1/*
2 * Copyright (c) 2013 Rob Clark <robdclark@gmail.com>
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 */
23
24#ifndef IR3_H_
25#define IR3_H_
26
27#include <stdint.h>
28#include <stdbool.h>
29
30#include "compiler/shader_enums.h"
31
32#include "util/bitscan.h"
33#include "util/list.h"
34#include "util/u_debug.h"
35
36#include "instr-a3xx.h"
37
38/* low level intermediate representation of an adreno shader program */
39
40struct ir3_compiler;
41struct ir3;
42struct ir3_instruction;
43struct ir3_block;
44
45struct ir3_info {
46	uint32_t gpu_id;
47	uint16_t sizedwords;
48	uint16_t instrs_count;   /* expanded to account for rpt's */
49	/* NOTE: max_reg, etc, does not include registers not touched
50	 * by the shader (ie. vertex fetched via VFD_DECODE but not
51	 * touched by shader)
52	 */
53	int8_t   max_reg;   /* highest GPR # used by shader */
54	int8_t   max_half_reg;
55	int16_t  max_const;
56
57	/* number of sync bits: */
58	uint16_t ss, sy;
59};
60
61struct ir3_register {
62	enum {
63		IR3_REG_CONST  = 0x001,
64		IR3_REG_IMMED  = 0x002,
65		IR3_REG_HALF   = 0x004,
66		/* high registers are used for some things in compute shaders,
67		 * for example.  Seems to be for things that are global to all
68		 * threads in a wave, so possibly these are global/shared by
69		 * all the threads in the wave?
70		 */
71		IR3_REG_HIGH   = 0x008,
72		IR3_REG_RELATIV= 0x010,
73		IR3_REG_R      = 0x020,
74		/* Most instructions, it seems, can do float abs/neg but not
75		 * integer.  The CP pass needs to know what is intended (int or
76		 * float) in order to do the right thing.  For this reason the
77		 * abs/neg flags are split out into float and int variants.  In
78		 * addition, .b (bitwise) operations, the negate is actually a
79		 * bitwise not, so split that out into a new flag to make it
80		 * more clear.
81		 */
82		IR3_REG_FNEG   = 0x040,
83		IR3_REG_FABS   = 0x080,
84		IR3_REG_SNEG   = 0x100,
85		IR3_REG_SABS   = 0x200,
86		IR3_REG_BNOT   = 0x400,
87		IR3_REG_EVEN   = 0x800,
88		IR3_REG_POS_INF= 0x1000,
89		/* (ei) flag, end-input?  Set on last bary, presumably to signal
90		 * that the shader needs no more input:
91		 */
92		IR3_REG_EI     = 0x2000,
93		/* meta-flags, for intermediate stages of IR, ie.
94		 * before register assignment is done:
95		 */
96		IR3_REG_SSA    = 0x4000,   /* 'instr' is ptr to assigning instr */
97		IR3_REG_ARRAY  = 0x8000,
98
99	} flags;
100
101	bool merged : 1;    /* half-regs conflict with full regs (ie >= a6xx) */
102
103	/* normal registers:
104	 * the component is in the low two bits of the reg #, so
105	 * rN.x becomes: (N << 2) | x
106	 */
107	uint16_t num;
108	union {
109		/* immediate: */
110		int32_t  iim_val;
111		uint32_t uim_val;
112		float    fim_val;
113		/* relative: */
114		struct {
115			uint16_t id;
116			int16_t offset;
117		} array;
118	};
119
120	/* For IR3_REG_SSA, src registers contain ptr back to assigning
121	 * instruction.
122	 *
123	 * For IR3_REG_ARRAY, the pointer is back to the last dependent
124	 * array access (although the net effect is the same, it points
125	 * back to a previous instruction that we depend on).
126	 */
127	struct ir3_instruction *instr;
128
129	union {
130		/* used for cat5 instructions, but also for internal/IR level
131		 * tracking of what registers are read/written by an instruction.
132		 * wrmask may be a bad name since it is used to represent both
133		 * src and dst that touch multiple adjacent registers.
134		 */
135		unsigned wrmask;
136		/* for relative addressing, 32bits for array size is too small,
137		 * but otoh we don't need to deal with disjoint sets, so instead
138		 * use a simple size field (number of scalar components).
139		 */
140		unsigned size;
141	};
142};
143
144/*
145 * Stupid/simple growable array implementation:
146 */
147#define DECLARE_ARRAY(type, name) \
148	unsigned name ## _count, name ## _sz; \
149	type * name;
150
151#define array_insert(ctx, arr, val) do { \
152		if (arr ## _count == arr ## _sz) { \
153			arr ## _sz = MAX2(2 * arr ## _sz, 16); \
154			arr = reralloc_size(ctx, arr, arr ## _sz * sizeof(arr[0])); \
155		} \
156		arr[arr ##_count++] = val; \
157	} while (0)
158
159struct ir3_instruction {
160	struct ir3_block *block;
161	opc_t opc;
162	enum {
163		/* (sy) flag is set on first instruction, and after sample
164		 * instructions (probably just on RAW hazard).
165		 */
166		IR3_INSTR_SY    = 0x001,
167		/* (ss) flag is set on first instruction, and first instruction
168		 * to depend on the result of "long" instructions (RAW hazard):
169		 *
170		 *   rcp, rsq, log2, exp2, sin, cos, sqrt
171		 *
172		 * It seems to synchronize until all in-flight instructions are
173		 * completed, for example:
174		 *
175		 *   rsq hr1.w, hr1.w
176		 *   add.f hr2.z, (neg)hr2.z, hc0.y
177		 *   mul.f hr2.w, (neg)hr2.y, (neg)hr2.y
178		 *   rsq hr2.x, hr2.x
179		 *   (rpt1)nop
180		 *   mad.f16 hr2.w, hr2.z, hr2.z, hr2.w
181		 *   nop
182		 *   mad.f16 hr2.w, (neg)hr0.w, (neg)hr0.w, hr2.w
183		 *   (ss)(rpt2)mul.f hr1.x, (r)hr1.x, hr1.w
184		 *   (rpt2)mul.f hr0.x, (neg)(r)hr0.x, hr2.x
185		 *
186		 * The last mul.f does not have (ss) set, presumably because the
187		 * (ss) on the previous instruction does the job.
188		 *
189		 * The blob driver also seems to set it on WAR hazards, although
190		 * not really clear if this is needed or just blob compiler being
191		 * sloppy.  So far I haven't found a case where removing the (ss)
192		 * causes problems for WAR hazard, but I could just be getting
193		 * lucky:
194		 *
195		 *   rcp r1.y, r3.y
196		 *   (ss)(rpt2)mad.f32 r3.y, (r)c9.x, r1.x, (r)r3.z
197		 *
198		 */
199		IR3_INSTR_SS    = 0x002,
200		/* (jp) flag is set on jump targets:
201		 */
202		IR3_INSTR_JP    = 0x004,
203		IR3_INSTR_UL    = 0x008,
204		IR3_INSTR_3D    = 0x010,
205		IR3_INSTR_A     = 0x020,
206		IR3_INSTR_O     = 0x040,
207		IR3_INSTR_P     = 0x080,
208		IR3_INSTR_S     = 0x100,
209		IR3_INSTR_S2EN  = 0x200,
210		IR3_INSTR_G     = 0x400,
211		IR3_INSTR_SAT   = 0x800,
212		/* meta-flags, for intermediate stages of IR, ie.
213		 * before register assignment is done:
214		 */
215		IR3_INSTR_MARK  = 0x1000,
216		IR3_INSTR_UNUSED= 0x2000,
217	} flags;
218	uint8_t repeat;
219	uint8_t nop;
220#ifdef DEBUG
221	unsigned regs_max;
222#endif
223	unsigned regs_count;
224	struct ir3_register **regs;
225	union {
226		struct {
227			char inv;
228			char comp;
229			int  immed;
230			struct ir3_block *target;
231		} cat0;
232		struct {
233			type_t src_type, dst_type;
234		} cat1;
235		struct {
236			enum {
237				IR3_COND_LT = 0,
238				IR3_COND_LE = 1,
239				IR3_COND_GT = 2,
240				IR3_COND_GE = 3,
241				IR3_COND_EQ = 4,
242				IR3_COND_NE = 5,
243			} condition;
244		} cat2;
245		struct {
246			unsigned samp, tex;
247			type_t type;
248		} cat5;
249		struct {
250			type_t type;
251			int src_offset;
252			int dst_offset;
253			int iim_val : 3;      /* for ldgb/stgb, # of components */
254			unsigned d : 3;
255			bool typed : 1;
256		} cat6;
257		struct {
258			unsigned w : 1;       /* write */
259			unsigned r : 1;       /* read */
260			unsigned l : 1;       /* local */
261			unsigned g : 1;       /* global */
262		} cat7;
263		/* for meta-instructions, just used to hold extra data
264		 * before instruction scheduling, etc
265		 */
266		struct {
267			int off;              /* component/offset */
268		} fo;
269		struct {
270			struct ir3_block *block;
271		} inout;
272	};
273
274	/* transient values used during various algorithms: */
275	union {
276		/* The instruction depth is the max dependency distance to output.
277		 *
278		 * You can also think of it as the "cost", if we did any sort of
279		 * optimization for register footprint.  Ie. a value that is  just
280		 * result of moving a const to a reg would have a low cost,  so to
281		 * it could make sense to duplicate the instruction at various
282		 * points where the result is needed to reduce register footprint.
283		 */
284		unsigned depth;
285		/* When we get to the RA stage, we no longer need depth, but
286		 * we do need instruction's position/name:
287		 */
288		struct {
289			uint16_t ip;
290			uint16_t name;
291		};
292	};
293
294	/* used for per-pass extra instruction data.
295	 */
296	void *data;
297
298	int sun;            /* Sethi–Ullman number, used by sched */
299	int use_count;      /* currently just updated/used by cp */
300
301	/* Used during CP and RA stages.  For fanin and shader inputs/
302	 * outputs where we need a sequence of consecutive registers,
303	 * keep track of each src instructions left (ie 'n-1') and right
304	 * (ie 'n+1') neighbor.  The front-end must insert enough mov's
305	 * to ensure that each instruction has at most one left and at
306	 * most one right neighbor.  During the copy-propagation pass,
307	 * we only remove mov's when we can preserve this constraint.
308	 * And during the RA stage, we use the neighbor information to
309	 * allocate a block of registers in one shot.
310	 *
311	 * TODO: maybe just add something like:
312	 *   struct ir3_instruction_ref {
313	 *       struct ir3_instruction *instr;
314	 *       unsigned cnt;
315	 *   }
316	 *
317	 * Or can we get away without the refcnt stuff?  It seems like
318	 * it should be overkill..  the problem is if, potentially after
319	 * already eliminating some mov's, if you have a single mov that
320	 * needs to be grouped with it's neighbors in two different
321	 * places (ex. shader output and a fanin).
322	 */
323	struct {
324		struct ir3_instruction *left, *right;
325		uint16_t left_cnt, right_cnt;
326	} cp;
327
328	/* an instruction can reference at most one address register amongst
329	 * it's src/dst registers.  Beyond that, you need to insert mov's.
330	 *
331	 * NOTE: do not write this directly, use ir3_instr_set_address()
332	 */
333	struct ir3_instruction *address;
334
335	/* Tracking for additional dependent instructions.  Used to handle
336	 * barriers, WAR hazards for arrays/SSBOs/etc.
337	 */
338	DECLARE_ARRAY(struct ir3_instruction *, deps);
339
340	/*
341	 * From PoV of instruction scheduling, not execution (ie. ignores global/
342	 * local distinction):
343	 *                            shared  image  atomic  SSBO  everything
344	 *   barrier()/            -   R/W     R/W    R/W     R/W       X
345	 *     groupMemoryBarrier()
346	 *   memoryBarrier()       -           R/W    R/W
347	 *     (but only images declared coherent?)
348	 *   memoryBarrierAtomic() -                  R/W
349	 *   memoryBarrierBuffer() -                          R/W
350	 *   memoryBarrierImage()  -           R/W
351	 *   memoryBarrierShared() -   R/W
352	 *
353	 * TODO I think for SSBO/image/shared, in cases where we can determine
354	 * which variable is accessed, we don't need to care about accesses to
355	 * different variables (unless declared coherent??)
356	 */
357	enum {
358		IR3_BARRIER_EVERYTHING = 1 << 0,
359		IR3_BARRIER_SHARED_R   = 1 << 1,
360		IR3_BARRIER_SHARED_W   = 1 << 2,
361		IR3_BARRIER_IMAGE_R    = 1 << 3,
362		IR3_BARRIER_IMAGE_W    = 1 << 4,
363		IR3_BARRIER_BUFFER_R   = 1 << 5,
364		IR3_BARRIER_BUFFER_W   = 1 << 6,
365		IR3_BARRIER_ARRAY_R    = 1 << 7,
366		IR3_BARRIER_ARRAY_W    = 1 << 8,
367	} barrier_class, barrier_conflict;
368
369	/* Entry in ir3_block's instruction list: */
370	struct list_head node;
371
372#ifdef DEBUG
373	uint32_t serialno;
374#endif
375};
376
377static inline struct ir3_instruction *
378ir3_neighbor_first(struct ir3_instruction *instr)
379{
380	int cnt = 0;
381	while (instr->cp.left) {
382		instr = instr->cp.left;
383		if (++cnt > 0xffff) {
384			debug_assert(0);
385			break;
386		}
387	}
388	return instr;
389}
390
391static inline int ir3_neighbor_count(struct ir3_instruction *instr)
392{
393	int num = 1;
394
395	debug_assert(!instr->cp.left);
396
397	while (instr->cp.right) {
398		num++;
399		instr = instr->cp.right;
400		if (num > 0xffff) {
401			debug_assert(0);
402			break;
403		}
404	}
405
406	return num;
407}
408
409struct ir3 {
410	struct ir3_compiler *compiler;
411	gl_shader_stage type;
412
413	unsigned ninputs, noutputs;
414	struct ir3_instruction **inputs;
415	struct ir3_instruction **outputs;
416
417	/* Track bary.f (and ldlv) instructions.. this is needed in
418	 * scheduling to ensure that all varying fetches happen before
419	 * any potential kill instructions.  The hw gets grumpy if all
420	 * threads in a group are killed before the last bary.f gets
421	 * a chance to signal end of input (ei).
422	 */
423	DECLARE_ARRAY(struct ir3_instruction *, baryfs);
424
425	/* Track all indirect instructions (read and write).  To avoid
426	 * deadlock scenario where an address register gets scheduled,
427	 * but other dependent src instructions cannot be scheduled due
428	 * to dependency on a *different* address register value, the
429	 * scheduler needs to ensure that all dependencies other than
430	 * the instruction other than the address register are scheduled
431	 * before the one that writes the address register.  Having a
432	 * convenient list of instructions that reference some address
433	 * register simplifies this.
434	 */
435	DECLARE_ARRAY(struct ir3_instruction *, indirects);
436
437	/* and same for instructions that consume predicate register: */
438	DECLARE_ARRAY(struct ir3_instruction *, predicates);
439
440	/* Track texture sample instructions which need texture state
441	 * patched in (for astc-srgb workaround):
442	 */
443	DECLARE_ARRAY(struct ir3_instruction *, astc_srgb);
444
445	/* List of blocks: */
446	struct list_head block_list;
447
448	/* List of ir3_array's: */
449	struct list_head array_list;
450
451	unsigned max_sun;   /* max Sethi–Ullman number */
452
453#ifdef DEBUG
454	unsigned block_count, instr_count;
455#endif
456};
457
458struct ir3_array {
459	struct list_head node;
460	unsigned length;
461	unsigned id;
462
463	struct nir_register *r;
464
465	/* To avoid array write's from getting DCE'd, keep track of the
466	 * most recent write.  Any array access depends on the most
467	 * recent write.  This way, nothing depends on writes after the
468	 * last read.  But all the writes that happen before that have
469	 * something depending on them
470	 */
471	struct ir3_instruction *last_write;
472
473	/* extra stuff used in RA pass: */
474	unsigned base;      /* base vreg name */
475	unsigned reg;       /* base physical reg */
476	uint16_t start_ip, end_ip;
477};
478
479struct ir3_array * ir3_lookup_array(struct ir3 *ir, unsigned id);
480
481struct ir3_block {
482	struct list_head node;
483	struct ir3 *shader;
484
485	const struct nir_block *nblock;
486
487	struct list_head instr_list;  /* list of ir3_instruction */
488
489	/* each block has either one or two successors.. in case of
490	 * two successors, 'condition' decides which one to follow.
491	 * A block preceding an if/else has two successors.
492	 */
493	struct ir3_instruction *condition;
494	struct ir3_block *successors[2];
495
496	unsigned predecessors_count;
497	struct ir3_block **predecessors;
498
499	uint16_t start_ip, end_ip;
500
501	/* Track instructions which do not write a register but other-
502	 * wise must not be discarded (such as kill, stg, etc)
503	 */
504	DECLARE_ARRAY(struct ir3_instruction *, keeps);
505
506	/* used for per-pass extra block data.  Mainly used right
507	 * now in RA step to track livein/liveout.
508	 */
509	void *data;
510
511#ifdef DEBUG
512	uint32_t serialno;
513#endif
514};
515
516static inline uint32_t
517block_id(struct ir3_block *block)
518{
519#ifdef DEBUG
520	return block->serialno;
521#else
522	return (uint32_t)(unsigned long)block;
523#endif
524}
525
526struct ir3 * ir3_create(struct ir3_compiler *compiler,
527		gl_shader_stage type, unsigned nin, unsigned nout);
528void ir3_destroy(struct ir3 *shader);
529void * ir3_assemble(struct ir3 *shader,
530		struct ir3_info *info, uint32_t gpu_id);
531void * ir3_alloc(struct ir3 *shader, int sz);
532
533struct ir3_block * ir3_block_create(struct ir3 *shader);
534
535struct ir3_instruction * ir3_instr_create(struct ir3_block *block, opc_t opc);
536struct ir3_instruction * ir3_instr_create2(struct ir3_block *block,
537		opc_t opc, int nreg);
538struct ir3_instruction * ir3_instr_clone(struct ir3_instruction *instr);
539void ir3_instr_add_dep(struct ir3_instruction *instr, struct ir3_instruction *dep);
540const char *ir3_instr_name(struct ir3_instruction *instr);
541
542struct ir3_register * ir3_reg_create(struct ir3_instruction *instr,
543		int num, int flags);
544struct ir3_register * ir3_reg_clone(struct ir3 *shader,
545		struct ir3_register *reg);
546
547void ir3_instr_set_address(struct ir3_instruction *instr,
548		struct ir3_instruction *addr);
549
550static inline bool ir3_instr_check_mark(struct ir3_instruction *instr)
551{
552	if (instr->flags & IR3_INSTR_MARK)
553		return true;  /* already visited */
554	instr->flags |= IR3_INSTR_MARK;
555	return false;
556}
557
558void ir3_block_clear_mark(struct ir3_block *block);
559void ir3_clear_mark(struct ir3 *shader);
560
561unsigned ir3_count_instructions(struct ir3 *ir);
562
563static inline int ir3_instr_regno(struct ir3_instruction *instr,
564		struct ir3_register *reg)
565{
566	unsigned i;
567	for (i = 0; i < instr->regs_count; i++)
568		if (reg == instr->regs[i])
569			return i;
570	return -1;
571}
572
573
574#define MAX_ARRAYS 16
575
576/* comp:
577 *   0 - x
578 *   1 - y
579 *   2 - z
580 *   3 - w
581 */
582static inline uint32_t regid(int num, int comp)
583{
584	return (num << 2) | (comp & 0x3);
585}
586
587static inline uint32_t reg_num(struct ir3_register *reg)
588{
589	return reg->num >> 2;
590}
591
592static inline uint32_t reg_comp(struct ir3_register *reg)
593{
594	return reg->num & 0x3;
595}
596
597static inline bool is_flow(struct ir3_instruction *instr)
598{
599	return (opc_cat(instr->opc) == 0);
600}
601
602static inline bool is_kill(struct ir3_instruction *instr)
603{
604	return instr->opc == OPC_KILL;
605}
606
607static inline bool is_nop(struct ir3_instruction *instr)
608{
609	return instr->opc == OPC_NOP;
610}
611
612/* Is it a non-transformative (ie. not type changing) mov?  This can
613 * also include absneg.s/absneg.f, which for the most part can be
614 * treated as a mov (single src argument).
615 */
616static inline bool is_same_type_mov(struct ir3_instruction *instr)
617{
618	struct ir3_register *dst;
619
620	switch (instr->opc) {
621	case OPC_MOV:
622		if (instr->cat1.src_type != instr->cat1.dst_type)
623			return false;
624		break;
625	case OPC_ABSNEG_F:
626	case OPC_ABSNEG_S:
627		if (instr->flags & IR3_INSTR_SAT)
628			return false;
629		break;
630	default:
631		return false;
632	}
633
634	dst = instr->regs[0];
635
636	/* mov's that write to a0.x or p0.x are special: */
637	if (dst->num == regid(REG_P0, 0))
638		return false;
639	if (dst->num == regid(REG_A0, 0))
640		return false;
641
642	if (dst->flags & (IR3_REG_RELATIV | IR3_REG_ARRAY))
643		return false;
644
645	return true;
646}
647
648static inline bool is_alu(struct ir3_instruction *instr)
649{
650	return (1 <= opc_cat(instr->opc)) && (opc_cat(instr->opc) <= 3);
651}
652
653static inline bool is_sfu(struct ir3_instruction *instr)
654{
655	return (opc_cat(instr->opc) == 4);
656}
657
658static inline bool is_tex(struct ir3_instruction *instr)
659{
660	return (opc_cat(instr->opc) == 5);
661}
662
663static inline bool is_mem(struct ir3_instruction *instr)
664{
665	return (opc_cat(instr->opc) == 6);
666}
667
668static inline bool is_barrier(struct ir3_instruction *instr)
669{
670	return (opc_cat(instr->opc) == 7);
671}
672
673static inline bool
674is_store(struct ir3_instruction *instr)
675{
676	/* these instructions, the "destination" register is
677	 * actually a source, the address to store to.
678	 */
679	switch (instr->opc) {
680	case OPC_STG:
681	case OPC_STGB:
682	case OPC_STIB:
683	case OPC_STP:
684	case OPC_STL:
685	case OPC_STLW:
686	case OPC_L2G:
687	case OPC_G2L:
688		return true;
689	default:
690		return false;
691	}
692}
693
694static inline bool is_load(struct ir3_instruction *instr)
695{
696	switch (instr->opc) {
697	case OPC_LDG:
698	case OPC_LDGB:
699	case OPC_LDIB:
700	case OPC_LDL:
701	case OPC_LDP:
702	case OPC_L2G:
703	case OPC_LDLW:
704	case OPC_LDC:
705	case OPC_LDLV:
706		/* probably some others too.. */
707		return true;
708	default:
709		return false;
710	}
711}
712
713static inline bool is_input(struct ir3_instruction *instr)
714{
715	/* in some cases, ldlv is used to fetch varying without
716	 * interpolation.. fortunately inloc is the first src
717	 * register in either case
718	 */
719	switch (instr->opc) {
720	case OPC_LDLV:
721	case OPC_BARY_F:
722		return true;
723	default:
724		return false;
725	}
726}
727
728static inline bool is_bool(struct ir3_instruction *instr)
729{
730	switch (instr->opc) {
731	case OPC_CMPS_F:
732	case OPC_CMPS_S:
733	case OPC_CMPS_U:
734		return true;
735	default:
736		return false;
737	}
738}
739
740static inline bool is_meta(struct ir3_instruction *instr)
741{
742	/* TODO how should we count PHI (and maybe fan-in/out) which
743	 * might actually contribute some instructions to the final
744	 * result?
745	 */
746	return (opc_cat(instr->opc) == -1);
747}
748
749static inline unsigned dest_regs(struct ir3_instruction *instr)
750{
751	if ((instr->regs_count == 0) || is_store(instr))
752		return 0;
753
754	return util_last_bit(instr->regs[0]->wrmask);
755}
756
757static inline bool writes_addr(struct ir3_instruction *instr)
758{
759	if (instr->regs_count > 0) {
760		struct ir3_register *dst = instr->regs[0];
761		return reg_num(dst) == REG_A0;
762	}
763	return false;
764}
765
766static inline bool writes_pred(struct ir3_instruction *instr)
767{
768	if (instr->regs_count > 0) {
769		struct ir3_register *dst = instr->regs[0];
770		return reg_num(dst) == REG_P0;
771	}
772	return false;
773}
774
775/* returns defining instruction for reg */
776/* TODO better name */
777static inline struct ir3_instruction *ssa(struct ir3_register *reg)
778{
779	if (reg->flags & (IR3_REG_SSA | IR3_REG_ARRAY)) {
780		return reg->instr;
781	}
782	return NULL;
783}
784
785static inline bool conflicts(struct ir3_instruction *a,
786		struct ir3_instruction *b)
787{
788	return (a && b) && (a != b);
789}
790
791static inline bool reg_gpr(struct ir3_register *r)
792{
793	if (r->flags & (IR3_REG_CONST | IR3_REG_IMMED))
794		return false;
795	if ((reg_num(r) == REG_A0) || (reg_num(r) == REG_P0))
796		return false;
797	return true;
798}
799
800static inline type_t half_type(type_t type)
801{
802	switch (type) {
803	case TYPE_F32: return TYPE_F16;
804	case TYPE_U32: return TYPE_U16;
805	case TYPE_S32: return TYPE_S16;
806	case TYPE_F16:
807	case TYPE_U16:
808	case TYPE_S16:
809		return type;
810	default:
811		assert(0);
812		return ~0;
813	}
814}
815
816/* some cat2 instructions (ie. those which are not float) can embed an
817 * immediate:
818 */
819static inline bool ir3_cat2_int(opc_t opc)
820{
821	switch (opc) {
822	case OPC_ADD_U:
823	case OPC_ADD_S:
824	case OPC_SUB_U:
825	case OPC_SUB_S:
826	case OPC_CMPS_U:
827	case OPC_CMPS_S:
828	case OPC_MIN_U:
829	case OPC_MIN_S:
830	case OPC_MAX_U:
831	case OPC_MAX_S:
832	case OPC_CMPV_U:
833	case OPC_CMPV_S:
834	case OPC_MUL_U:
835	case OPC_MUL_S:
836	case OPC_MULL_U:
837	case OPC_CLZ_S:
838	case OPC_ABSNEG_S:
839	case OPC_AND_B:
840	case OPC_OR_B:
841	case OPC_NOT_B:
842	case OPC_XOR_B:
843	case OPC_BFREV_B:
844	case OPC_CLZ_B:
845	case OPC_SHL_B:
846	case OPC_SHR_B:
847	case OPC_ASHR_B:
848	case OPC_MGEN_B:
849	case OPC_GETBIT_B:
850	case OPC_CBITS_B:
851	case OPC_BARY_F:
852		return true;
853
854	default:
855		return false;
856	}
857}
858
859
860/* map cat2 instruction to valid abs/neg flags: */
861static inline unsigned ir3_cat2_absneg(opc_t opc)
862{
863	switch (opc) {
864	case OPC_ADD_F:
865	case OPC_MIN_F:
866	case OPC_MAX_F:
867	case OPC_MUL_F:
868	case OPC_SIGN_F:
869	case OPC_CMPS_F:
870	case OPC_ABSNEG_F:
871	case OPC_CMPV_F:
872	case OPC_FLOOR_F:
873	case OPC_CEIL_F:
874	case OPC_RNDNE_F:
875	case OPC_RNDAZ_F:
876	case OPC_TRUNC_F:
877	case OPC_BARY_F:
878		return IR3_REG_FABS | IR3_REG_FNEG;
879
880	case OPC_ADD_U:
881	case OPC_ADD_S:
882	case OPC_SUB_U:
883	case OPC_SUB_S:
884	case OPC_CMPS_U:
885	case OPC_CMPS_S:
886	case OPC_MIN_U:
887	case OPC_MIN_S:
888	case OPC_MAX_U:
889	case OPC_MAX_S:
890	case OPC_CMPV_U:
891	case OPC_CMPV_S:
892	case OPC_MUL_U:
893	case OPC_MUL_S:
894	case OPC_MULL_U:
895	case OPC_CLZ_S:
896		return 0;
897
898	case OPC_ABSNEG_S:
899		return IR3_REG_SABS | IR3_REG_SNEG;
900
901	case OPC_AND_B:
902	case OPC_OR_B:
903	case OPC_NOT_B:
904	case OPC_XOR_B:
905	case OPC_BFREV_B:
906	case OPC_CLZ_B:
907	case OPC_SHL_B:
908	case OPC_SHR_B:
909	case OPC_ASHR_B:
910	case OPC_MGEN_B:
911	case OPC_GETBIT_B:
912	case OPC_CBITS_B:
913		return IR3_REG_BNOT;
914
915	default:
916		return 0;
917	}
918}
919
920/* map cat3 instructions to valid abs/neg flags: */
921static inline unsigned ir3_cat3_absneg(opc_t opc)
922{
923	switch (opc) {
924	case OPC_MAD_F16:
925	case OPC_MAD_F32:
926	case OPC_SEL_F16:
927	case OPC_SEL_F32:
928		return IR3_REG_FNEG;
929
930	case OPC_MAD_U16:
931	case OPC_MADSH_U16:
932	case OPC_MAD_S16:
933	case OPC_MADSH_M16:
934	case OPC_MAD_U24:
935	case OPC_MAD_S24:
936	case OPC_SEL_S16:
937	case OPC_SEL_S32:
938	case OPC_SAD_S16:
939	case OPC_SAD_S32:
940		/* neg *may* work on 3rd src.. */
941
942	case OPC_SEL_B16:
943	case OPC_SEL_B32:
944
945	default:
946		return 0;
947	}
948}
949
950#define MASK(n) ((1 << (n)) - 1)
951
952/* iterator for an instructions's sources (reg), also returns src #: */
953#define foreach_src_n(__srcreg, __n, __instr) \
954	if ((__instr)->regs_count) \
955		for (unsigned __cnt = (__instr)->regs_count - 1, __n = 0; __n < __cnt; __n++) \
956			if ((__srcreg = (__instr)->regs[__n + 1]))
957
958/* iterator for an instructions's sources (reg): */
959#define foreach_src(__srcreg, __instr) \
960	foreach_src_n(__srcreg, __i, __instr)
961
962static inline unsigned __ssa_src_cnt(struct ir3_instruction *instr)
963{
964	unsigned cnt = instr->regs_count + instr->deps_count;
965	if (instr->address)
966		cnt++;
967	return cnt;
968}
969
970static inline struct ir3_instruction * __ssa_src_n(struct ir3_instruction *instr, unsigned n)
971{
972	if (n == (instr->regs_count + instr->deps_count))
973		return instr->address;
974	if (n >= instr->regs_count)
975		return instr->deps[n - instr->regs_count];
976	return ssa(instr->regs[n]);
977}
978
979static inline bool __is_false_dep(struct ir3_instruction *instr, unsigned n)
980{
981	if (n == (instr->regs_count + instr->deps_count))
982		return false;
983	if (n >= instr->regs_count)
984		return true;
985	return false;
986}
987
988#define __src_cnt(__instr) ((__instr)->address ? (__instr)->regs_count : (__instr)->regs_count - 1)
989
990/* iterator for an instruction's SSA sources (instr), also returns src #: */
991#define foreach_ssa_src_n(__srcinst, __n, __instr) \
992	for (unsigned __cnt = __ssa_src_cnt(__instr), __n = 0; __n < __cnt; __n++) \
993		if ((__srcinst = __ssa_src_n(__instr, __n)))
994
995/* iterator for an instruction's SSA sources (instr): */
996#define foreach_ssa_src(__srcinst, __instr) \
997	foreach_ssa_src_n(__srcinst, __i, __instr)
998
999
1000/* dump: */
1001void ir3_print(struct ir3 *ir);
1002void ir3_print_instr(struct ir3_instruction *instr);
1003
1004/* depth calculation: */
1005int ir3_delayslots(struct ir3_instruction *assigner,
1006		struct ir3_instruction *consumer, unsigned n);
1007void ir3_insert_by_depth(struct ir3_instruction *instr, struct list_head *list);
1008void ir3_depth(struct ir3 *ir);
1009
1010/* copy-propagate: */
1011struct ir3_shader_variant;
1012void ir3_cp(struct ir3 *ir, struct ir3_shader_variant *so);
1013
1014/* group neighbors and insert mov's to resolve conflicts: */
1015void ir3_group(struct ir3 *ir);
1016
1017/* Sethi–Ullman numbering: */
1018void ir3_sun(struct ir3 *ir);
1019
1020/* scheduling: */
1021void ir3_sched_add_deps(struct ir3 *ir);
1022int ir3_sched(struct ir3 *ir);
1023
1024void ir3_a6xx_fixup_atomic_dests(struct ir3 *ir, struct ir3_shader_variant *so);
1025
1026/* register assignment: */
1027struct ir3_ra_reg_set * ir3_ra_alloc_reg_set(struct ir3_compiler *compiler);
1028int ir3_ra(struct ir3 *ir3, gl_shader_stage type,
1029		bool frag_coord, bool frag_face);
1030
1031/* legalize: */
1032void ir3_legalize(struct ir3 *ir, bool *has_ssbo, bool *need_pixlod, int *max_bary);
1033
1034/* ************************************************************************* */
1035/* instruction helpers */
1036
1037static inline struct ir3_instruction *
1038create_immed_typed(struct ir3_block *block, uint32_t val, type_t type)
1039{
1040	struct ir3_instruction *mov;
1041	unsigned flags = (type_size(type) < 32) ? IR3_REG_HALF : 0;
1042
1043	mov = ir3_instr_create(block, OPC_MOV);
1044	mov->cat1.src_type = type;
1045	mov->cat1.dst_type = type;
1046	ir3_reg_create(mov, 0, flags);
1047	ir3_reg_create(mov, 0, IR3_REG_IMMED)->uim_val = val;
1048
1049	return mov;
1050}
1051
1052static inline struct ir3_instruction *
1053create_immed(struct ir3_block *block, uint32_t val)
1054{
1055	return create_immed_typed(block, val, TYPE_U32);
1056}
1057
1058static inline struct ir3_instruction *
1059create_uniform(struct ir3_block *block, unsigned n)
1060{
1061	struct ir3_instruction *mov;
1062
1063	mov = ir3_instr_create(block, OPC_MOV);
1064	/* TODO get types right? */
1065	mov->cat1.src_type = TYPE_F32;
1066	mov->cat1.dst_type = TYPE_F32;
1067	ir3_reg_create(mov, 0, 0);
1068	ir3_reg_create(mov, n, IR3_REG_CONST);
1069
1070	return mov;
1071}
1072
1073static inline struct ir3_instruction *
1074create_uniform_indirect(struct ir3_block *block, int n,
1075		struct ir3_instruction *address)
1076{
1077	struct ir3_instruction *mov;
1078
1079	mov = ir3_instr_create(block, OPC_MOV);
1080	mov->cat1.src_type = TYPE_U32;
1081	mov->cat1.dst_type = TYPE_U32;
1082	ir3_reg_create(mov, 0, 0);
1083	ir3_reg_create(mov, 0, IR3_REG_CONST | IR3_REG_RELATIV)->array.offset = n;
1084
1085	ir3_instr_set_address(mov, address);
1086
1087	return mov;
1088}
1089
1090/* creates SSA src of correct type (ie. half vs full precision) */
1091static inline struct ir3_register * __ssa_src(struct ir3_instruction *instr,
1092		struct ir3_instruction *src, unsigned flags)
1093{
1094	struct ir3_register *reg;
1095	if (src->regs[0]->flags & IR3_REG_HALF)
1096		flags |= IR3_REG_HALF;
1097	reg = ir3_reg_create(instr, 0, IR3_REG_SSA | flags);
1098	reg->instr = src;
1099	reg->wrmask = src->regs[0]->wrmask;
1100	return reg;
1101}
1102
1103static inline struct ir3_instruction *
1104ir3_MOV(struct ir3_block *block, struct ir3_instruction *src, type_t type)
1105{
1106	struct ir3_instruction *instr = ir3_instr_create(block, OPC_MOV);
1107	ir3_reg_create(instr, 0, 0);   /* dst */
1108	if (src->regs[0]->flags & IR3_REG_ARRAY) {
1109		struct ir3_register *src_reg = __ssa_src(instr, src, IR3_REG_ARRAY);
1110		src_reg->array = src->regs[0]->array;
1111	} else {
1112		__ssa_src(instr, src, src->regs[0]->flags & IR3_REG_HIGH);
1113	}
1114	debug_assert(!(src->regs[0]->flags & IR3_REG_RELATIV));
1115	instr->cat1.src_type = type;
1116	instr->cat1.dst_type = type;
1117	return instr;
1118}
1119
1120static inline struct ir3_instruction *
1121ir3_COV(struct ir3_block *block, struct ir3_instruction *src,
1122		type_t src_type, type_t dst_type)
1123{
1124	struct ir3_instruction *instr = ir3_instr_create(block, OPC_MOV);
1125	unsigned dst_flags = (type_size(dst_type) < 32) ? IR3_REG_HALF : 0;
1126	unsigned src_flags = (type_size(src_type) < 32) ? IR3_REG_HALF : 0;
1127
1128	debug_assert((src->regs[0]->flags & IR3_REG_HALF) == src_flags);
1129
1130	ir3_reg_create(instr, 0, dst_flags);   /* dst */
1131	__ssa_src(instr, src, 0);
1132	instr->cat1.src_type = src_type;
1133	instr->cat1.dst_type = dst_type;
1134	debug_assert(!(src->regs[0]->flags & IR3_REG_ARRAY));
1135	return instr;
1136}
1137
1138static inline struct ir3_instruction *
1139ir3_NOP(struct ir3_block *block)
1140{
1141	return ir3_instr_create(block, OPC_NOP);
1142}
1143
1144#define IR3_INSTR_0 0
1145
1146#define __INSTR0(flag, name, opc)                                        \
1147static inline struct ir3_instruction *                                   \
1148ir3_##name(struct ir3_block *block)                                      \
1149{                                                                        \
1150	struct ir3_instruction *instr =                                      \
1151		ir3_instr_create(block, opc);                                    \
1152	instr->flags |= flag;                                                \
1153	return instr;                                                        \
1154}
1155#define INSTR0F(f, name)    __INSTR0(IR3_INSTR_##f, name##_##f, OPC_##name)
1156#define INSTR0(name)        __INSTR0(0, name, OPC_##name)
1157
1158#define __INSTR1(flag, name, opc)                                        \
1159static inline struct ir3_instruction *                                   \
1160ir3_##name(struct ir3_block *block,                                      \
1161		struct ir3_instruction *a, unsigned aflags)                      \
1162{                                                                        \
1163	struct ir3_instruction *instr =                                      \
1164		ir3_instr_create(block, opc);                                    \
1165	ir3_reg_create(instr, 0, 0);   /* dst */                             \
1166	__ssa_src(instr, a, aflags);                                         \
1167	instr->flags |= flag;                                                \
1168	return instr;                                                        \
1169}
1170#define INSTR1F(f, name)    __INSTR1(IR3_INSTR_##f, name##_##f, OPC_##name)
1171#define INSTR1(name)        __INSTR1(0, name, OPC_##name)
1172
1173#define __INSTR2(flag, name, opc)                                        \
1174static inline struct ir3_instruction *                                   \
1175ir3_##name(struct ir3_block *block,                                      \
1176		struct ir3_instruction *a, unsigned aflags,                      \
1177		struct ir3_instruction *b, unsigned bflags)                      \
1178{                                                                        \
1179	struct ir3_instruction *instr =                                      \
1180		ir3_instr_create(block, opc);                                    \
1181	ir3_reg_create(instr, 0, 0);   /* dst */                             \
1182	__ssa_src(instr, a, aflags);                                         \
1183	__ssa_src(instr, b, bflags);                                         \
1184	instr->flags |= flag;                                                \
1185	return instr;                                                        \
1186}
1187#define INSTR2F(f, name)    __INSTR2(IR3_INSTR_##f, name##_##f, OPC_##name)
1188#define INSTR2(name)        __INSTR2(0, name, OPC_##name)
1189
1190#define __INSTR3(flag, name, opc)                                        \
1191static inline struct ir3_instruction *                                   \
1192ir3_##name(struct ir3_block *block,                                      \
1193		struct ir3_instruction *a, unsigned aflags,                      \
1194		struct ir3_instruction *b, unsigned bflags,                      \
1195		struct ir3_instruction *c, unsigned cflags)                      \
1196{                                                                        \
1197	struct ir3_instruction *instr =                                      \
1198		ir3_instr_create2(block, opc, 4);                                \
1199	ir3_reg_create(instr, 0, 0);   /* dst */                             \
1200	__ssa_src(instr, a, aflags);                                         \
1201	__ssa_src(instr, b, bflags);                                         \
1202	__ssa_src(instr, c, cflags);                                         \
1203	instr->flags |= flag;                                                \
1204	return instr;                                                        \
1205}
1206#define INSTR3F(f, name)    __INSTR3(IR3_INSTR_##f, name##_##f, OPC_##name)
1207#define INSTR3(name)        __INSTR3(0, name, OPC_##name)
1208
1209#define __INSTR4(flag, name, opc)                                        \
1210static inline struct ir3_instruction *                                   \
1211ir3_##name(struct ir3_block *block,                                      \
1212		struct ir3_instruction *a, unsigned aflags,                      \
1213		struct ir3_instruction *b, unsigned bflags,                      \
1214		struct ir3_instruction *c, unsigned cflags,                      \
1215		struct ir3_instruction *d, unsigned dflags)                      \
1216{                                                                        \
1217	struct ir3_instruction *instr =                                      \
1218		ir3_instr_create2(block, opc, 5);                                \
1219	ir3_reg_create(instr, 0, 0);   /* dst */                             \
1220	__ssa_src(instr, a, aflags);                                         \
1221	__ssa_src(instr, b, bflags);                                         \
1222	__ssa_src(instr, c, cflags);                                         \
1223	__ssa_src(instr, d, dflags);                                         \
1224	instr->flags |= flag;                                                \
1225	return instr;                                                        \
1226}
1227#define INSTR4F(f, name)    __INSTR4(IR3_INSTR_##f, name##_##f, OPC_##name)
1228#define INSTR4(name)        __INSTR4(0, name, OPC_##name)
1229
1230/* cat0 instructions: */
1231INSTR0(BR)
1232INSTR0(JUMP)
1233INSTR1(KILL)
1234INSTR0(END)
1235
1236/* cat2 instructions, most 2 src but some 1 src: */
1237INSTR2(ADD_F)
1238INSTR2(MIN_F)
1239INSTR2(MAX_F)
1240INSTR2(MUL_F)
1241INSTR1(SIGN_F)
1242INSTR2(CMPS_F)
1243INSTR1(ABSNEG_F)
1244INSTR2(CMPV_F)
1245INSTR1(FLOOR_F)
1246INSTR1(CEIL_F)
1247INSTR1(RNDNE_F)
1248INSTR1(RNDAZ_F)
1249INSTR1(TRUNC_F)
1250INSTR2(ADD_U)
1251INSTR2(ADD_S)
1252INSTR2(SUB_U)
1253INSTR2(SUB_S)
1254INSTR2(CMPS_U)
1255INSTR2(CMPS_S)
1256INSTR2(MIN_U)
1257INSTR2(MIN_S)
1258INSTR2(MAX_U)
1259INSTR2(MAX_S)
1260INSTR1(ABSNEG_S)
1261INSTR2(AND_B)
1262INSTR2(OR_B)
1263INSTR1(NOT_B)
1264INSTR2(XOR_B)
1265INSTR2(CMPV_U)
1266INSTR2(CMPV_S)
1267INSTR2(MUL_U)
1268INSTR2(MUL_S)
1269INSTR2(MULL_U)
1270INSTR1(BFREV_B)
1271INSTR1(CLZ_S)
1272INSTR1(CLZ_B)
1273INSTR2(SHL_B)
1274INSTR2(SHR_B)
1275INSTR2(ASHR_B)
1276INSTR2(BARY_F)
1277INSTR2(MGEN_B)
1278INSTR2(GETBIT_B)
1279INSTR1(SETRM)
1280INSTR1(CBITS_B)
1281INSTR2(SHB)
1282INSTR2(MSAD)
1283
1284/* cat3 instructions: */
1285INSTR3(MAD_U16)
1286INSTR3(MADSH_U16)
1287INSTR3(MAD_S16)
1288INSTR3(MADSH_M16)
1289INSTR3(MAD_U24)
1290INSTR3(MAD_S24)
1291INSTR3(MAD_F16)
1292INSTR3(MAD_F32)
1293INSTR3(SEL_B16)
1294INSTR3(SEL_B32)
1295INSTR3(SEL_S16)
1296INSTR3(SEL_S32)
1297INSTR3(SEL_F16)
1298INSTR3(SEL_F32)
1299INSTR3(SAD_S16)
1300INSTR3(SAD_S32)
1301
1302/* cat4 instructions: */
1303INSTR1(RCP)
1304INSTR1(RSQ)
1305INSTR1(LOG2)
1306INSTR1(EXP2)
1307INSTR1(SIN)
1308INSTR1(COS)
1309INSTR1(SQRT)
1310
1311/* cat5 instructions: */
1312INSTR1(DSX)
1313INSTR1(DSY)
1314INSTR1F(3D, DSX)
1315INSTR1F(3D, DSY)
1316INSTR1(RGETPOS)
1317
1318static inline struct ir3_instruction *
1319ir3_SAM(struct ir3_block *block, opc_t opc, type_t type,
1320		unsigned wrmask, unsigned flags, struct ir3_instruction *samp_tex,
1321		struct ir3_instruction *src0, struct ir3_instruction *src1)
1322{
1323	struct ir3_instruction *sam;
1324	struct ir3_register *reg;
1325
1326	sam = ir3_instr_create(block, opc);
1327	sam->flags |= flags | IR3_INSTR_S2EN;
1328	ir3_reg_create(sam, 0, 0)->wrmask = wrmask;
1329	__ssa_src(sam, samp_tex, IR3_REG_HALF);
1330	if (src0) {
1331		reg = ir3_reg_create(sam, 0, IR3_REG_SSA);
1332		reg->wrmask = (1 << (src0->regs_count - 1)) - 1;
1333		reg->instr = src0;
1334	}
1335	if (src1) {
1336		reg = ir3_reg_create(sam, 0, IR3_REG_SSA);
1337		reg->instr = src1;
1338		reg->wrmask = (1 << (src1->regs_count - 1)) - 1;
1339	}
1340	sam->cat5.type  = type;
1341
1342	return sam;
1343}
1344
1345/* cat6 instructions: */
1346INSTR2(LDLV)
1347INSTR2(LDG)
1348INSTR2(LDL)
1349INSTR3(STG)
1350INSTR3(STL)
1351INSTR1(RESINFO)
1352INSTR1(RESFMT)
1353INSTR2(ATOMIC_ADD)
1354INSTR2(ATOMIC_SUB)
1355INSTR2(ATOMIC_XCHG)
1356INSTR2(ATOMIC_INC)
1357INSTR2(ATOMIC_DEC)
1358INSTR2(ATOMIC_CMPXCHG)
1359INSTR2(ATOMIC_MIN)
1360INSTR2(ATOMIC_MAX)
1361INSTR2(ATOMIC_AND)
1362INSTR2(ATOMIC_OR)
1363INSTR2(ATOMIC_XOR)
1364#if GPU >= 600
1365INSTR3(STIB);
1366INSTR2(LDIB);
1367INSTR3F(G, ATOMIC_ADD)
1368INSTR3F(G, ATOMIC_SUB)
1369INSTR3F(G, ATOMIC_XCHG)
1370INSTR3F(G, ATOMIC_INC)
1371INSTR3F(G, ATOMIC_DEC)
1372INSTR3F(G, ATOMIC_CMPXCHG)
1373INSTR3F(G, ATOMIC_MIN)
1374INSTR3F(G, ATOMIC_MAX)
1375INSTR3F(G, ATOMIC_AND)
1376INSTR3F(G, ATOMIC_OR)
1377INSTR3F(G, ATOMIC_XOR)
1378#elif GPU >= 400
1379INSTR3(LDGB)
1380INSTR4(STGB)
1381INSTR4(STIB)
1382INSTR4F(G, ATOMIC_ADD)
1383INSTR4F(G, ATOMIC_SUB)
1384INSTR4F(G, ATOMIC_XCHG)
1385INSTR4F(G, ATOMIC_INC)
1386INSTR4F(G, ATOMIC_DEC)
1387INSTR4F(G, ATOMIC_CMPXCHG)
1388INSTR4F(G, ATOMIC_MIN)
1389INSTR4F(G, ATOMIC_MAX)
1390INSTR4F(G, ATOMIC_AND)
1391INSTR4F(G, ATOMIC_OR)
1392INSTR4F(G, ATOMIC_XOR)
1393#endif
1394
1395/* cat7 instructions: */
1396INSTR0(BAR)
1397INSTR0(FENCE)
1398
1399/* ************************************************************************* */
1400/* split this out or find some helper to use.. like main/bitset.h.. */
1401
1402#include <string.h>
1403
1404#define MAX_REG 256
1405
1406typedef uint8_t regmask_t[2 * MAX_REG / 8];
1407
1408static inline unsigned regmask_idx(struct ir3_register *reg)
1409{
1410	unsigned num = (reg->flags & IR3_REG_RELATIV) ? reg->array.offset : reg->num;
1411	debug_assert(num < MAX_REG);
1412	if (reg->flags & IR3_REG_HALF) {
1413		if (reg->merged) {
1414			num /= 2;
1415		} else {
1416			num += MAX_REG;
1417		}
1418	}
1419	return num;
1420}
1421
1422static inline void regmask_init(regmask_t *regmask)
1423{
1424	memset(regmask, 0, sizeof(*regmask));
1425}
1426
1427static inline void regmask_set(regmask_t *regmask, struct ir3_register *reg)
1428{
1429	unsigned idx = regmask_idx(reg);
1430	if (reg->flags & IR3_REG_RELATIV) {
1431		unsigned i;
1432		for (i = 0; i < reg->size; i++, idx++)
1433			(*regmask)[idx / 8] |= 1 << (idx % 8);
1434	} else {
1435		unsigned mask;
1436		for (mask = reg->wrmask; mask; mask >>= 1, idx++)
1437			if (mask & 1)
1438				(*regmask)[idx / 8] |= 1 << (idx % 8);
1439	}
1440}
1441
1442static inline void regmask_or(regmask_t *dst, regmask_t *a, regmask_t *b)
1443{
1444	unsigned i;
1445	for (i = 0; i < ARRAY_SIZE(*dst); i++)
1446		(*dst)[i] = (*a)[i] | (*b)[i];
1447}
1448
1449/* set bits in a if not set in b, conceptually:
1450 *   a |= (reg & ~b)
1451 */
1452static inline void regmask_set_if_not(regmask_t *a,
1453		struct ir3_register *reg, regmask_t *b)
1454{
1455	unsigned idx = regmask_idx(reg);
1456	if (reg->flags & IR3_REG_RELATIV) {
1457		unsigned i;
1458		for (i = 0; i < reg->size; i++, idx++)
1459			if (!((*b)[idx / 8] & (1 << (idx % 8))))
1460				(*a)[idx / 8] |= 1 << (idx % 8);
1461	} else {
1462		unsigned mask;
1463		for (mask = reg->wrmask; mask; mask >>= 1, idx++)
1464			if (mask & 1)
1465				if (!((*b)[idx / 8] & (1 << (idx % 8))))
1466					(*a)[idx / 8] |= 1 << (idx % 8);
1467	}
1468}
1469
1470static inline bool regmask_get(regmask_t *regmask,
1471		struct ir3_register *reg)
1472{
1473	unsigned idx = regmask_idx(reg);
1474	if (reg->flags & IR3_REG_RELATIV) {
1475		unsigned i;
1476		for (i = 0; i < reg->size; i++, idx++)
1477			if ((*regmask)[idx / 8] & (1 << (idx % 8)))
1478				return true;
1479	} else {
1480		unsigned mask;
1481		for (mask = reg->wrmask; mask; mask >>= 1, idx++)
1482			if (mask & 1)
1483				if ((*regmask)[idx / 8] & (1 << (idx % 8)))
1484					return true;
1485	}
1486	return false;
1487}
1488
1489/* ************************************************************************* */
1490
1491#endif /* IR3_H_ */
1492