUMBER (op
[5]));
9577 targetm
.asm_out
.internal_label (file
, "L",
9578 CODE_LABEL_NUMBER (op
[6]));
9579 output_asm_insn (".long\t%2", op
);
9583 targetm
.asm_out
.internal_label (file
, "L",
9584 CODE_LABEL_NUMBER (op
[7]));
9585 output_asm_insn (".long\t%3", op
);
9590 /* Setup base pointer if required. */
9592 || (!DISP_IN_RANGE (delta
)
9593 && !CONST_OK_FOR_K (delta
)
9594 && !CONST_OK_FOR_Os (delta
))
9595 || (!DISP_IN_RANGE (delta
)
9596 && !CONST_OK_FOR_K (vcall_offset
)
9597 && !CONST_OK_FOR_Os (vcall_offset
)))
9599 op
[5] = gen_label_rtx ();
9600 output_asm_insn ("basr\t%4,0", op
);
9601 targetm
.asm_out
.internal_label (file
, "L",
9602 CODE_LABEL_NUMBER (op
[5]));
9605 /* Add DELTA to this pointer. */
9608 if (CONST_OK_FOR_J (delta
))
9609 output_asm_insn ("la\t%1,%2(%1)", op
);
9610 else if (DISP_IN_RANGE (delta
))
9611 output_asm_insn ("lay\t%1,%2(%1)", op
);
9612 else if (CONST_OK_FOR_K (delta
))
9613 output_asm_insn ("ahi\t%1,%2", op
);
9614 else if (CONST_OK_FOR_Os (delta
))
9615 output_asm_insn ("afi\t%1,%2", op
);
9618 op
[6] = gen_label_rtx ();
9619 output_asm_insn ("a\t%1,%6-%5(%4)", op
);
9623 /* Perform vcall adjustment. */
9626 if (CONST_OK_FOR_J (vcall_offset
))
9628 output_asm_insn ("l\t%4,0(%1)", op
);
9629 output_asm_insn ("a\t%1,%3(%4)", op
);
9631 else if (DISP_IN_RANGE (vcall_offset
))
9633 output_asm_insn ("l\t%4,0(%1)", op
);
9634 output_asm_insn ("ay\t%1,%3(%4)", op
);
9636 else if (CONST_OK_FOR_K (vcall_offset
))
9638 output_asm_insn ("lhi\t%4,%3", op
);
9639 output_asm_insn ("a\t%4,0(%1)", op
);
9640 output_asm_insn ("a\t%1,0(%4)", op
);
9642 else if (CONST_OK_FOR_Os (vcall_offset
))
9644 output_asm_insn ("iilf\t%4,%3", op
);
9645 output_asm_insn ("a\t%4,0(%1)", op
);
9646 output_asm_insn ("a\t%1,0(%4)", op
);
9650 op
[7] = gen_label_rtx ();
9651 output_asm_insn ("l\t%4,%7-%5(%4)", op
);
9652 output_asm_insn ("a\t%4,0(%1)", op
);
9653 output_asm_insn ("a\t%1,0(%4)", op
);
9656 /* We had to clobber the base pointer register.
9657 Re-setup the base pointer (with a different base). */
9658 op
[5] = gen_label_rtx ();
9659 output_asm_insn ("basr\t%4,0", op
);
9660 targetm
.asm_out
.internal_label (file
, "L",
9661 CODE_LABEL_NUMBER (op
[5]));
9664 /* Jump to target. */
9665 op
[8] = gen_label_rtx ();
9668 output_asm_insn ("l\t%4,%8-%5(%4)", op
);
9670 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9671 /* We cannot call through .plt, since .plt requires %r12 loaded. */
9672 else if (flag_pic
== 1)
9674 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9675 output_asm_insn ("l\t%4,%0(%4)", op
);
9677 else if (flag_pic
== 2)
9679 op
[9] = gen_rtx_REG (Pmode
, 0);
9680 output_asm_insn ("l\t%9,%8-4-%5(%4)", op
);
9681 output_asm_insn ("a\t%4,%8-%5(%4)", op
);
9682 output_asm_insn ("ar\t%4,%9", op
);
9683 output_asm_insn ("l\t%4,0(%4)", op
);
9686 output_asm_insn ("br\t%4", op
);
9688 /* Output literal pool. */
9689 output_asm_insn (".align\t4", op
);
9691 if (nonlocal
&& flag_pic
== 2)
9692 output_asm_insn (".long\t%0", op
);
9695 op
[0] = gen_rtx_SYMBOL_REF (Pmode
, "_GLOBAL_OFFSET_TABLE_");
9696 SYMBOL_REF_FLAGS (op
[0]) = SYMBOL_FLAG_LOCAL
;
9699 targetm
.asm_out
.internal_label (file
, "L", CODE_LABEL_NUMBER (op
[8]));
9701 output_asm_insn (".long\t%0", op
);
9703 output_asm_insn (".long\t%0-%5", op
);
9707 targetm
.asm_out
.internal_label (file
, "L",
9708 CODE_LABEL_NUMBER (op
[6]));
9709 output_asm_insn (".long\t%2", op
);
9713 targetm
.asm_out
.internal_label (file
, "L",
9714 CODE_LABEL_NUMBER (op
[7]));
9715 output_asm_insn (".long\t%3", op
);
9718 final_end_function ();
9722 s390_valid_pointer_mode (enum machine_mode mode
)
9724 return (mode
== SImode
|| (TARGET_64BIT
&& mode
== DImode
));
9727 /* Checks whether the given CALL_EXPR would use a caller
9728 saved register. This is used to decide whether sibling call
9729 optimization could be performed on the respective function
9733 s390_call_saved_register_used (tree call_expr
)
9735 CUMULATIVE_ARGS cum
;
9737 enum machine_mode mode
;
9742 INIT_CUMULATIVE_ARGS (cum
, NULL
, NULL
, 0, 0);
9744 for (i
= 0; i
< call_expr_nargs (call_expr
); i
++)
9746 parameter
= CALL_EXPR_ARG (call_expr
, i
);
9747 gcc_assert (parameter
);
9749 /* For an undeclared variable passed as parameter we will get
9750 an ERROR_MARK node here. */
9751 if (TREE_CODE (parameter
) == ERROR_MARK
)
9754 type
= TREE_TYPE (parameter
);
9757 mode
= TYPE_MODE (type
);
9760 if (pass_by_reference (&cum
, mode
, type
, true))
9763 type
= build_pointer_type (type
);
9766 parm_rtx
= s390_function_arg (&cum
, mode
, type
, 0);
9768 s390_function_arg_advance (&cum
, mode
, type
, 0);
9773 if (REG_P (parm_rtx
))
9776 reg
< HARD_REGNO_NREGS (REGNO (parm_rtx
), GET_MODE (parm_rtx
));
9778 if (!call_used_regs
[reg
+ REGNO (parm_rtx
)])
9782 if (GET_CODE (parm_rtx
) == PARALLEL
)
9786 for (i
= 0; i
< XVECLEN (parm_rtx
, 0); i
++)
9788 rtx r
= XEXP (XVECEXP (parm_rtx
, 0, i
), 0);
9790 gcc_assert (REG_P (r
));
9793 reg
< HARD_REGNO_NREGS (REGNO (r
), GET_MODE (r
));
9795 if (!call_used_regs
[reg
+ REGNO (r
)])
9804 /* Return true if the given call expression can be
9805 turned into a sibling call.
9806 DECL holds the declaration of the function to be called whereas
9807 EXP is the call expression itself. */
9810 s390_function_ok_for_sibcall (tree decl
, tree exp
)
9812 /* The TPF epilogue uses register 1. */
9813 if (TARGET_TPF_PROFILING
)
9816 /* The 31 bit PLT code uses register 12 (GOT pointer - caller saved)
9817 which would have to be restored before the sibcall. */
9818 if (!TARGET_64BIT
&& flag_pic
&& decl
&& !targetm
.binds_local_p (decl
))
9821 /* Register 6 on s390 is available as an argument register but unfortunately
9822 "caller saved". This makes functions needing this register for arguments
9823 not suitable for sibcalls. */
9824 return !s390_call_saved_register_used (exp
);
9827 /* Return the fixed registers used for condition codes. */
9830 s390_fixed_condition_code_regs (unsigned int *p1
, unsigned int *p2
)
9833 *p2
= INVALID_REGNUM
;
9838 /* This function is used by the call expanders of the machine description.
9839 It emits the call insn itself together with the necessary operations
9840 to adjust the target address and returns the emitted insn.
9841 ADDR_LOCATION is the target address rtx
9842 TLS_CALL the location of the thread-local symbol
9843 RESULT_REG the register where the result of the call should be stored
9844 RETADDR_REG the register where the return address should be stored
9845 If this parameter is NULL_RTX the call is considered
9846 to be a sibling call. */
9849 s390_emit_call (rtx addr_location
, rtx tls_call
, rtx result_reg
,
9852 bool plt_call
= false;
9858 /* Direct function calls need special treatment. */
9859 if (GET_CODE (addr_location
) == SYMBOL_REF
)
9861 /* When calling a global routine in PIC mode, we must
9862 replace the symbol itself with the PLT stub. */
9863 if (flag_pic
&& !SYMBOL_REF_LOCAL_P (addr_location
))
9865 if (retaddr_reg
!= NULL_RTX
)
9867 addr_location
= gen_rtx_UNSPEC (Pmode
,
9868 gen_rtvec (1, addr_location
),
9870 addr_location
= gen_rtx_CONST (Pmode
, addr_location
);
9874 /* For -fpic code the PLT entries might use r12 which is
9875 call-saved. Therefore we cannot do a sibcall when
9876 calling directly using a symbol ref. When reaching
9877 this point we decided (in s390_function_ok_for_sibcall)
9878 to do a sibcall for a function pointer but one of the
9879 optimizers was able to get rid of the function pointer
9880 by propagating the symbol ref into the call. This
9881 optimization is illegal for S/390 so we turn the direct
9882 call into a indirect call again. */
9883 addr_location
= force_reg (Pmode
, addr_location
);
9886 /* Unless we can use the bras(l) insn, force the
9887 routine address into a register. */
9888 if (!TARGET_SMALL_EXEC
&& !TARGET_CPU_ZARCH
)
9891 addr_location
= legitimize_pic_address (addr_location
, 0);
9893 addr_location
= force_reg (Pmode
, addr_location
);
9897 /* If it is already an indirect call or the code above moved the
9898 SYMBOL_REF to somewhere else make sure the address can be found in
9900 if (retaddr_reg
== NULL_RTX
9901 && GET_CODE (addr_location
) != SYMBOL_REF
9904 emit_move_insn (gen_rtx_REG (Pmode
, SIBCALL_REGNUM
), addr_location
);
9905 addr_location
= gen_rtx_REG (Pmode
, SIBCALL_REGNUM
);
9908 addr_location
= gen_rtx_MEM (QImode
, addr_location
);
9909 call
= gen_rtx_CALL (VOIDmode
, addr_location
, const0_rtx
);
9911 if (result_reg
!= NULL_RTX
)
9912 call
= gen_rtx_SET (VOIDmode
, result_reg
, call
);
9914 if (retaddr_reg
!= NULL_RTX
)
9916 clobber
= gen_rtx_CLOBBER (VOIDmode
, retaddr_reg
);
9918 if (tls_call
!= NULL_RTX
)
9919 vec
= gen_rtvec (3, call
, clobber
,
9920 gen_rtx_USE (VOIDmode
, tls_call
));
9922 vec
= gen_rtvec (2, call
, clobber
);
9924 call
= gen_rtx_PARALLEL (VOIDmode
, vec
);
9927 insn
= emit_call_insn (call
);
9929 /* 31-bit PLT stubs and tls calls use the GOT register implicitly. */
9930 if ((!TARGET_64BIT
&& plt_call
) || tls_call
!= NULL_RTX
)
9932 /* s390_function_ok_for_sibcall should
9933 have denied sibcalls in this case. */
9934 gcc_assert (retaddr_reg
!= NULL_RTX
);
9936 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), pic_offset_table_rtx
);
9941 /* Implement TARGET_CONDITIONAL_REGISTER_USAGE. */
9944 s390_conditional_register_usage (void)
9950 fixed_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
9951 call_used_regs
[PIC_OFFSET_TABLE_REGNUM
] = 1;
9953 if (TARGET_CPU_ZARCH
)
9955 fixed_regs
[BASE_REGNUM
] = 0;
9956 call_used_regs
[BASE_REGNUM
] = 0;
9957 fixed_regs
[RETURN_REGNUM
] = 0;
9958 call_used_regs
[RETURN_REGNUM
] = 0;
9962 for (i
= 24; i
< 32; i
++)
9963 call_used_regs
[i
] = call_really_used_regs
[i
] = 0;
9967 for (i
= 18; i
< 20; i
++)
9968 call_used_regs
[i
] = call_really_used_regs
[i
] = 0;
9971 if (TARGET_SOFT_FLOAT
)
9973 for (i
= 16; i
< 32; i
++)
9974 call_used_regs
[i
] = fixed_regs
[i
] = 1;
9978 /* Corresponding function to eh_return expander. */
9980 static GTY(()) rtx s390_tpf_eh_return_symbol
;
9982 s390_emit_tpf_eh_return (rtx target
)
9986 if (!s390_tpf_eh_return_symbol
)
9987 s390_tpf_eh_return_symbol
= gen_rtx_SYMBOL_REF (Pmode
, "__tpf_eh_return");
9989 reg
= gen_rtx_REG (Pmode
, 2);
9991 emit_move_insn (reg
, target
);
9992 insn
= s390_emit_call (s390_tpf_eh_return_symbol
, NULL_RTX
, reg
,
9993 gen_rtx_REG (Pmode
, RETURN_REGNUM
));
9994 use_reg (&CALL_INSN_FUNCTION_USAGE (insn
), reg
);
9996 emit_move_insn (EH_RETURN_HANDLER_RTX
, reg
);
9999 /* Rework the prologue/epilogue to avoid saving/restoring
10000 registers unnecessarily. */
10003 s390_optimize_prologue (void)
10005 rtx insn
, new_insn
, next_insn
;
10007 /* Do a final recompute of the frame-related data. */
10009 s390_update_frame_layout ();
10011 /* If all special registers are in fact used, there's nothing we
10012 can do, so no point in walking the insn list. */
10014 if (cfun_frame_layout
.first_save_gpr
<= BASE_REGNUM
10015 && cfun_frame_layout
.last_save_gpr
>= BASE_REGNUM
10016 && (TARGET_CPU_ZARCH
10017 || (cfun_frame_layout
.first_save_gpr
<= RETURN_REGNUM
10018 && cfun_frame_layout
.last_save_gpr
>= RETURN_REGNUM
)))
10021 /* Search for prologue/epilogue insns and replace them. */
10023 for (insn
= get_insns (); insn
; insn
= next_insn
)
10025 int first
, last
, off
;
10026 rtx set
, base
, offset
;
10028 next_insn
= NEXT_INSN (insn
);
10030 if (GET_CODE (insn
) != INSN
)
10033 if (GET_CODE (PATTERN (insn
)) == PARALLEL
10034 && store_multiple_operation (PATTERN (insn
), VOIDmode
))
10036 set
= XVECEXP (PATTERN (insn
), 0, 0);
10037 first
= REGNO (SET_SRC (set
));
10038 last
= first
+ XVECLEN (PATTERN (insn
), 0) - 1;
10039 offset
= const0_rtx
;
10040 base
= eliminate_constant_term (XEXP (SET_DEST (set
), 0), &offset
);
10041 off
= INTVAL (offset
);
10043 if (GET_CODE (base
) != REG
|| off
< 0)
10045 if (cfun_frame_layout
.first_save_gpr
!= -1
10046 && (cfun_frame_layout
.first_save_gpr
< first
10047 || cfun_frame_layout
.last_save_gpr
> last
))
10049 if (REGNO (base
) != STACK_POINTER_REGNUM
10050 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10052 if (first
> BASE_REGNUM
|| last
< BASE_REGNUM
)
10055 if (cfun_frame_layout
.first_save_gpr
!= -1)
10057 new_insn
= save_gprs (base
,
10058 off
+ (cfun_frame_layout
.first_save_gpr
10059 - first
) * UNITS_PER_LONG
,
10060 cfun_frame_layout
.first_save_gpr
,
10061 cfun_frame_layout
.last_save_gpr
);
10062 new_insn
= emit_insn_before (new_insn
, insn
);
10063 INSN_ADDRESSES_NEW (new_insn
, -1);
10066 remove_insn (insn
);
10070 if (cfun_frame_layout
.first_save_gpr
== -1
10071 && GET_CODE (PATTERN (insn
)) == SET
10072 && GET_CODE (SET_SRC (PATTERN (insn
))) == REG
10073 && (REGNO (SET_SRC (PATTERN (insn
))) == BASE_REGNUM
10074 || (!TARGET_CPU_ZARCH
10075 && REGNO (SET_SRC (PATTERN (insn
))) == RETURN_REGNUM
))
10076 && GET_CODE (SET_DEST (PATTERN (insn
))) == MEM
)
10078 set
= PATTERN (insn
);
10079 first
= REGNO (SET_SRC (set
));
10080 offset
= const0_rtx
;
10081 base
= eliminate_constant_term (XEXP (SET_DEST (set
), 0), &offset
);
10082 off
= INTVAL (offset
);
10084 if (GET_CODE (base
) != REG
|| off
< 0)
10086 if (REGNO (base
) != STACK_POINTER_REGNUM
10087 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10090 remove_insn (insn
);
10094 if (GET_CODE (PATTERN (insn
)) == PARALLEL
10095 && load_multiple_operation (PATTERN (insn
), VOIDmode
))
10097 set
= XVECEXP (PATTERN (insn
), 0, 0);
10098 first
= REGNO (SET_DEST (set
));
10099 last
= first
+ XVECLEN (PATTERN (insn
), 0) - 1;
10100 offset
= const0_rtx
;
10101 base
= eliminate_constant_term (XEXP (SET_SRC (set
), 0), &offset
);
10102 off
= INTVAL (offset
);
10104 if (GET_CODE (base
) != REG
|| off
< 0)
10106 if (cfun_frame_layout
.first_restore_gpr
!= -1
10107 && (cfun_frame_layout
.first_restore_gpr
< first
10108 || cfun_frame_layout
.last_restore_gpr
> last
))
10110 if (REGNO (base
) != STACK_POINTER_REGNUM
10111 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10113 if (first
> BASE_REGNUM
|| last
< BASE_REGNUM
)
10116 if (cfun_frame_layout
.first_restore_gpr
!= -1)
10118 new_insn
= restore_gprs (base
,
10119 off
+ (cfun_frame_layout
.first_restore_gpr
10120 - first
) * UNITS_PER_LONG
,
10121 cfun_frame_layout
.first_restore_gpr
,
10122 cfun_frame_layout
.last_restore_gpr
);
10123 new_insn
= emit_insn_before (new_insn
, insn
);
10124 INSN_ADDRESSES_NEW (new_insn
, -1);
10127 remove_insn (insn
);
10131 if (cfun_frame_layout
.first_restore_gpr
== -1
10132 && GET_CODE (PATTERN (insn
)) == SET
10133 && GET_CODE (SET_DEST (PATTERN (insn
))) == REG
10134 && (REGNO (SET_DEST (PATTERN (insn
))) == BASE_REGNUM
10135 || (!TARGET_CPU_ZARCH
10136 && REGNO (SET_DEST (PATTERN (insn
))) == RETURN_REGNUM
))
10137 && GET_CODE (SET_SRC (PATTERN (insn
))) == MEM
)
10139 set
= PATTERN (insn
);
10140 first
= REGNO (SET_DEST (set
));
10141 offset
= const0_rtx
;
10142 base
= eliminate_constant_term (XEXP (SET_SRC (set
), 0), &offset
);
10143 off
= INTVAL (offset
);
10145 if (GET_CODE (base
) != REG
|| off
< 0)
10147 if (REGNO (base
) != STACK_POINTER_REGNUM
10148 && REGNO (base
) != HARD_FRAME_POINTER_REGNUM
)
10151 remove_insn (insn
);
10157 /* On z10 and later the dynamic branch prediction must see the
10158 backward jump within a certain windows. If not it falls back to
10159 the static prediction. This function rearranges the loop backward
10160 branch in a way which makes the static prediction always correct.
10161 The function returns true if it added an instruction. */
10163 s390_fix_long_loop_prediction (rtx insn
)
10165 rtx set
= single_set (insn
);
10166 rtx code_label
, label_ref
, new_label
;
10172 /* This will exclude branch on count and branch on index patterns
10173 since these are correctly statically predicted. */
10175 || SET_DEST (set
) != pc_rtx
10176 || GET_CODE (SET_SRC(set
)) != IF_THEN_ELSE
)
10179 label_ref
= (GET_CODE (XEXP (SET_SRC (set
), 1)) == LABEL_REF
?
10180 XEXP (SET_SRC (set
), 1) : XEXP (SET_SRC (set
), 2));
10182 gcc_assert (GET_CODE (label_ref
) == LABEL_REF
);
10184 code_label
= XEXP (label_ref
, 0);
10186 if (INSN_ADDRESSES (INSN_UID (code_label
)) == -1
10187 || INSN_ADDRESSES (INSN_UID (insn
)) == -1
10188 || (INSN_ADDRESSES (INSN_UID (insn
))
10189 - INSN_ADDRESSES (INSN_UID (code_label
)) < PREDICT_DISTANCE
))
10192 for (distance
= 0, cur_insn
= PREV_INSN (insn
);
10193 distance
< PREDICT_DISTANCE
- 6;
10194 distance
+= get_attr_length (cur_insn
), cur_insn
= PREV_INSN (cur_insn
))
10195 if (!cur_insn
|| JUMP_P (cur_insn
) || LABEL_P (cur_insn
))
10198 new_label
= gen_label_rtx ();
10199 uncond_jump
= emit_jump_insn_after (
10200 gen_rtx_SET (VOIDmode
, pc_rtx
,
10201 gen_rtx_LABEL_REF (VOIDmode
, code_label
)),
10203 emit_label_after (new_label
, uncond_jump
);
10205 tmp
= XEXP (SET_SRC (set
), 1);
10206 XEXP (SET_SRC (set
), 1) = XEXP (SET_SRC (set
), 2);
10207 XEXP (SET_SRC (set
), 2) = tmp
;
10208 INSN_CODE (insn
) = -1;
10210 XEXP (label_ref
, 0) = new_label
;
10211 JUMP_LABEL (insn
) = new_label
;
10212 JUMP_LABEL (uncond_jump
) = code_label
;
10217 /* Returns 1 if INSN reads the value of REG for purposes not related
10218 to addressing of memory, and 0 otherwise. */
10220 s390_non_addr_reg_read_p (rtx reg
, rtx insn
)
10222 return reg_referenced_p (reg
, PATTERN (insn
))
10223 && !reg_used_in_mem_p (REGNO (reg
), PATTERN (insn
));
10226 /* Starting from INSN find_cond_jump looks downwards in the insn
10227 stream for a single jump insn which is the last user of the
10228 condition code set in INSN. */
10230 find_cond_jump (rtx insn
)
10232 for (; insn
; insn
= NEXT_INSN (insn
))
10236 if (LABEL_P (insn
))
10239 if (!JUMP_P (insn
))
10241 if (reg_mentioned_p (gen_rtx_REG (CCmode
, CC_REGNUM
), insn
))
10246 /* This will be triggered by a return. */
10247 if (GET_CODE (PATTERN (insn
)) != SET
)
10250 gcc_assert (SET_DEST (PATTERN (insn
)) == pc_rtx
);
10251 ite
= SET_SRC (PATTERN (insn
));
10253 if (GET_CODE (ite
) != IF_THEN_ELSE
)
10256 cc
= XEXP (XEXP (ite
, 0), 0);
10257 if (!REG_P (cc
) || !CC_REGNO_P (REGNO (cc
)))
10260 if (find_reg_note (insn
, REG_DEAD
, cc
))
10268 /* Swap the condition in COND and the operands in OP0 and OP1 so that
10269 the semantics does not change. If NULL_RTX is passed as COND the
10270 function tries to find the conditional jump starting with INSN. */
10272 s390_swap_cmp (rtx cond
, rtx
*op0
, rtx
*op1
, rtx insn
)
10276 if (cond
== NULL_RTX
)
10278 rtx jump
= find_cond_jump (NEXT_INSN (insn
));
10279 jump
= jump
? single_set (jump
) : NULL_RTX
;
10281 if (jump
== NULL_RTX
)
10284 cond
= XEXP (XEXP (jump
, 1), 0);
10289 PUT_CODE (cond
, swap_condition (GET_CODE (cond
)));
10292 /* On z10, instructions of the compare-and-branch family have the
10293 property to access the register occurring as second operand with
10294 its bits complemented. If such a compare is grouped with a second
10295 instruction that accesses the same register non-complemented, and
10296 if that register's value is delivered via a bypass, then the
10297 pipeline recycles, thereby causing significant performance decline.
10298 This function locates such situations and exchanges the two
10299 operands of the compare. The function return true whenever it
10302 s390_z10_optimize_cmp (rtx insn
)
10304 rtx prev_insn
, next_insn
;
10305 bool insn_added_p
= false;
10306 rtx cond
, *op0
, *op1
;
10308 if (GET_CODE (PATTERN (insn
)) == PARALLEL
)
10310 /* Handle compare and branch and branch on count
10312 rtx pattern
= single_set (insn
);
10315 || SET_DEST (pattern
) != pc_rtx
10316 || GET_CODE (SET_SRC (pattern
)) != IF_THEN_ELSE
)
10319 cond
= XEXP (SET_SRC (pattern
), 0);
10320 op0
= &XEXP (cond
, 0);
10321 op1
= &XEXP (cond
, 1);
10323 else if (GET_CODE (PATTERN (insn
)) == SET
)
10327 /* Handle normal compare instructions. */
10328 src
= SET_SRC (PATTERN (insn
));
10329 dest
= SET_DEST (PATTERN (insn
));
10332 || !CC_REGNO_P (REGNO (dest
))
10333 || GET_CODE (src
) != COMPARE
)
10336 /* s390_swap_cmp will try to find the conditional
10337 jump when passing NULL_RTX as condition. */
10339 op0
= &XEXP (src
, 0);
10340 op1
= &XEXP (src
, 1);
10345 if (!REG_P (*op0
) || !REG_P (*op1
))
10348 if (GET_MODE_CLASS (GET_MODE (*op0
)) != MODE_INT
)
10351 /* Swap the COMPARE arguments and its mask if there is a
10352 conflicting access in the previous insn. */
10353 prev_insn
= prev_active_insn (insn
);
10354 if (prev_insn
!= NULL_RTX
&& INSN_P (prev_insn
)
10355 && reg_referenced_p (*op1
, PATTERN (prev_insn
)))
10356 s390_swap_cmp (cond
, op0
, op1
, insn
);
10358 /* Check if there is a conflict with the next insn. If there
10359 was no conflict with the previous insn, then swap the
10360 COMPARE arguments and its mask. If we already swapped
10361 the operands, or if swapping them would cause a conflict
10362 with the previous insn, issue a NOP after the COMPARE in
10363 order to separate the two instuctions. */
10364 next_insn
= next_active_insn (insn
);
10365 if (next_insn
!= NULL_RTX
&& INSN_P (next_insn
)
10366 && s390_non_addr_reg_read_p (*op1
, next_insn
))
10368 if (prev_insn
!= NULL_RTX
&& INSN_P (prev_insn
)
10369 && s390_non_addr_reg_read_p (*op0
, prev_insn
))
10371 if (REGNO (*op1
) == 0)
10372 emit_insn_after (gen_nop1 (), insn
);
10374 emit_insn_after (gen_nop (), insn
);
10375 insn_added_p
= true;
10378 s390_swap_cmp (cond
, op0
, op1
, insn
);
10380 return insn_added_p
;
10383 /* Perform machine-dependent processing. */
10388 bool pool_overflow
= false;
10390 /* Make sure all splits have been performed; splits after
10391 machine_dependent_reorg might confuse insn length counts. */
10392 split_all_insns_noflow ();
10394 /* Install the main literal pool and the associated base
10395 register load insns.
10397 In addition, there are two problematic situations we need
10400 - the literal pool might be > 4096 bytes in size, so that
10401 some of its elements cannot be directly accessed
10403 - a branch target might be > 64K away from the branch, so that
10404 it is not possible to use a PC-relative instruction.
10406 To fix those, we split the single literal pool into multiple
10407 pool chunks, reloading the pool base register at various
10408 points throughout the function to ensure it always points to
10409 the pool chunk the following code expects, and / or replace
10410 PC-relative branches by absolute branches.
10412 However, the two problems are interdependent: splitting the
10413 literal pool can move a branch further away from its target,
10414 causing the 64K limit to overflow, and on the other hand,
10415 replacing a PC-relative branch by an absolute branch means
10416 we need to put the branch target address into the literal
10417 pool, possibly causing it to overflow.
10419 So, we loop trying to fix up both problems until we manage
10420 to satisfy both conditions at the same time. Note that the
10421 loop is guaranteed to terminate as every pass of the loop
10422 strictly decreases the total number of PC-relative branches
10423 in the function. (This is not completely true as there
10424 might be branch-over-pool insns introduced by chunkify_start.
10425 Those never need to be split however.) */
10429 struct constant_pool
*pool
= NULL
;
10431 /* Collect the literal pool. */
10432 if (!pool_overflow
)
10434 pool
= s390_mainpool_start ();
10436 pool_overflow
= true;
10439 /* If literal pool overflowed, start to chunkify it. */
10441 pool
= s390_chunkify_start ();
10443 /* Split out-of-range branches. If this has created new
10444 literal pool entries, cancel current chunk list and
10445 recompute it. zSeries machines have large branch
10446 instructions, so we never need to split a branch. */
10447 if (!TARGET_CPU_ZARCH
&& s390_split_branches ())
10450 s390_chunkify_cancel (pool
);
10452 s390_mainpool_cancel (pool
);
10457 /* If we made it up to here, both conditions are satisfied.
10458 Finish up literal pool related changes. */
10460 s390_chunkify_finish (pool
);
10462 s390_mainpool_finish (pool
);
10464 /* We're done splitting branches. */
10465 cfun
->machine
->split_branches_pending_p
= false;
10469 /* Generate out-of-pool execute target insns. */
10470 if (TARGET_CPU_ZARCH
)
10472 rtx insn
, label
, target
;
10474 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
10476 label
= s390_execute_label (insn
);
10480 gcc_assert (label
!= const0_rtx
);
10482 target
= emit_label (XEXP (label
, 0));
10483 INSN_ADDRESSES_NEW (target
, -1);
10485 target
= emit_insn (s390_execute_target (insn
));
10486 INSN_ADDRESSES_NEW (target
, -1);
10490 /* Try to optimize prologue and epilogue further. */
10491 s390_optimize_prologue ();
10493 /* Walk over the insns and do some >=z10 specific changes. */
10494 if (s390_tune
== PROCESSOR_2097_Z10
10495 || s390_tune
== PROCESSOR_2817_Z196
)
10498 bool insn_added_p
= false;
10500 /* The insn lengths and addresses have to be up to date for the
10501 following manipulations. */
10502 shorten_branches (get_insns ());
10504 for (insn
= get_insns (); insn
; insn
= NEXT_INSN (insn
))
10506 if (!INSN_P (insn
) || INSN_CODE (insn
) <= 0)
10510 insn_added_p
|= s390_fix_long_loop_prediction (insn
);
10512 if ((GET_CODE (PATTERN (insn
)) == PARALLEL
10513 || GET_CODE (PATTERN (insn
)) == SET
)
10514 && s390_tune
== PROCESSOR_2097_Z10
)
10515 insn_added_p
|= s390_z10_optimize_cmp (insn
);
10518 /* Adjust branches if we added new instructions. */
10520 shorten_branches (get_insns ());
10524 /* Return true if INSN is a fp load insn writing register REGNO. */
10526 s390_fpload_toreg (rtx insn
, unsigned int regno
)
10529 enum attr_type flag
= s390_safe_attr_type (insn
);
10531 if (flag
!= TYPE_FLOADSF
&& flag
!= TYPE_FLOADDF
)
10534 set
= single_set (insn
);
10536 if (set
== NULL_RTX
)
10539 if (!REG_P (SET_DEST (set
)) || !MEM_P (SET_SRC (set
)))
10542 if (REGNO (SET_DEST (set
)) != regno
)
10548 /* This value describes the distance to be avoided between an
10549 aritmetic fp instruction and an fp load writing the same register.
10550 Z10_EARLYLOAD_DISTANCE - 1 as well as Z10_EARLYLOAD_DISTANCE + 1 is
10551 fine but the exact value has to be avoided. Otherwise the FP
10552 pipeline will throw an exception causing a major penalty. */
10553 #define Z10_EARLYLOAD_DISTANCE 7
10555 /* Rearrange the ready list in order to avoid the situation described
10556 for Z10_EARLYLOAD_DISTANCE. A problematic load instruction is
10557 moved to the very end of the ready list. */
10559 s390_z10_prevent_earlyload_conflicts (rtx
*ready
, int *nready_p
)
10561 unsigned int regno
;
10562 int nready
= *nready_p
;
10567 enum attr_type flag
;
10570 /* Skip DISTANCE - 1 active insns. */
10571 for (insn
= last_scheduled_insn
, distance
= Z10_EARLYLOAD_DISTANCE
- 1;
10572 distance
> 0 && insn
!= NULL_RTX
;
10573 distance
--, insn
= prev_active_insn (insn
))
10574 if (CALL_P (insn
) || JUMP_P (insn
))
10577 if (insn
== NULL_RTX
)
10580 set
= single_set (insn
);
10582 if (set
== NULL_RTX
|| !REG_P (SET_DEST (set
))
10583 || GET_MODE_CLASS (GET_MODE (SET_DEST (set
))) != MODE_FLOAT
)
10586 flag
= s390_safe_attr_type (insn
);
10588 if (flag
== TYPE_FLOADSF
|| flag
== TYPE_FLOADDF
)
10591 regno
= REGNO (SET_DEST (set
));
10594 while (!s390_fpload_toreg (ready
[i
], regno
) && i
> 0)
10601 memmove (&ready
[1], &ready
[0], sizeof (rtx
) * i
);
10605 /* This function is called via hook TARGET_SCHED_REORDER before
10606 issueing one insn from list READY which contains *NREADYP entries.
10607 For target z10 it reorders load instructions to avoid early load
10608 conflicts in the floating point pipeline */
10610 s390_sched_reorder (FILE *file ATTRIBUTE_UNUSED
, int verbose ATTRIBUTE_UNUSED
,
10611 rtx
*ready
, int *nreadyp
, int clock ATTRIBUTE_UNUSED
)
10613 if (s390_tune
== PROCESSOR_2097_Z10
)
10614 if (reload_completed
&& *nreadyp
> 1)
10615 s390_z10_prevent_earlyload_conflicts (ready
, nreadyp
);
10617 return s390_issue_rate ();
10620 /* This function is called via hook TARGET_SCHED_VARIABLE_ISSUE after
10621 the scheduler has issued INSN. It stores the last issued insn into
10622 last_scheduled_insn in order to make it available for
10623 s390_sched_reorder. */
10625 s390_sched_variable_issue (FILE *file ATTRIBUTE_UNUSED
,
10626 int verbose ATTRIBUTE_UNUSED
,
10627 rtx insn
, int more
)
10629 last_scheduled_insn
= insn
;
10631 if (GET_CODE (PATTERN (insn
)) != USE
10632 && GET_CODE (PATTERN (insn
)) != CLOBBER
)
10639 s390_sched_init (FILE *file ATTRIBUTE_UNUSED
,
10640 int verbose ATTRIBUTE_UNUSED
,
10641 int max_ready ATTRIBUTE_UNUSED
)
10643 last_scheduled_insn
= NULL_RTX
;
10646 /* This function checks the whole of insn X for memory references. The
10647 function always returns zero because the framework it is called
10648 from would stop recursively analyzing the insn upon a return value
10649 other than zero. The real result of this function is updating
10650 counter variable MEM_COUNT. */
10652 check_dpu (rtx
*x
, unsigned *mem_count
)
10654 if (*x
!= NULL_RTX
&& MEM_P (*x
))
10659 /* This target hook implementation for TARGET_LOOP_UNROLL_ADJUST calculates
10660 a new number struct loop *loop should be unrolled if tuned for cpus with
10661 a built-in stride prefetcher.
10662 The loop is analyzed for memory accesses by calling check_dpu for
10663 each rtx of the loop. Depending on the loop_depth and the amount of
10664 memory accesses a new number <=nunroll is returned to improve the
10665 behaviour of the hardware prefetch unit. */
10667 s390_loop_unroll_adjust (unsigned nunroll
, struct loop
*loop
)
10672 unsigned mem_count
= 0;
10674 if (s390_tune
!= PROCESSOR_2097_Z10
&& s390_tune
!= PROCESSOR_2817_Z196
)
10677 /* Count the number of memory references within the loop body. */
10678 bbs
= get_loop_body (loop
);
10679 for (i
= 0; i
< loop
->num_nodes
; i
++)
10681 for (insn
= BB_HEAD (bbs
[i
]); insn
!= BB_END (bbs
[i
]); insn
= NEXT_INSN (insn
))
10682 if (INSN_P (insn
) && INSN_CODE (insn
) != -1)
10683 for_each_rtx (&insn
, (rtx_function
) check_dpu
, &mem_count
);
10687 /* Prevent division by zero, and we do not need to adjust nunroll in this case. */
10688 if (mem_count
== 0)
10691 switch (loop_depth(loop
))
10694 return MIN (nunroll
, 28 / mem_count
);
10696 return MIN (nunroll
, 22 / mem_count
);
10698 return MIN (nunroll
, 16 / mem_count
);
10702 /* Initialize GCC target structure. */
10704 #undef TARGET_ASM_ALIGNED_HI_OP
10705 #define TARGET_ASM_ALIGNED_HI_OP "\t.word\t"
10706 #undef TARGET_ASM_ALIGNED_DI_OP
10707 #define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t"
10708 #undef TARGET_ASM_INTEGER
10709 #define TARGET_ASM_INTEGER s390_assemble_integer
10711 #undef TARGET_ASM_OPEN_PAREN
10712 #define TARGET_ASM_OPEN_PAREN ""
10714 #undef TARGET_ASM_CLOSE_PAREN
10715 #define TARGET_ASM_CLOSE_PAREN ""
10717 #undef TARGET_DEFAULT_TARGET_FLAGS
10718 #define TARGET_DEFAULT_TARGET_FLAGS (TARGET_DEFAULT)
10720 #undef TARGET_HANDLE_OPTION
10721 #define TARGET_HANDLE_OPTION s390_handle_option
10723 #undef TARGET_OPTION_OVERRIDE
10724 #define TARGET_OPTION_OVERRIDE s390_option_override
10726 #undef TARGET_OPTION_OPTIMIZATION_TABLE
10727 #define TARGET_OPTION_OPTIMIZATION_TABLE s390_option_optimization_table
10729 #undef TARGET_OPTION_INIT_STRUCT
10730 #define TARGET_OPTION_INIT_STRUCT s390_option_init_struct
10732 #undef TARGET_ENCODE_SECTION_INFO
10733 #define TARGET_ENCODE_SECTION_INFO s390_encode_section_info
10735 #undef TARGET_SCALAR_MODE_SUPPORTED_P
10736 #define TARGET_SCALAR_MODE_SUPPORTED_P s390_scalar_mode_supported_p
10739 #undef TARGET_HAVE_TLS
10740 #define TARGET_HAVE_TLS true
10742 #undef TARGET_CANNOT_FORCE_CONST_MEM
10743 #define TARGET_CANNOT_FORCE_CONST_MEM s390_cannot_force_const_mem
10745 #undef TARGET_DELEGITIMIZE_ADDRESS
10746 #define TARGET_DELEGITIMIZE_ADDRESS s390_delegitimize_address
10748 #undef TARGET_LEGITIMIZE_ADDRESS
10749 #define TARGET_LEGITIMIZE_ADDRESS s390_legitimize_address
10751 #undef TARGET_RETURN_IN_MEMORY
10752 #define TARGET_RETURN_IN_MEMORY s390_return_in_memory
10754 #undef TARGET_INIT_BUILTINS
10755 #define TARGET_INIT_BUILTINS s390_init_builtins
10756 #undef TARGET_EXPAND_BUILTIN
10757 #define TARGET_EXPAND_BUILTIN s390_expand_builtin
10759 #undef TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA
10760 #define TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA s390_output_addr_const_extra
10762 #undef TARGET_ASM_OUTPUT_MI_THUNK
10763 #define TARGET_ASM_OUTPUT_MI_THUNK s390_output_mi_thunk
10764 #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
10765 #define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true
10767 #undef TARGET_SCHED_ADJUST_PRIORITY
10768 #define TARGET_SCHED_ADJUST_PRIORITY s390_adjust_priority
10769 #undef TARGET_SCHED_ISSUE_RATE
10770 #define TARGET_SCHED_ISSUE_RATE s390_issue_rate
10771 #undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
10772 #define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD s390_first_cycle_multipass_dfa_lookahead
10774 #undef TARGET_SCHED_VARIABLE_ISSUE
10775 #define TARGET_SCHED_VARIABLE_ISSUE s390_sched_variable_issue
10776 #undef TARGET_SCHED_REORDER
10777 #define TARGET_SCHED_REORDER s390_sched_reorder
10778 #undef TARGET_SCHED_INIT
10779 #define TARGET_SCHED_INIT s390_sched_init
10781 #undef TARGET_CANNOT_COPY_INSN_P
10782 #define TARGET_CANNOT_COPY_INSN_P s390_cannot_copy_insn_p
10783 #undef TARGET_RTX_COSTS
10784 #define TARGET_RTX_COSTS s390_rtx_costs
10785 #undef TARGET_ADDRESS_COST
10786 #define TARGET_ADDRESS_COST s390_address_cost
10787 #undef TARGET_REGISTER_MOVE_COST
10788 #define TARGET_REGISTER_MOVE_COST s390_register_move_cost
10789 #undef TARGET_MEMORY_MOVE_COST
10790 #define TARGET_MEMORY_MOVE_COST s390_memory_move_cost
10792 #undef TARGET_MACHINE_DEPENDENT_REORG
10793 #define TARGET_MACHINE_DEPENDENT_REORG s390_reorg
10795 #undef TARGET_VALID_POINTER_MODE
10796 #define TARGET_VALID_POINTER_MODE s390_valid_pointer_mode
10798 #undef TARGET_BUILD_BUILTIN_VA_LIST
10799 #define TARGET_BUILD_BUILTIN_VA_LIST s390_build_builtin_va_list
10800 #undef TARGET_EXPAND_BUILTIN_VA_START
10801 #define TARGET_EXPAND_BUILTIN_VA_START s390_va_start
10802 #undef TARGET_GIMPLIFY_VA_ARG_EXPR
10803 #define TARGET_GIMPLIFY_VA_ARG_EXPR s390_gimplify_va_arg
10805 #undef TARGET_PROMOTE_FUNCTION_MODE
10806 #define TARGET_PROMOTE_FUNCTION_MODE s390_promote_function_mode
10807 #undef TARGET_PASS_BY_REFERENCE
10808 #define TARGET_PASS_BY_REFERENCE s390_pass_by_reference
10810 #undef TARGET_FUNCTION_OK_FOR_SIBCALL
10811 #define TARGET_FUNCTION_OK_FOR_SIBCALL s390_function_ok_for_sibcall
10812 #undef TARGET_FUNCTION_ARG
10813 #define TARGET_FUNCTION_ARG s390_function_arg
10814 #undef TARGET_FUNCTION_ARG_ADVANCE
10815 #define TARGET_FUNCTION_ARG_ADVANCE s390_function_arg_advance
10816 #undef TARGET_FUNCTION_VALUE
10817 #define TARGET_FUNCTION_VALUE s390_function_value
10818 #undef TARGET_LIBCALL_VALUE
10819 #define TARGET_LIBCALL_VALUE s390_libcall_value
10821 #undef TARGET_FIXED_CONDITION_CODE_REGS
10822 #define TARGET_FIXED_CONDITION_CODE_REGS s390_fixed_condition_code_regs
10824 #undef TARGET_CC_MODES_COMPATIBLE
10825 #define TARGET_CC_MODES_COMPATIBLE s390_cc_modes_compatible
10827 #undef TARGET_INVALID_WITHIN_DOLOOP
10828 #define TARGET_INVALID_WITHIN_DOLOOP hook_constcharptr_const_rtx_null
10831 #undef TARGET_ASM_OUTPUT_DWARF_DTPREL
10832 #define TARGET_ASM_OUTPUT_DWARF_DTPREL s390_output_dwarf_dtprel
10835 #ifdef TARGET_ALTERNATE_LONG_DOUBLE_MANGLING
10836 #undef TARGET_MANGLE_TYPE
10837 #define TARGET_MANGLE_TYPE s390_mangle_type
10840 #undef TARGET_SCALAR_MODE_SUPPORTED_P
10841 #define TARGET_SCALAR_MODE_SUPPORTED_P s390_scalar_mode_supported_p
10843 #undef TARGET_PREFERRED_RELOAD_CLASS
10844 #define TARGET_PREFERRED_RELOAD_CLASS s390_preferred_reload_class
10846 #undef TARGET_SECONDARY_RELOAD
10847 #define TARGET_SECONDARY_RELOAD s390_secondary_reload
10849 #undef TARGET_LIBGCC_CMP_RETURN_MODE
10850 #define TARGET_LIBGCC_CMP_RETURN_MODE s390_libgcc_cmp_return_mode
10852 #undef TARGET_LIBGCC_SHIFT_COUNT_MODE
10853 #define TARGET_LIBGCC_SHIFT_COUNT_MODE s390_libgcc_shift_count_mode
10855 #undef TARGET_LEGITIMATE_ADDRESS_P
10856 #define TARGET_LEGITIMATE_ADDRESS_P s390_legitimate_address_p
10858 #undef TARGET_CAN_ELIMINATE
10859 #define TARGET_CAN_ELIMINATE s390_can_eliminate
10861 #undef TARGET_CONDITIONAL_REGISTER_USAGE
10862 #define TARGET_CONDITIONAL_REGISTER_USAGE s390_conditional_register_usage
10864 #undef TARGET_LOOP_UNROLL_ADJUST
10865 #define TARGET_LOOP_UNROLL_ADJUST s390_loop_unroll_adjust
10867 #undef TARGET_ASM_TRAMPOLINE_TEMPLATE
10868 #define TARGET_ASM_TRAMPOLINE_TEMPLATE s390_asm_trampoline_template
10869 #undef TARGET_TRAMPOLINE_INIT
10870 #define TARGET_TRAMPOLINE_INIT s390_trampoline_init
10872 #undef TARGET_UNWIND_WORD_MODE
10873 #define TARGET_UNWIND_WORD_MODE s390_unwind_word_mode
10875 struct gcc_target targetm
= TARGET_INITIALIZER
;
10877 #include "gt-s390.h"