ir3_cp.c revision 7e102996
1/*
2 * Copyright (C) 2014 Rob Clark <robclark@freedesktop.org>
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 * Authors:
24 *    Rob Clark <robclark@freedesktop.org>
25 */
26
27#include <math.h>
28
29#include "ir3.h"
30#include "ir3_compiler.h"
31#include "ir3_shader.h"
32
33/*
34 * Copy Propagate:
35 */
36
37struct ir3_cp_ctx {
38	struct ir3 *shader;
39	struct ir3_shader_variant *so;
40	unsigned immediate_idx;
41};
42
43/* is it a type preserving mov, with ok flags? */
44static bool is_eligible_mov(struct ir3_instruction *instr, bool allow_flags)
45{
46	if (is_same_type_mov(instr)) {
47		struct ir3_register *dst = instr->regs[0];
48		struct ir3_register *src = instr->regs[1];
49		struct ir3_instruction *src_instr = ssa(src);
50
51		/* only if mov src is SSA (not const/immed): */
52		if (!src_instr)
53			return false;
54
55		/* no indirect: */
56		if (dst->flags & IR3_REG_RELATIV)
57			return false;
58		if (src->flags & IR3_REG_RELATIV)
59			return false;
60
61		if (src->flags & IR3_REG_ARRAY)
62			return false;
63
64		if (!allow_flags)
65			if (src->flags & (IR3_REG_FABS | IR3_REG_FNEG |
66					IR3_REG_SABS | IR3_REG_SNEG | IR3_REG_BNOT))
67				return false;
68
69		/* TODO: remove this hack: */
70		if (src_instr->opc == OPC_META_FO)
71			return false;
72
73		return true;
74	}
75	return false;
76}
77
78static unsigned cp_flags(unsigned flags)
79{
80	/* only considering these flags (at least for now): */
81	flags &= (IR3_REG_CONST | IR3_REG_IMMED |
82			IR3_REG_FNEG | IR3_REG_FABS |
83			IR3_REG_SNEG | IR3_REG_SABS |
84			IR3_REG_BNOT | IR3_REG_RELATIV);
85	return flags;
86}
87
88static bool valid_flags(struct ir3_instruction *instr, unsigned n,
89		unsigned flags)
90{
91	unsigned valid_flags;
92
93	if ((flags & IR3_REG_HIGH) &&
94			(opc_cat(instr->opc) > 1) &&
95			(instr->block->shader->compiler->gpu_id >= 600))
96		return false;
97
98	flags = cp_flags(flags);
99
100	/* If destination is indirect, then source cannot be.. at least
101	 * I don't think so..
102	 */
103	if ((instr->regs[0]->flags & IR3_REG_RELATIV) &&
104			(flags & IR3_REG_RELATIV))
105		return false;
106
107	/* TODO it seems to *mostly* work to cp RELATIV, except we get some
108	 * intermittent piglit variable-indexing fails.  Newer blob driver
109	 * doesn't seem to cp these.  Possibly this is hw workaround?  Not
110	 * sure, but until that is understood better, lets just switch off
111	 * cp for indirect src's:
112	 */
113	if (flags & IR3_REG_RELATIV)
114		return false;
115
116	switch (opc_cat(instr->opc)) {
117	case 1:
118		valid_flags = IR3_REG_IMMED | IR3_REG_CONST | IR3_REG_RELATIV;
119		if (flags & ~valid_flags)
120			return false;
121		break;
122	case 2:
123		valid_flags = ir3_cat2_absneg(instr->opc) |
124				IR3_REG_CONST | IR3_REG_RELATIV;
125
126		if (ir3_cat2_int(instr->opc))
127			valid_flags |= IR3_REG_IMMED;
128
129		if (flags & ~valid_flags)
130			return false;
131
132		if (flags & (IR3_REG_CONST | IR3_REG_IMMED)) {
133			unsigned m = (n ^ 1) + 1;
134			/* cannot deal w/ const in both srcs:
135			 * (note that some cat2 actually only have a single src)
136			 */
137			if (m < instr->regs_count) {
138				struct ir3_register *reg = instr->regs[m];
139				if ((flags & IR3_REG_CONST) && (reg->flags & IR3_REG_CONST))
140					return false;
141				if ((flags & IR3_REG_IMMED) && (reg->flags & IR3_REG_IMMED))
142					return false;
143			}
144			/* cannot be const + ABS|NEG: */
145			if (flags & (IR3_REG_FABS | IR3_REG_FNEG |
146					IR3_REG_SABS | IR3_REG_SNEG | IR3_REG_BNOT))
147				return false;
148		}
149		break;
150	case 3:
151		valid_flags = ir3_cat3_absneg(instr->opc) |
152				IR3_REG_CONST | IR3_REG_RELATIV;
153
154		if (flags & ~valid_flags)
155			return false;
156
157		if (flags & (IR3_REG_CONST | IR3_REG_RELATIV)) {
158			/* cannot deal w/ const/relativ in 2nd src: */
159			if (n == 1)
160				return false;
161		}
162
163		if (flags & IR3_REG_CONST) {
164			/* cannot be const + ABS|NEG: */
165			if (flags & (IR3_REG_FABS | IR3_REG_FNEG |
166					IR3_REG_SABS | IR3_REG_SNEG | IR3_REG_BNOT))
167				return false;
168		}
169		break;
170	case 4:
171		/* seems like blob compiler avoids const as src.. */
172		/* TODO double check if this is still the case on a4xx */
173		if (flags & (IR3_REG_CONST | IR3_REG_IMMED))
174			return false;
175		if (flags & (IR3_REG_SABS | IR3_REG_SNEG))
176			return false;
177		break;
178	case 5:
179		/* no flags allowed */
180		if (flags)
181			return false;
182		break;
183	case 6:
184		valid_flags = IR3_REG_IMMED;
185		if (flags & ~valid_flags)
186			return false;
187
188		if (flags & IR3_REG_IMMED) {
189			/* doesn't seem like we can have immediate src for store
190			 * instructions:
191			 *
192			 * TODO this restriction could also apply to load instructions,
193			 * but for load instructions this arg is the address (and not
194			 * really sure any good way to test a hard-coded immed addr src)
195			 */
196			if (is_store(instr) && (n == 1))
197				return false;
198
199			if ((instr->opc == OPC_LDL) && (n != 1))
200				return false;
201
202			if ((instr->opc == OPC_STL) && (n != 2))
203				return false;
204
205			/* disallow CP into anything but the SSBO slot argument for
206			 * atomics:
207			 */
208			if (is_atomic(instr->opc) && (n != 0))
209				return false;
210
211			if (is_atomic(instr->opc) && !(instr->flags & IR3_INSTR_G))
212				return false;
213
214			/* as with atomics, ldib on a6xx can only have immediate for
215			 * SSBO slot argument
216			 */
217			if ((instr->opc == OPC_LDIB) && (n != 0))
218				return false;
219		}
220
221		break;
222	}
223
224	return true;
225}
226
227/* propagate register flags from src to dst.. negates need special
228 * handling to cancel each other out.
229 */
230static void combine_flags(unsigned *dstflags, struct ir3_instruction *src)
231{
232	unsigned srcflags = src->regs[1]->flags;
233
234	/* if what we are combining into already has (abs) flags,
235	 * we can drop (neg) from src:
236	 */
237	if (*dstflags & IR3_REG_FABS)
238		srcflags &= ~IR3_REG_FNEG;
239	if (*dstflags & IR3_REG_SABS)
240		srcflags &= ~IR3_REG_SNEG;
241
242	if (srcflags & IR3_REG_FABS)
243		*dstflags |= IR3_REG_FABS;
244	if (srcflags & IR3_REG_SABS)
245		*dstflags |= IR3_REG_SABS;
246	if (srcflags & IR3_REG_FNEG)
247		*dstflags ^= IR3_REG_FNEG;
248	if (srcflags & IR3_REG_SNEG)
249		*dstflags ^= IR3_REG_SNEG;
250	if (srcflags & IR3_REG_BNOT)
251		*dstflags ^= IR3_REG_BNOT;
252
253	*dstflags &= ~IR3_REG_SSA;
254	*dstflags |= srcflags & IR3_REG_SSA;
255	*dstflags |= srcflags & IR3_REG_CONST;
256	*dstflags |= srcflags & IR3_REG_IMMED;
257	*dstflags |= srcflags & IR3_REG_RELATIV;
258	*dstflags |= srcflags & IR3_REG_ARRAY;
259	*dstflags |= srcflags & IR3_REG_HIGH;
260
261	/* if src of the src is boolean we can drop the (abs) since we know
262	 * the source value is already a postitive integer.  This cleans
263	 * up the absnegs that get inserted when converting between nir and
264	 * native boolean (see ir3_b2n/n2b)
265	 */
266	struct ir3_instruction *srcsrc = ssa(src->regs[1]);
267	if (srcsrc && is_bool(srcsrc))
268		*dstflags &= ~IR3_REG_SABS;
269}
270
271static struct ir3_register *
272lower_immed(struct ir3_cp_ctx *ctx, struct ir3_register *reg, unsigned new_flags)
273{
274	unsigned swiz, idx, i;
275
276	reg = ir3_reg_clone(ctx->shader, reg);
277
278	/* in some cases, there are restrictions on (abs)/(neg) plus const..
279	 * so just evaluate those and clear the flags:
280	 */
281	if (new_flags & IR3_REG_SABS) {
282		reg->iim_val = abs(reg->iim_val);
283		new_flags &= ~IR3_REG_SABS;
284	}
285
286	if (new_flags & IR3_REG_FABS) {
287		reg->fim_val = fabs(reg->fim_val);
288		new_flags &= ~IR3_REG_FABS;
289	}
290
291	if (new_flags & IR3_REG_SNEG) {
292		reg->iim_val = -reg->iim_val;
293		new_flags &= ~IR3_REG_SNEG;
294	}
295
296	if (new_flags & IR3_REG_FNEG) {
297		reg->fim_val = -reg->fim_val;
298		new_flags &= ~IR3_REG_FNEG;
299	}
300
301	/* Reallocate for 4 more elements whenever it's necessary */
302	if (ctx->immediate_idx == ctx->so->immediates_size * 4) {
303		ctx->so->immediates_size += 4;
304		ctx->so->immediates = realloc (ctx->so->immediates,
305			ctx->so->immediates_size * sizeof (ctx->so->immediates[0]));
306	}
307
308	for (i = 0; i < ctx->immediate_idx; i++) {
309		swiz = i % 4;
310		idx  = i / 4;
311
312		if (ctx->so->immediates[idx].val[swiz] == reg->uim_val) {
313			break;
314		}
315	}
316
317	if (i == ctx->immediate_idx) {
318		/* need to generate a new immediate: */
319		swiz = i % 4;
320		idx  = i / 4;
321		ctx->so->immediates[idx].val[swiz] = reg->uim_val;
322		ctx->so->immediates_count = idx + 1;
323		ctx->immediate_idx++;
324	}
325
326	new_flags &= ~IR3_REG_IMMED;
327	new_flags |= IR3_REG_CONST;
328	reg->flags = new_flags;
329	reg->num = i + (4 * ctx->so->constbase.immediate);
330
331	return reg;
332}
333
334static void
335unuse(struct ir3_instruction *instr)
336{
337	debug_assert(instr->use_count > 0);
338
339	if (--instr->use_count == 0) {
340		struct ir3_block *block = instr->block;
341
342		instr->barrier_class = 0;
343		instr->barrier_conflict = 0;
344
345		/* we don't want to remove anything in keeps (which could
346		 * be things like array store's)
347		 */
348		for (unsigned i = 0; i < block->keeps_count; i++) {
349			debug_assert(block->keeps[i] != instr);
350		}
351	}
352}
353
354/**
355 * Handle cp for a given src register.  This additionally handles
356 * the cases of collapsing immedate/const (which replace the src
357 * register with a non-ssa src) or collapsing mov's from relative
358 * src (which needs to also fixup the address src reference by the
359 * instruction).
360 */
361static void
362reg_cp(struct ir3_cp_ctx *ctx, struct ir3_instruction *instr,
363		struct ir3_register *reg, unsigned n)
364{
365	struct ir3_instruction *src = ssa(reg);
366
367	if (is_eligible_mov(src, true)) {
368		/* simple case, no immed/const/relativ, only mov's w/ ssa src: */
369		struct ir3_register *src_reg = src->regs[1];
370		unsigned new_flags = reg->flags;
371
372		combine_flags(&new_flags, src);
373
374		if (valid_flags(instr, n, new_flags)) {
375			if (new_flags & IR3_REG_ARRAY) {
376				debug_assert(!(reg->flags & IR3_REG_ARRAY));
377				reg->array = src_reg->array;
378			}
379			reg->flags = new_flags;
380			reg->instr = ssa(src_reg);
381
382			instr->barrier_class |= src->barrier_class;
383			instr->barrier_conflict |= src->barrier_conflict;
384
385			unuse(src);
386			reg->instr->use_count++;
387		}
388
389	} else if (is_same_type_mov(src) &&
390			/* cannot collapse const/immed/etc into meta instrs: */
391			!is_meta(instr)) {
392		/* immed/const/etc cases, which require some special handling: */
393		struct ir3_register *src_reg = src->regs[1];
394		unsigned new_flags = reg->flags;
395
396		combine_flags(&new_flags, src);
397
398		if (!valid_flags(instr, n, new_flags)) {
399			/* See if lowering an immediate to const would help. */
400			if (valid_flags(instr, n, (new_flags & ~IR3_REG_IMMED) | IR3_REG_CONST)) {
401				debug_assert(new_flags & IR3_REG_IMMED);
402				instr->regs[n + 1] = lower_immed(ctx, src_reg, new_flags);
403				return;
404			}
405
406			/* special case for "normal" mad instructions, we can
407			 * try swapping the first two args if that fits better.
408			 *
409			 * the "plain" MAD's (ie. the ones that don't shift first
410			 * src prior to multiply) can swap their first two srcs if
411			 * src[0] is !CONST and src[1] is CONST:
412			 */
413			if ((n == 1) && is_mad(instr->opc) &&
414					!(instr->regs[0 + 1]->flags & (IR3_REG_CONST | IR3_REG_RELATIV)) &&
415					valid_flags(instr, 0, new_flags & ~IR3_REG_IMMED)) {
416				/* swap src[0] and src[1]: */
417				struct ir3_register *tmp;
418				tmp = instr->regs[0 + 1];
419				instr->regs[0 + 1] = instr->regs[1 + 1];
420				instr->regs[1 + 1] = tmp;
421
422				n = 0;
423			} else {
424				return;
425			}
426		}
427
428		/* Here we handle the special case of mov from
429		 * CONST and/or RELATIV.  These need to be handled
430		 * specially, because in the case of move from CONST
431		 * there is no src ir3_instruction so we need to
432		 * replace the ir3_register.  And in the case of
433		 * RELATIV we need to handle the address register
434		 * dependency.
435		 */
436		if (src_reg->flags & IR3_REG_CONST) {
437			/* an instruction cannot reference two different
438			 * address registers:
439			 */
440			if ((src_reg->flags & IR3_REG_RELATIV) &&
441					conflicts(instr->address, reg->instr->address))
442				return;
443
444			/* This seems to be a hw bug, or something where the timings
445			 * just somehow don't work out.  This restriction may only
446			 * apply if the first src is also CONST.
447			 */
448			if ((opc_cat(instr->opc) == 3) && (n == 2) &&
449					(src_reg->flags & IR3_REG_RELATIV) &&
450					(src_reg->array.offset == 0))
451				return;
452
453			src_reg = ir3_reg_clone(instr->block->shader, src_reg);
454			src_reg->flags = new_flags;
455			instr->regs[n+1] = src_reg;
456
457			if (src_reg->flags & IR3_REG_RELATIV)
458				ir3_instr_set_address(instr, reg->instr->address);
459
460			return;
461		}
462
463		if ((src_reg->flags & IR3_REG_RELATIV) &&
464				!conflicts(instr->address, reg->instr->address)) {
465			src_reg = ir3_reg_clone(instr->block->shader, src_reg);
466			src_reg->flags = new_flags;
467			instr->regs[n+1] = src_reg;
468			ir3_instr_set_address(instr, reg->instr->address);
469
470			return;
471		}
472
473		/* NOTE: seems we can only do immed integers, so don't
474		 * need to care about float.  But we do need to handle
475		 * abs/neg *before* checking that the immediate requires
476		 * few enough bits to encode:
477		 *
478		 * TODO: do we need to do something to avoid accidentally
479		 * catching a float immed?
480		 */
481		if (src_reg->flags & IR3_REG_IMMED) {
482			int32_t iim_val = src_reg->iim_val;
483
484			debug_assert((opc_cat(instr->opc) == 1) ||
485					(opc_cat(instr->opc) == 6) ||
486					ir3_cat2_int(instr->opc) ||
487					(is_mad(instr->opc) && (n == 0)));
488
489			if (new_flags & IR3_REG_SABS)
490				iim_val = abs(iim_val);
491
492			if (new_flags & IR3_REG_SNEG)
493				iim_val = -iim_val;
494
495			if (new_flags & IR3_REG_BNOT)
496				iim_val = ~iim_val;
497
498			/* other than category 1 (mov) we can only encode up to 10 bits: */
499			if ((instr->opc == OPC_MOV) ||
500					!((iim_val & ~0x3ff) && (-iim_val & ~0x3ff))) {
501				new_flags &= ~(IR3_REG_SABS | IR3_REG_SNEG | IR3_REG_BNOT);
502				src_reg = ir3_reg_clone(instr->block->shader, src_reg);
503				src_reg->flags = new_flags;
504				src_reg->iim_val = iim_val;
505				instr->regs[n+1] = src_reg;
506			} else if (valid_flags(instr, n, (new_flags & ~IR3_REG_IMMED) | IR3_REG_CONST)) {
507				/* See if lowering an immediate to const would help. */
508				instr->regs[n+1] = lower_immed(ctx, src_reg, new_flags);
509			}
510
511			return;
512		}
513	}
514}
515
516/* Handle special case of eliminating output mov, and similar cases where
517 * there isn't a normal "consuming" instruction.  In this case we cannot
518 * collapse flags (ie. output mov from const, or w/ abs/neg flags, cannot
519 * be eliminated)
520 */
521static struct ir3_instruction *
522eliminate_output_mov(struct ir3_instruction *instr)
523{
524	if (is_eligible_mov(instr, false)) {
525		struct ir3_register *reg = instr->regs[1];
526		if (!(reg->flags & IR3_REG_ARRAY)) {
527			struct ir3_instruction *src_instr = ssa(reg);
528			debug_assert(src_instr);
529			return src_instr;
530		}
531	}
532	return instr;
533}
534
535/**
536 * Find instruction src's which are mov's that can be collapsed, replacing
537 * the mov dst with the mov src
538 */
539static void
540instr_cp(struct ir3_cp_ctx *ctx, struct ir3_instruction *instr)
541{
542	struct ir3_register *reg;
543
544	if (instr->regs_count == 0)
545		return;
546
547	if (ir3_instr_check_mark(instr))
548		return;
549
550	/* walk down the graph from each src: */
551	foreach_src_n(reg, n, instr) {
552		struct ir3_instruction *src = ssa(reg);
553
554		if (!src)
555			continue;
556
557		instr_cp(ctx, src);
558
559		/* TODO non-indirect access we could figure out which register
560		 * we actually want and allow cp..
561		 */
562		if (reg->flags & IR3_REG_ARRAY)
563			continue;
564
565		/* Don't CP absneg into meta instructions, that won't end well: */
566		if (is_meta(instr) && (src->opc != OPC_MOV))
567			continue;
568
569		reg_cp(ctx, instr, reg, n);
570	}
571
572	if (instr->regs[0]->flags & IR3_REG_ARRAY) {
573		struct ir3_instruction *src = ssa(instr->regs[0]);
574		if (src)
575			instr_cp(ctx, src);
576	}
577
578	if (instr->address) {
579		instr_cp(ctx, instr->address);
580		ir3_instr_set_address(instr, eliminate_output_mov(instr->address));
581	}
582
583	/* we can end up with extra cmps.s from frontend, which uses a
584	 *
585	 *    cmps.s p0.x, cond, 0
586	 *
587	 * as a way to mov into the predicate register.  But frequently 'cond'
588	 * is itself a cmps.s/cmps.f/cmps.u.  So detect this special case and
589	 * just re-write the instruction writing predicate register to get rid
590	 * of the double cmps.
591	 */
592	if ((instr->opc == OPC_CMPS_S) &&
593			(instr->regs[0]->num == regid(REG_P0, 0)) &&
594			ssa(instr->regs[1]) &&
595			(instr->regs[2]->flags & IR3_REG_IMMED) &&
596			(instr->regs[2]->iim_val == 0)) {
597		struct ir3_instruction *cond = ssa(instr->regs[1]);
598		switch (cond->opc) {
599		case OPC_CMPS_S:
600		case OPC_CMPS_F:
601		case OPC_CMPS_U:
602			instr->opc   = cond->opc;
603			instr->flags = cond->flags;
604			instr->cat2  = cond->cat2;
605			instr->address = cond->address;
606			instr->regs[1] = cond->regs[1];
607			instr->regs[2] = cond->regs[2];
608			instr->barrier_class |= cond->barrier_class;
609			instr->barrier_conflict |= cond->barrier_conflict;
610			unuse(cond);
611			break;
612		default:
613			break;
614		}
615	}
616
617	/* Handle converting a sam.s2en (taking samp/tex idx params via
618	 * register) into a normal sam (encoding immediate samp/tex idx)
619	 * if they are immediate.  This saves some instructions and regs
620	 * in the common case where we know samp/tex at compile time:
621	 */
622	if (is_tex(instr) && (instr->flags & IR3_INSTR_S2EN) &&
623			!(ir3_shader_debug & IR3_DBG_FORCES2EN)) {
624		/* The first src will be a fan-in (collect), if both of it's
625		 * two sources are mov from imm, then we can
626		 */
627		struct ir3_instruction *samp_tex = ssa(instr->regs[1]);
628
629		debug_assert(samp_tex->opc == OPC_META_FI);
630
631		struct ir3_instruction *samp = ssa(samp_tex->regs[1]);
632		struct ir3_instruction *tex  = ssa(samp_tex->regs[2]);
633
634		if ((samp->opc == OPC_MOV) &&
635				(samp->regs[1]->flags & IR3_REG_IMMED) &&
636				(tex->opc == OPC_MOV) &&
637				(tex->regs[1]->flags & IR3_REG_IMMED)) {
638			instr->flags &= ~IR3_INSTR_S2EN;
639			instr->cat5.samp = samp->regs[1]->iim_val;
640			instr->cat5.tex  = tex->regs[1]->iim_val;
641			instr->regs[1]->instr = NULL;
642		}
643	}
644}
645
646void
647ir3_cp(struct ir3 *ir, struct ir3_shader_variant *so)
648{
649	struct ir3_cp_ctx ctx = {
650			.shader = ir,
651			.so = so,
652	};
653
654	/* This is a bit annoying, and probably wouldn't be necessary if we
655	 * tracked a reverse link from producing instruction to consumer.
656	 * But we need to know when we've eliminated the last consumer of
657	 * a mov, so we need to do a pass to first count consumers of a
658	 * mov.
659	 */
660	list_for_each_entry (struct ir3_block, block, &ir->block_list, node) {
661		list_for_each_entry (struct ir3_instruction, instr, &block->instr_list, node) {
662			struct ir3_instruction *src;
663
664			/* by the way, we don't account for false-dep's, so the CP
665			 * pass should always happen before false-dep's are inserted
666			 */
667			debug_assert(instr->deps_count == 0);
668
669			foreach_ssa_src(src, instr) {
670				src->use_count++;
671			}
672		}
673	}
674
675	ir3_clear_mark(ir);
676
677	for (unsigned i = 0; i < ir->noutputs; i++) {
678		if (ir->outputs[i]) {
679			instr_cp(&ctx, ir->outputs[i]);
680			ir->outputs[i] = eliminate_output_mov(ir->outputs[i]);
681		}
682	}
683
684	list_for_each_entry (struct ir3_block, block, &ir->block_list, node) {
685		if (block->condition) {
686			instr_cp(&ctx, block->condition);
687			block->condition = eliminate_output_mov(block->condition);
688		}
689
690		for (unsigned i = 0; i < block->keeps_count; i++) {
691			instr_cp(&ctx, block->keeps[i]);
692			block->keeps[i] = eliminate_output_mov(block->keeps[i]);
693		}
694	}
695}
696