1/*
2 * Copyright (C) 2015 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 <stdarg.h>
28
29#include "util/u_string.h"
30#include "util/u_memory.h"
31#include "util/u_math.h"
32
33#include "ir3_compiler.h"
34#include "ir3_image.h"
35#include "ir3_shader.h"
36#include "ir3_nir.h"
37
38#include "instr-a3xx.h"
39#include "ir3.h"
40#include "ir3_context.h"
41
42
43static struct ir3_instruction *
44create_indirect_load(struct ir3_context *ctx, unsigned arrsz, int n,
45		struct ir3_instruction *address, struct ir3_instruction *collect)
46{
47	struct ir3_block *block = ctx->block;
48	struct ir3_instruction *mov;
49	struct ir3_register *src;
50
51	mov = ir3_instr_create(block, OPC_MOV);
52	mov->cat1.src_type = TYPE_U32;
53	mov->cat1.dst_type = TYPE_U32;
54	ir3_reg_create(mov, 0, 0);
55	src = ir3_reg_create(mov, 0, IR3_REG_SSA | IR3_REG_RELATIV);
56	src->instr = collect;
57	src->size  = arrsz;
58	src->array.offset = n;
59
60	ir3_instr_set_address(mov, address);
61
62	return mov;
63}
64
65static struct ir3_instruction *
66create_input_compmask(struct ir3_context *ctx, unsigned n, unsigned compmask)
67{
68	struct ir3_instruction *in;
69
70	in = ir3_instr_create(ctx->in_block, OPC_META_INPUT);
71	in->inout.block = ctx->in_block;
72	ir3_reg_create(in, n, 0);
73
74	in->regs[0]->wrmask = compmask;
75
76	return in;
77}
78
79static struct ir3_instruction *
80create_input(struct ir3_context *ctx, unsigned n)
81{
82	return create_input_compmask(ctx, n, 0x1);
83}
84
85static struct ir3_instruction *
86create_frag_input(struct ir3_context *ctx, bool use_ldlv, unsigned n)
87{
88	struct ir3_block *block = ctx->block;
89	struct ir3_instruction *instr;
90	/* packed inloc is fixed up later: */
91	struct ir3_instruction *inloc = create_immed(block, n);
92
93	if (use_ldlv) {
94		instr = ir3_LDLV(block, inloc, 0, create_immed(block, 1), 0);
95		instr->cat6.type = TYPE_U32;
96		instr->cat6.iim_val = 1;
97	} else {
98		instr = ir3_BARY_F(block, inloc, 0, ctx->ij_pixel, 0);
99		instr->regs[2]->wrmask = 0x3;
100	}
101
102	return instr;
103}
104
105static struct ir3_instruction *
106create_driver_param(struct ir3_context *ctx, enum ir3_driver_param dp)
107{
108	/* first four vec4 sysval's reserved for UBOs: */
109	/* NOTE: dp is in scalar, but there can be >4 dp components: */
110	unsigned n = ctx->so->constbase.driver_param;
111	unsigned r = regid(n + dp / 4, dp % 4);
112	return create_uniform(ctx->block, r);
113}
114
115/*
116 * Adreno uses uint rather than having dedicated bool type,
117 * which (potentially) requires some conversion, in particular
118 * when using output of an bool instr to int input, or visa
119 * versa.
120 *
121 *         | Adreno  |  NIR  |
122 *  -------+---------+-------+-
123 *   true  |    1    |  ~0   |
124 *   false |    0    |   0   |
125 *
126 * To convert from an adreno bool (uint) to nir, use:
127 *
128 *    absneg.s dst, (neg)src
129 *
130 * To convert back in the other direction:
131 *
132 *    absneg.s dst, (abs)arc
133 *
134 * The CP step can clean up the absneg.s that cancel each other
135 * out, and with a slight bit of extra cleverness (to recognize
136 * the instructions which produce either a 0 or 1) can eliminate
137 * the absneg.s's completely when an instruction that wants
138 * 0/1 consumes the result.  For example, when a nir 'bcsel'
139 * consumes the result of 'feq'.  So we should be able to get by
140 * without a boolean resolve step, and without incuring any
141 * extra penalty in instruction count.
142 */
143
144/* NIR bool -> native (adreno): */
145static struct ir3_instruction *
146ir3_b2n(struct ir3_block *block, struct ir3_instruction *instr)
147{
148	return ir3_ABSNEG_S(block, instr, IR3_REG_SABS);
149}
150
151/* native (adreno) -> NIR bool: */
152static struct ir3_instruction *
153ir3_n2b(struct ir3_block *block, struct ir3_instruction *instr)
154{
155	return ir3_ABSNEG_S(block, instr, IR3_REG_SNEG);
156}
157
158/*
159 * alu/sfu instructions:
160 */
161
162static struct ir3_instruction *
163create_cov(struct ir3_context *ctx, struct ir3_instruction *src,
164		unsigned src_bitsize, nir_op op)
165{
166	type_t src_type, dst_type;
167
168	switch (op) {
169	case nir_op_f2f32:
170	case nir_op_f2f16_rtne:
171	case nir_op_f2f16_rtz:
172	case nir_op_f2f16:
173	case nir_op_f2i32:
174	case nir_op_f2i16:
175	case nir_op_f2i8:
176	case nir_op_f2u32:
177	case nir_op_f2u16:
178	case nir_op_f2u8:
179		switch (src_bitsize) {
180		case 32:
181			src_type = TYPE_F32;
182			break;
183		case 16:
184			src_type = TYPE_F16;
185			break;
186		default:
187			ir3_context_error(ctx, "invalid src bit size: %u", src_bitsize);
188		}
189		break;
190
191	case nir_op_i2f32:
192	case nir_op_i2f16:
193	case nir_op_i2i32:
194	case nir_op_i2i16:
195	case nir_op_i2i8:
196		switch (src_bitsize) {
197		case 32:
198			src_type = TYPE_S32;
199			break;
200		case 16:
201			src_type = TYPE_S16;
202			break;
203		case 8:
204			src_type = TYPE_S8;
205			break;
206		default:
207			ir3_context_error(ctx, "invalid src bit size: %u", src_bitsize);
208		}
209		break;
210
211	case nir_op_u2f32:
212	case nir_op_u2f16:
213	case nir_op_u2u32:
214	case nir_op_u2u16:
215	case nir_op_u2u8:
216		switch (src_bitsize) {
217		case 32:
218			src_type = TYPE_U32;
219			break;
220		case 16:
221			src_type = TYPE_U16;
222			break;
223		case 8:
224			src_type = TYPE_U8;
225			break;
226		default:
227			ir3_context_error(ctx, "invalid src bit size: %u", src_bitsize);
228		}
229		break;
230
231	default:
232		ir3_context_error(ctx, "invalid conversion op: %u", op);
233	}
234
235	switch (op) {
236	case nir_op_f2f32:
237	case nir_op_i2f32:
238	case nir_op_u2f32:
239		dst_type = TYPE_F32;
240		break;
241
242	case nir_op_f2f16_rtne:
243	case nir_op_f2f16_rtz:
244	case nir_op_f2f16:
245		/* TODO how to handle rounding mode? */
246	case nir_op_i2f16:
247	case nir_op_u2f16:
248		dst_type = TYPE_F16;
249		break;
250
251	case nir_op_f2i32:
252	case nir_op_i2i32:
253		dst_type = TYPE_S32;
254		break;
255
256	case nir_op_f2i16:
257	case nir_op_i2i16:
258		dst_type = TYPE_S16;
259		break;
260
261	case nir_op_f2i8:
262	case nir_op_i2i8:
263		dst_type = TYPE_S8;
264		break;
265
266	case nir_op_f2u32:
267	case nir_op_u2u32:
268		dst_type = TYPE_U32;
269		break;
270
271	case nir_op_f2u16:
272	case nir_op_u2u16:
273		dst_type = TYPE_U16;
274		break;
275
276	case nir_op_f2u8:
277	case nir_op_u2u8:
278		dst_type = TYPE_U8;
279		break;
280
281	default:
282		ir3_context_error(ctx, "invalid conversion op: %u", op);
283	}
284
285	return ir3_COV(ctx->block, src, src_type, dst_type);
286}
287
288static void
289emit_alu(struct ir3_context *ctx, nir_alu_instr *alu)
290{
291	const nir_op_info *info = &nir_op_infos[alu->op];
292	struct ir3_instruction **dst, *src[info->num_inputs];
293	unsigned bs[info->num_inputs];     /* bit size */
294	struct ir3_block *b = ctx->block;
295	unsigned dst_sz, wrmask;
296
297	if (alu->dest.dest.is_ssa) {
298		dst_sz = alu->dest.dest.ssa.num_components;
299		wrmask = (1 << dst_sz) - 1;
300	} else {
301		dst_sz = alu->dest.dest.reg.reg->num_components;
302		wrmask = alu->dest.write_mask;
303	}
304
305	dst = ir3_get_dst(ctx, &alu->dest.dest, dst_sz);
306
307	/* Vectors are special in that they have non-scalarized writemasks,
308	 * and just take the first swizzle channel for each argument in
309	 * order into each writemask channel.
310	 */
311	if ((alu->op == nir_op_vec2) ||
312			(alu->op == nir_op_vec3) ||
313			(alu->op == nir_op_vec4)) {
314
315		for (int i = 0; i < info->num_inputs; i++) {
316			nir_alu_src *asrc = &alu->src[i];
317
318			compile_assert(ctx, !asrc->abs);
319			compile_assert(ctx, !asrc->negate);
320
321			src[i] = ir3_get_src(ctx, &asrc->src)[asrc->swizzle[0]];
322			if (!src[i])
323				src[i] = create_immed(ctx->block, 0);
324			dst[i] = ir3_MOV(b, src[i], TYPE_U32);
325		}
326
327		ir3_put_dst(ctx, &alu->dest.dest);
328		return;
329	}
330
331	/* We also get mov's with more than one component for mov's so
332	 * handle those specially:
333	 */
334	if ((alu->op == nir_op_imov) || (alu->op == nir_op_fmov)) {
335		type_t type = (alu->op == nir_op_imov) ? TYPE_U32 : TYPE_F32;
336		nir_alu_src *asrc = &alu->src[0];
337		struct ir3_instruction *const *src0 = ir3_get_src(ctx, &asrc->src);
338
339		for (unsigned i = 0; i < dst_sz; i++) {
340			if (wrmask & (1 << i)) {
341				dst[i] = ir3_MOV(b, src0[asrc->swizzle[i]], type);
342			} else {
343				dst[i] = NULL;
344			}
345		}
346
347		ir3_put_dst(ctx, &alu->dest.dest);
348		return;
349	}
350
351	/* General case: We can just grab the one used channel per src. */
352	for (int i = 0; i < info->num_inputs; i++) {
353		unsigned chan = ffs(alu->dest.write_mask) - 1;
354		nir_alu_src *asrc = &alu->src[i];
355
356		compile_assert(ctx, !asrc->abs);
357		compile_assert(ctx, !asrc->negate);
358
359		src[i] = ir3_get_src(ctx, &asrc->src)[asrc->swizzle[chan]];
360		bs[i] = nir_src_bit_size(asrc->src);
361
362		compile_assert(ctx, src[i]);
363	}
364
365	switch (alu->op) {
366	case nir_op_f2f32:
367	case nir_op_f2f16_rtne:
368	case nir_op_f2f16_rtz:
369	case nir_op_f2f16:
370	case nir_op_f2i32:
371	case nir_op_f2i16:
372	case nir_op_f2i8:
373	case nir_op_f2u32:
374	case nir_op_f2u16:
375	case nir_op_f2u8:
376	case nir_op_i2f32:
377	case nir_op_i2f16:
378	case nir_op_i2i32:
379	case nir_op_i2i16:
380	case nir_op_i2i8:
381	case nir_op_u2f32:
382	case nir_op_u2f16:
383	case nir_op_u2u32:
384	case nir_op_u2u16:
385	case nir_op_u2u8:
386		dst[0] = create_cov(ctx, src[0], bs[0], alu->op);
387		break;
388	case nir_op_f2b32:
389		dst[0] = ir3_CMPS_F(b, src[0], 0, create_immed(b, fui(0.0)), 0);
390		dst[0]->cat2.condition = IR3_COND_NE;
391		dst[0] = ir3_n2b(b, dst[0]);
392		break;
393	case nir_op_b2f16:
394	case nir_op_b2f32:
395		dst[0] = ir3_COV(b, ir3_b2n(b, src[0]), TYPE_U32, TYPE_F32);
396		break;
397	case nir_op_b2i8:
398	case nir_op_b2i16:
399	case nir_op_b2i32:
400		dst[0] = ir3_b2n(b, src[0]);
401		break;
402	case nir_op_i2b32:
403		dst[0] = ir3_CMPS_S(b, src[0], 0, create_immed(b, 0), 0);
404		dst[0]->cat2.condition = IR3_COND_NE;
405		dst[0] = ir3_n2b(b, dst[0]);
406		break;
407
408	case nir_op_fneg:
409		dst[0] = ir3_ABSNEG_F(b, src[0], IR3_REG_FNEG);
410		break;
411	case nir_op_fabs:
412		dst[0] = ir3_ABSNEG_F(b, src[0], IR3_REG_FABS);
413		break;
414	case nir_op_fmax:
415		dst[0] = ir3_MAX_F(b, src[0], 0, src[1], 0);
416		break;
417	case nir_op_fmin:
418		dst[0] = ir3_MIN_F(b, src[0], 0, src[1], 0);
419		break;
420	case nir_op_fsat:
421		/* if there is just a single use of the src, and it supports
422		 * (sat) bit, we can just fold the (sat) flag back to the
423		 * src instruction and create a mov.  This is easier for cp
424		 * to eliminate.
425		 *
426		 * TODO probably opc_cat==4 is ok too
427		 */
428		if (alu->src[0].src.is_ssa &&
429				(list_length(&alu->src[0].src.ssa->uses) == 1) &&
430				((opc_cat(src[0]->opc) == 2) || (opc_cat(src[0]->opc) == 3))) {
431			src[0]->flags |= IR3_INSTR_SAT;
432			dst[0] = ir3_MOV(b, src[0], TYPE_U32);
433		} else {
434			/* otherwise generate a max.f that saturates.. blob does
435			 * similar (generating a cat2 mov using max.f)
436			 */
437			dst[0] = ir3_MAX_F(b, src[0], 0, src[0], 0);
438			dst[0]->flags |= IR3_INSTR_SAT;
439		}
440		break;
441	case nir_op_fmul:
442		dst[0] = ir3_MUL_F(b, src[0], 0, src[1], 0);
443		break;
444	case nir_op_fadd:
445		dst[0] = ir3_ADD_F(b, src[0], 0, src[1], 0);
446		break;
447	case nir_op_fsub:
448		dst[0] = ir3_ADD_F(b, src[0], 0, src[1], IR3_REG_FNEG);
449		break;
450	case nir_op_ffma:
451		dst[0] = ir3_MAD_F32(b, src[0], 0, src[1], 0, src[2], 0);
452		break;
453	case nir_op_fddx:
454		dst[0] = ir3_DSX(b, src[0], 0);
455		dst[0]->cat5.type = TYPE_F32;
456		break;
457	case nir_op_fddy:
458		dst[0] = ir3_DSY(b, src[0], 0);
459		dst[0]->cat5.type = TYPE_F32;
460		break;
461		break;
462	case nir_op_flt32:
463		dst[0] = ir3_CMPS_F(b, src[0], 0, src[1], 0);
464		dst[0]->cat2.condition = IR3_COND_LT;
465		dst[0] = ir3_n2b(b, dst[0]);
466		break;
467	case nir_op_fge32:
468		dst[0] = ir3_CMPS_F(b, src[0], 0, src[1], 0);
469		dst[0]->cat2.condition = IR3_COND_GE;
470		dst[0] = ir3_n2b(b, dst[0]);
471		break;
472	case nir_op_feq32:
473		dst[0] = ir3_CMPS_F(b, src[0], 0, src[1], 0);
474		dst[0]->cat2.condition = IR3_COND_EQ;
475		dst[0] = ir3_n2b(b, dst[0]);
476		break;
477	case nir_op_fne32:
478		dst[0] = ir3_CMPS_F(b, src[0], 0, src[1], 0);
479		dst[0]->cat2.condition = IR3_COND_NE;
480		dst[0] = ir3_n2b(b, dst[0]);
481		break;
482	case nir_op_fceil:
483		dst[0] = ir3_CEIL_F(b, src[0], 0);
484		break;
485	case nir_op_ffloor:
486		dst[0] = ir3_FLOOR_F(b, src[0], 0);
487		break;
488	case nir_op_ftrunc:
489		dst[0] = ir3_TRUNC_F(b, src[0], 0);
490		break;
491	case nir_op_fround_even:
492		dst[0] = ir3_RNDNE_F(b, src[0], 0);
493		break;
494	case nir_op_fsign:
495		dst[0] = ir3_SIGN_F(b, src[0], 0);
496		break;
497
498	case nir_op_fsin:
499		dst[0] = ir3_SIN(b, src[0], 0);
500		break;
501	case nir_op_fcos:
502		dst[0] = ir3_COS(b, src[0], 0);
503		break;
504	case nir_op_frsq:
505		dst[0] = ir3_RSQ(b, src[0], 0);
506		break;
507	case nir_op_frcp:
508		dst[0] = ir3_RCP(b, src[0], 0);
509		break;
510	case nir_op_flog2:
511		dst[0] = ir3_LOG2(b, src[0], 0);
512		break;
513	case nir_op_fexp2:
514		dst[0] = ir3_EXP2(b, src[0], 0);
515		break;
516	case nir_op_fsqrt:
517		dst[0] = ir3_SQRT(b, src[0], 0);
518		break;
519
520	case nir_op_iabs:
521		dst[0] = ir3_ABSNEG_S(b, src[0], IR3_REG_SABS);
522		break;
523	case nir_op_iadd:
524		dst[0] = ir3_ADD_U(b, src[0], 0, src[1], 0);
525		break;
526	case nir_op_iand:
527		dst[0] = ir3_AND_B(b, src[0], 0, src[1], 0);
528		break;
529	case nir_op_imax:
530		dst[0] = ir3_MAX_S(b, src[0], 0, src[1], 0);
531		break;
532	case nir_op_umax:
533		dst[0] = ir3_MAX_U(b, src[0], 0, src[1], 0);
534		break;
535	case nir_op_imin:
536		dst[0] = ir3_MIN_S(b, src[0], 0, src[1], 0);
537		break;
538	case nir_op_umin:
539		dst[0] = ir3_MIN_U(b, src[0], 0, src[1], 0);
540		break;
541	case nir_op_imul:
542		/*
543		 * dst = (al * bl) + (ah * bl << 16) + (al * bh << 16)
544		 *   mull.u tmp0, a, b           ; mul low, i.e. al * bl
545		 *   madsh.m16 tmp1, a, b, tmp0  ; mul-add shift high mix, i.e. ah * bl << 16
546		 *   madsh.m16 dst, b, a, tmp1   ; i.e. al * bh << 16
547		 */
548		dst[0] = ir3_MADSH_M16(b, src[1], 0, src[0], 0,
549					ir3_MADSH_M16(b, src[0], 0, src[1], 0,
550						ir3_MULL_U(b, src[0], 0, src[1], 0), 0), 0);
551		break;
552	case nir_op_ineg:
553		dst[0] = ir3_ABSNEG_S(b, src[0], IR3_REG_SNEG);
554		break;
555	case nir_op_inot:
556		dst[0] = ir3_NOT_B(b, src[0], 0);
557		break;
558	case nir_op_ior:
559		dst[0] = ir3_OR_B(b, src[0], 0, src[1], 0);
560		break;
561	case nir_op_ishl:
562		dst[0] = ir3_SHL_B(b, src[0], 0, src[1], 0);
563		break;
564	case nir_op_ishr:
565		dst[0] = ir3_ASHR_B(b, src[0], 0, src[1], 0);
566		break;
567	case nir_op_isub:
568		dst[0] = ir3_SUB_U(b, src[0], 0, src[1], 0);
569		break;
570	case nir_op_ixor:
571		dst[0] = ir3_XOR_B(b, src[0], 0, src[1], 0);
572		break;
573	case nir_op_ushr:
574		dst[0] = ir3_SHR_B(b, src[0], 0, src[1], 0);
575		break;
576	case nir_op_ilt32:
577		dst[0] = ir3_CMPS_S(b, src[0], 0, src[1], 0);
578		dst[0]->cat2.condition = IR3_COND_LT;
579		dst[0] = ir3_n2b(b, dst[0]);
580		break;
581	case nir_op_ige32:
582		dst[0] = ir3_CMPS_S(b, src[0], 0, src[1], 0);
583		dst[0]->cat2.condition = IR3_COND_GE;
584		dst[0] = ir3_n2b(b, dst[0]);
585		break;
586	case nir_op_ieq32:
587		dst[0] = ir3_CMPS_S(b, src[0], 0, src[1], 0);
588		dst[0]->cat2.condition = IR3_COND_EQ;
589		dst[0] = ir3_n2b(b, dst[0]);
590		break;
591	case nir_op_ine32:
592		dst[0] = ir3_CMPS_S(b, src[0], 0, src[1], 0);
593		dst[0]->cat2.condition = IR3_COND_NE;
594		dst[0] = ir3_n2b(b, dst[0]);
595		break;
596	case nir_op_ult32:
597		dst[0] = ir3_CMPS_U(b, src[0], 0, src[1], 0);
598		dst[0]->cat2.condition = IR3_COND_LT;
599		dst[0] = ir3_n2b(b, dst[0]);
600		break;
601	case nir_op_uge32:
602		dst[0] = ir3_CMPS_U(b, src[0], 0, src[1], 0);
603		dst[0]->cat2.condition = IR3_COND_GE;
604		dst[0] = ir3_n2b(b, dst[0]);
605		break;
606
607	case nir_op_b32csel: {
608		struct ir3_instruction *cond = ir3_b2n(b, src[0]);
609		compile_assert(ctx, bs[1] == bs[2]);
610		/* the boolean condition is 32b even if src[1] and src[2] are
611		 * half-precision, but sel.b16 wants all three src's to be the
612		 * same type.
613		 */
614		if (bs[1] < 32)
615			cond = ir3_COV(b, cond, TYPE_U32, TYPE_U16);
616		dst[0] = ir3_SEL_B32(b, src[1], 0, cond, 0, src[2], 0);
617		break;
618	}
619	case nir_op_bit_count: {
620		// TODO, we need to do this 16b at a time on a5xx+a6xx.. need to
621		// double check on earlier gen's.  Once half-precision support is
622		// in place, this should probably move to a NIR lowering pass:
623		struct ir3_instruction *hi, *lo;
624
625		hi = ir3_COV(b, ir3_SHR_B(b, src[0], 0, create_immed(b, 16), 0),
626				TYPE_U32, TYPE_U16);
627		lo = ir3_COV(b, src[0], TYPE_U32, TYPE_U16);
628
629		hi = ir3_CBITS_B(b, hi, 0);
630		lo = ir3_CBITS_B(b, lo, 0);
631
632		// TODO maybe the builders should default to making dst half-precision
633		// if the src's were half precision, to make this less awkward.. otoh
634		// we should probably just do this lowering in NIR.
635		hi->regs[0]->flags |= IR3_REG_HALF;
636		lo->regs[0]->flags |= IR3_REG_HALF;
637
638		dst[0] = ir3_ADD_S(b, hi, 0, lo, 0);
639		dst[0]->regs[0]->flags |= IR3_REG_HALF;
640		dst[0] = ir3_COV(b, dst[0], TYPE_U16, TYPE_U32);
641		break;
642	}
643	case nir_op_ifind_msb: {
644		struct ir3_instruction *cmp;
645		dst[0] = ir3_CLZ_S(b, src[0], 0);
646		cmp = ir3_CMPS_S(b, dst[0], 0, create_immed(b, 0), 0);
647		cmp->cat2.condition = IR3_COND_GE;
648		dst[0] = ir3_SEL_B32(b,
649				ir3_SUB_U(b, create_immed(b, 31), 0, dst[0], 0), 0,
650				cmp, 0, dst[0], 0);
651		break;
652	}
653	case nir_op_ufind_msb:
654		dst[0] = ir3_CLZ_B(b, src[0], 0);
655		dst[0] = ir3_SEL_B32(b,
656				ir3_SUB_U(b, create_immed(b, 31), 0, dst[0], 0), 0,
657				src[0], 0, dst[0], 0);
658		break;
659	case nir_op_find_lsb:
660		dst[0] = ir3_BFREV_B(b, src[0], 0);
661		dst[0] = ir3_CLZ_B(b, dst[0], 0);
662		break;
663	case nir_op_bitfield_reverse:
664		dst[0] = ir3_BFREV_B(b, src[0], 0);
665		break;
666
667	default:
668		ir3_context_error(ctx, "Unhandled ALU op: %s\n",
669				nir_op_infos[alu->op].name);
670		break;
671	}
672
673	ir3_put_dst(ctx, &alu->dest.dest);
674}
675
676/* handles direct/indirect UBO reads: */
677static void
678emit_intrinsic_load_ubo(struct ir3_context *ctx, nir_intrinsic_instr *intr,
679		struct ir3_instruction **dst)
680{
681	struct ir3_block *b = ctx->block;
682	struct ir3_instruction *base_lo, *base_hi, *addr, *src0, *src1;
683	/* UBO addresses are the first driver params, but subtract 2 here to
684	 * account for nir_lower_uniforms_to_ubo rebasing the UBOs such that UBO 0
685	 * is the uniforms: */
686	unsigned ubo = regid(ctx->so->constbase.ubo, 0) - 2;
687	const unsigned ptrsz = ir3_pointer_size(ctx);
688
689	int off = 0;
690
691	/* First src is ubo index, which could either be an immed or not: */
692	src0 = ir3_get_src(ctx, &intr->src[0])[0];
693	if (is_same_type_mov(src0) &&
694			(src0->regs[1]->flags & IR3_REG_IMMED)) {
695		base_lo = create_uniform(b, ubo + (src0->regs[1]->iim_val * ptrsz));
696		base_hi = create_uniform(b, ubo + (src0->regs[1]->iim_val * ptrsz) + 1);
697	} else {
698		base_lo = create_uniform_indirect(b, ubo, ir3_get_addr(ctx, src0, ptrsz));
699		base_hi = create_uniform_indirect(b, ubo + 1, ir3_get_addr(ctx, src0, ptrsz));
700	}
701
702	/* note: on 32bit gpu's base_hi is ignored and DCE'd */
703	addr = base_lo;
704
705	if (nir_src_is_const(intr->src[1])) {
706		off += nir_src_as_uint(intr->src[1]);
707	} else {
708		/* For load_ubo_indirect, second src is indirect offset: */
709		src1 = ir3_get_src(ctx, &intr->src[1])[0];
710
711		/* and add offset to addr: */
712		addr = ir3_ADD_S(b, addr, 0, src1, 0);
713	}
714
715	/* if offset is to large to encode in the ldg, split it out: */
716	if ((off + (intr->num_components * 4)) > 1024) {
717		/* split out the minimal amount to improve the odds that
718		 * cp can fit the immediate in the add.s instruction:
719		 */
720		unsigned off2 = off + (intr->num_components * 4) - 1024;
721		addr = ir3_ADD_S(b, addr, 0, create_immed(b, off2), 0);
722		off -= off2;
723	}
724
725	if (ptrsz == 2) {
726		struct ir3_instruction *carry;
727
728		/* handle 32b rollover, ie:
729		 *   if (addr < base_lo)
730		 *      base_hi++
731		 */
732		carry = ir3_CMPS_U(b, addr, 0, base_lo, 0);
733		carry->cat2.condition = IR3_COND_LT;
734		base_hi = ir3_ADD_S(b, base_hi, 0, carry, 0);
735
736		addr = ir3_create_collect(ctx, (struct ir3_instruction*[]){ addr, base_hi }, 2);
737	}
738
739	for (int i = 0; i < intr->num_components; i++) {
740		struct ir3_instruction *load =
741				ir3_LDG(b, addr, 0, create_immed(b, 1), 0);
742		load->cat6.type = TYPE_U32;
743		load->cat6.src_offset = off + i * 4;     /* byte offset */
744		dst[i] = load;
745	}
746}
747
748/* src[] = { block_index } */
749static void
750emit_intrinsic_ssbo_size(struct ir3_context *ctx, nir_intrinsic_instr *intr,
751		struct ir3_instruction **dst)
752{
753	/* SSBO size stored as a const starting at ssbo_sizes: */
754	unsigned blk_idx = nir_src_as_uint(intr->src[0]);
755	unsigned idx = regid(ctx->so->constbase.ssbo_sizes, 0) +
756		ctx->so->const_layout.ssbo_size.off[blk_idx];
757
758	debug_assert(ctx->so->const_layout.ssbo_size.mask & (1 << blk_idx));
759
760	dst[0] = create_uniform(ctx->block, idx);
761}
762
763/* src[] = { offset }. const_index[] = { base } */
764static void
765emit_intrinsic_load_shared(struct ir3_context *ctx, nir_intrinsic_instr *intr,
766		struct ir3_instruction **dst)
767{
768	struct ir3_block *b = ctx->block;
769	struct ir3_instruction *ldl, *offset;
770	unsigned base;
771
772	offset = ir3_get_src(ctx, &intr->src[0])[0];
773	base   = nir_intrinsic_base(intr);
774
775	ldl = ir3_LDL(b, offset, 0, create_immed(b, intr->num_components), 0);
776	ldl->cat6.src_offset = base;
777	ldl->cat6.type = utype_dst(intr->dest);
778	ldl->regs[0]->wrmask = MASK(intr->num_components);
779
780	ldl->barrier_class = IR3_BARRIER_SHARED_R;
781	ldl->barrier_conflict = IR3_BARRIER_SHARED_W;
782
783	ir3_split_dest(b, dst, ldl, 0, intr->num_components);
784}
785
786/* src[] = { value, offset }. const_index[] = { base, write_mask } */
787static void
788emit_intrinsic_store_shared(struct ir3_context *ctx, nir_intrinsic_instr *intr)
789{
790	struct ir3_block *b = ctx->block;
791	struct ir3_instruction *stl, *offset;
792	struct ir3_instruction * const *value;
793	unsigned base, wrmask;
794
795	value  = ir3_get_src(ctx, &intr->src[0]);
796	offset = ir3_get_src(ctx, &intr->src[1])[0];
797
798	base   = nir_intrinsic_base(intr);
799	wrmask = nir_intrinsic_write_mask(intr);
800
801	/* Combine groups of consecutive enabled channels in one write
802	 * message. We use ffs to find the first enabled channel and then ffs on
803	 * the bit-inverse, down-shifted writemask to determine the length of
804	 * the block of enabled bits.
805	 *
806	 * (trick stolen from i965's fs_visitor::nir_emit_cs_intrinsic())
807	 */
808	while (wrmask) {
809		unsigned first_component = ffs(wrmask) - 1;
810		unsigned length = ffs(~(wrmask >> first_component)) - 1;
811
812		stl = ir3_STL(b, offset, 0,
813			ir3_create_collect(ctx, &value[first_component], length), 0,
814			create_immed(b, length), 0);
815		stl->cat6.dst_offset = first_component + base;
816		stl->cat6.type = utype_src(intr->src[0]);
817		stl->barrier_class = IR3_BARRIER_SHARED_W;
818		stl->barrier_conflict = IR3_BARRIER_SHARED_R | IR3_BARRIER_SHARED_W;
819
820		array_insert(b, b->keeps, stl);
821
822		/* Clear the bits in the writemask that we just wrote, then try
823		 * again to see if more channels are left.
824		 */
825		wrmask &= (15 << (first_component + length));
826	}
827}
828
829/*
830 * CS shared variable atomic intrinsics
831 *
832 * All of the shared variable atomic memory operations read a value from
833 * memory, compute a new value using one of the operations below, write the
834 * new value to memory, and return the original value read.
835 *
836 * All operations take 2 sources except CompSwap that takes 3. These
837 * sources represent:
838 *
839 * 0: The offset into the shared variable storage region that the atomic
840 *    operation will operate on.
841 * 1: The data parameter to the atomic function (i.e. the value to add
842 *    in shared_atomic_add, etc).
843 * 2: For CompSwap only: the second data parameter.
844 */
845static struct ir3_instruction *
846emit_intrinsic_atomic_shared(struct ir3_context *ctx, nir_intrinsic_instr *intr)
847{
848	struct ir3_block *b = ctx->block;
849	struct ir3_instruction *atomic, *src0, *src1;
850	type_t type = TYPE_U32;
851
852	src0 = ir3_get_src(ctx, &intr->src[0])[0];   /* offset */
853	src1 = ir3_get_src(ctx, &intr->src[1])[0];   /* value */
854
855	switch (intr->intrinsic) {
856	case nir_intrinsic_shared_atomic_add:
857		atomic = ir3_ATOMIC_ADD(b, src0, 0, src1, 0);
858		break;
859	case nir_intrinsic_shared_atomic_imin:
860		atomic = ir3_ATOMIC_MIN(b, src0, 0, src1, 0);
861		type = TYPE_S32;
862		break;
863	case nir_intrinsic_shared_atomic_umin:
864		atomic = ir3_ATOMIC_MIN(b, src0, 0, src1, 0);
865		break;
866	case nir_intrinsic_shared_atomic_imax:
867		atomic = ir3_ATOMIC_MAX(b, src0, 0, src1, 0);
868		type = TYPE_S32;
869		break;
870	case nir_intrinsic_shared_atomic_umax:
871		atomic = ir3_ATOMIC_MAX(b, src0, 0, src1, 0);
872		break;
873	case nir_intrinsic_shared_atomic_and:
874		atomic = ir3_ATOMIC_AND(b, src0, 0, src1, 0);
875		break;
876	case nir_intrinsic_shared_atomic_or:
877		atomic = ir3_ATOMIC_OR(b, src0, 0, src1, 0);
878		break;
879	case nir_intrinsic_shared_atomic_xor:
880		atomic = ir3_ATOMIC_XOR(b, src0, 0, src1, 0);
881		break;
882	case nir_intrinsic_shared_atomic_exchange:
883		atomic = ir3_ATOMIC_XCHG(b, src0, 0, src1, 0);
884		break;
885	case nir_intrinsic_shared_atomic_comp_swap:
886		/* for cmpxchg, src1 is [ui]vec2(data, compare): */
887		src1 = ir3_create_collect(ctx, (struct ir3_instruction*[]){
888			ir3_get_src(ctx, &intr->src[2])[0],
889			src1,
890		}, 2);
891		atomic = ir3_ATOMIC_CMPXCHG(b, src0, 0, src1, 0);
892		break;
893	default:
894		unreachable("boo");
895	}
896
897	atomic->cat6.iim_val = 1;
898	atomic->cat6.d = 1;
899	atomic->cat6.type = type;
900	atomic->barrier_class = IR3_BARRIER_SHARED_W;
901	atomic->barrier_conflict = IR3_BARRIER_SHARED_R | IR3_BARRIER_SHARED_W;
902
903	/* even if nothing consume the result, we can't DCE the instruction: */
904	array_insert(b, b->keeps, atomic);
905
906	return atomic;
907}
908
909/* TODO handle actual indirect/dynamic case.. which is going to be weird
910 * to handle with the image_mapping table..
911 */
912static struct ir3_instruction *
913get_image_samp_tex_src(struct ir3_context *ctx, nir_intrinsic_instr *intr)
914{
915	unsigned slot = ir3_get_image_slot(nir_src_as_deref(intr->src[0]));
916	unsigned tex_idx = ir3_image_to_tex(&ctx->so->image_mapping, slot);
917	struct ir3_instruction *texture, *sampler;
918
919	texture = create_immed_typed(ctx->block, tex_idx, TYPE_U16);
920	sampler = create_immed_typed(ctx->block, tex_idx, TYPE_U16);
921
922	return ir3_create_collect(ctx, (struct ir3_instruction*[]){
923		sampler,
924		texture,
925	}, 2);
926}
927
928/* src[] = { deref, coord, sample_index }. const_index[] = {} */
929static void
930emit_intrinsic_load_image(struct ir3_context *ctx, nir_intrinsic_instr *intr,
931		struct ir3_instruction **dst)
932{
933	struct ir3_block *b = ctx->block;
934	const nir_variable *var = nir_intrinsic_get_var(intr, 0);
935	struct ir3_instruction *samp_tex = get_image_samp_tex_src(ctx, intr);
936	struct ir3_instruction *sam;
937	struct ir3_instruction * const *src0 = ir3_get_src(ctx, &intr->src[1]);
938	struct ir3_instruction *coords[4];
939	unsigned flags, ncoords = ir3_get_image_coords(var, &flags);
940	type_t type = ir3_get_image_type(var);
941
942	/* hmm, this seems a bit odd, but it is what blob does and (at least
943	 * a5xx) just faults on bogus addresses otherwise:
944	 */
945	if (flags & IR3_INSTR_3D) {
946		flags &= ~IR3_INSTR_3D;
947		flags |= IR3_INSTR_A;
948	}
949
950	for (unsigned i = 0; i < ncoords; i++)
951		coords[i] = src0[i];
952
953	if (ncoords == 1)
954		coords[ncoords++] = create_immed(b, 0);
955
956	sam = ir3_SAM(b, OPC_ISAM, type, 0b1111, flags,
957			samp_tex, ir3_create_collect(ctx, coords, ncoords), NULL);
958
959	sam->barrier_class = IR3_BARRIER_IMAGE_R;
960	sam->barrier_conflict = IR3_BARRIER_IMAGE_W;
961
962	ir3_split_dest(b, dst, sam, 0, 4);
963}
964
965static void
966emit_intrinsic_image_size(struct ir3_context *ctx, nir_intrinsic_instr *intr,
967		struct ir3_instruction **dst)
968{
969	struct ir3_block *b = ctx->block;
970	const nir_variable *var = nir_intrinsic_get_var(intr, 0);
971	struct ir3_instruction *samp_tex = get_image_samp_tex_src(ctx, intr);
972	struct ir3_instruction *sam, *lod;
973	unsigned flags, ncoords = ir3_get_image_coords(var, &flags);
974
975	lod = create_immed(b, 0);
976	sam = ir3_SAM(b, OPC_GETSIZE, TYPE_U32, 0b1111, flags,
977			samp_tex, lod, NULL);
978
979	/* Array size actually ends up in .w rather than .z. This doesn't
980	 * matter for miplevel 0, but for higher mips the value in z is
981	 * minified whereas w stays. Also, the value in TEX_CONST_3_DEPTH is
982	 * returned, which means that we have to add 1 to it for arrays for
983	 * a3xx.
984	 *
985	 * Note use a temporary dst and then copy, since the size of the dst
986	 * array that is passed in is based on nir's understanding of the
987	 * result size, not the hardware's
988	 */
989	struct ir3_instruction *tmp[4];
990
991	ir3_split_dest(b, tmp, sam, 0, 4);
992
993	/* get_size instruction returns size in bytes instead of texels
994	 * for imageBuffer, so we need to divide it by the pixel size
995	 * of the image format.
996	 *
997	 * TODO: This is at least true on a5xx. Check other gens.
998	 */
999	enum glsl_sampler_dim dim =
1000		glsl_get_sampler_dim(glsl_without_array(var->type));
1001	if (dim == GLSL_SAMPLER_DIM_BUF) {
1002		/* Since all the possible values the divisor can take are
1003		 * power-of-two (4, 8, or 16), the division is implemented
1004		 * as a shift-right.
1005		 * During shader setup, the log2 of the image format's
1006		 * bytes-per-pixel should have been emitted in 2nd slot of
1007		 * image_dims. See ir3_shader::emit_image_dims().
1008		 */
1009		unsigned cb = regid(ctx->so->constbase.image_dims, 0) +
1010			ctx->so->const_layout.image_dims.off[var->data.driver_location];
1011		struct ir3_instruction *aux = create_uniform(b, cb + 1);
1012
1013		tmp[0] = ir3_SHR_B(b, tmp[0], 0, aux, 0);
1014	}
1015
1016	for (unsigned i = 0; i < ncoords; i++)
1017		dst[i] = tmp[i];
1018
1019	if (flags & IR3_INSTR_A) {
1020		if (ctx->compiler->levels_add_one) {
1021			dst[ncoords-1] = ir3_ADD_U(b, tmp[3], 0, create_immed(b, 1), 0);
1022		} else {
1023			dst[ncoords-1] = ir3_MOV(b, tmp[3], TYPE_U32);
1024		}
1025	}
1026}
1027
1028static void
1029emit_intrinsic_barrier(struct ir3_context *ctx, nir_intrinsic_instr *intr)
1030{
1031	struct ir3_block *b = ctx->block;
1032	struct ir3_instruction *barrier;
1033
1034	switch (intr->intrinsic) {
1035	case nir_intrinsic_barrier:
1036		barrier = ir3_BAR(b);
1037		barrier->cat7.g = true;
1038		barrier->cat7.l = true;
1039		barrier->flags = IR3_INSTR_SS | IR3_INSTR_SY;
1040		barrier->barrier_class = IR3_BARRIER_EVERYTHING;
1041		break;
1042	case nir_intrinsic_memory_barrier:
1043		barrier = ir3_FENCE(b);
1044		barrier->cat7.g = true;
1045		barrier->cat7.r = true;
1046		barrier->cat7.w = true;
1047		barrier->cat7.l = true;
1048		barrier->barrier_class = IR3_BARRIER_IMAGE_W |
1049				IR3_BARRIER_BUFFER_W;
1050		barrier->barrier_conflict =
1051				IR3_BARRIER_IMAGE_R | IR3_BARRIER_IMAGE_W |
1052				IR3_BARRIER_BUFFER_R | IR3_BARRIER_BUFFER_W;
1053		break;
1054	case nir_intrinsic_memory_barrier_atomic_counter:
1055	case nir_intrinsic_memory_barrier_buffer:
1056		barrier = ir3_FENCE(b);
1057		barrier->cat7.g = true;
1058		barrier->cat7.r = true;
1059		barrier->cat7.w = true;
1060		barrier->barrier_class = IR3_BARRIER_BUFFER_W;
1061		barrier->barrier_conflict = IR3_BARRIER_BUFFER_R |
1062				IR3_BARRIER_BUFFER_W;
1063		break;
1064	case nir_intrinsic_memory_barrier_image:
1065		// TODO double check if this should have .g set
1066		barrier = ir3_FENCE(b);
1067		barrier->cat7.g = true;
1068		barrier->cat7.r = true;
1069		barrier->cat7.w = true;
1070		barrier->barrier_class = IR3_BARRIER_IMAGE_W;
1071		barrier->barrier_conflict = IR3_BARRIER_IMAGE_R |
1072				IR3_BARRIER_IMAGE_W;
1073		break;
1074	case nir_intrinsic_memory_barrier_shared:
1075		barrier = ir3_FENCE(b);
1076		barrier->cat7.g = true;
1077		barrier->cat7.l = true;
1078		barrier->cat7.r = true;
1079		barrier->cat7.w = true;
1080		barrier->barrier_class = IR3_BARRIER_SHARED_W;
1081		barrier->barrier_conflict = IR3_BARRIER_SHARED_R |
1082				IR3_BARRIER_SHARED_W;
1083		break;
1084	case nir_intrinsic_group_memory_barrier:
1085		barrier = ir3_FENCE(b);
1086		barrier->cat7.g = true;
1087		barrier->cat7.l = true;
1088		barrier->cat7.r = true;
1089		barrier->cat7.w = true;
1090		barrier->barrier_class = IR3_BARRIER_SHARED_W |
1091				IR3_BARRIER_IMAGE_W |
1092				IR3_BARRIER_BUFFER_W;
1093		barrier->barrier_conflict =
1094				IR3_BARRIER_SHARED_R | IR3_BARRIER_SHARED_W |
1095				IR3_BARRIER_IMAGE_R | IR3_BARRIER_IMAGE_W |
1096				IR3_BARRIER_BUFFER_R | IR3_BARRIER_BUFFER_W;
1097		break;
1098	default:
1099		unreachable("boo");
1100	}
1101
1102	/* make sure barrier doesn't get DCE'd */
1103	array_insert(b, b->keeps, barrier);
1104}
1105
1106static void add_sysval_input_compmask(struct ir3_context *ctx,
1107		gl_system_value slot, unsigned compmask,
1108		struct ir3_instruction *instr)
1109{
1110	struct ir3_shader_variant *so = ctx->so;
1111	unsigned r = regid(so->inputs_count, 0);
1112	unsigned n = so->inputs_count++;
1113
1114	so->inputs[n].sysval = true;
1115	so->inputs[n].slot = slot;
1116	so->inputs[n].compmask = compmask;
1117	so->inputs[n].regid = r;
1118	so->inputs[n].interpolate = INTERP_MODE_FLAT;
1119	so->total_in++;
1120
1121	ctx->ir->ninputs = MAX2(ctx->ir->ninputs, r + 1);
1122	ctx->ir->inputs[r] = instr;
1123}
1124
1125static void add_sysval_input(struct ir3_context *ctx, gl_system_value slot,
1126		struct ir3_instruction *instr)
1127{
1128	add_sysval_input_compmask(ctx, slot, 0x1, instr);
1129}
1130
1131static struct ir3_instruction *
1132get_barycentric_centroid(struct ir3_context *ctx)
1133{
1134	if (!ctx->ij_centroid) {
1135		struct ir3_instruction *xy[2];
1136		struct ir3_instruction *ij;
1137
1138		ij = create_input_compmask(ctx, 0, 0x3);
1139		ir3_split_dest(ctx->block, xy, ij, 0, 2);
1140
1141		ctx->ij_centroid = ir3_create_collect(ctx, xy, 2);
1142
1143		add_sysval_input_compmask(ctx,
1144				SYSTEM_VALUE_BARYCENTRIC_CENTROID,
1145				0x3, ij);
1146	}
1147
1148	return ctx->ij_centroid;
1149}
1150
1151static struct ir3_instruction *
1152get_barycentric_sample(struct ir3_context *ctx)
1153{
1154	if (!ctx->ij_sample) {
1155		struct ir3_instruction *xy[2];
1156		struct ir3_instruction *ij;
1157
1158		ij = create_input_compmask(ctx, 0, 0x3);
1159		ir3_split_dest(ctx->block, xy, ij, 0, 2);
1160
1161		ctx->ij_sample = ir3_create_collect(ctx, xy, 2);
1162
1163		add_sysval_input_compmask(ctx,
1164				SYSTEM_VALUE_BARYCENTRIC_SAMPLE,
1165				0x3, ij);
1166	}
1167
1168	return ctx->ij_sample;
1169}
1170
1171static struct ir3_instruction  *
1172get_barycentric_pixel(struct ir3_context *ctx)
1173{
1174	/* TODO when tgsi_to_nir supports "new-style" FS inputs switch
1175	 * this to create ij_pixel only on demand:
1176	 */
1177	return ctx->ij_pixel;
1178}
1179
1180static struct ir3_instruction *
1181get_frag_coord(struct ir3_context *ctx)
1182{
1183	if (!ctx->frag_coord) {
1184		struct ir3_block *b = ctx->block;
1185		struct ir3_instruction *xyzw[4];
1186		struct ir3_instruction *hw_frag_coord;
1187
1188		hw_frag_coord = create_input_compmask(ctx, 0, 0xf);
1189		ir3_split_dest(ctx->block, xyzw, hw_frag_coord, 0, 4);
1190
1191		/* for frag_coord.xy, we get unsigned values.. we need
1192		 * to subtract (integer) 8 and divide by 16 (right-
1193		 * shift by 4) then convert to float:
1194		 *
1195		 *    sub.s tmp, src, 8
1196		 *    shr.b tmp, tmp, 4
1197		 *    mov.u32f32 dst, tmp
1198		 *
1199		 */
1200		for (int i = 0; i < 2; i++) {
1201			xyzw[i] = ir3_SUB_S(b, xyzw[i], 0,
1202					create_immed(b, 8), 0);
1203			xyzw[i] = ir3_SHR_B(b, xyzw[i], 0,
1204					create_immed(b, 4), 0);
1205			xyzw[i] = ir3_COV(b, xyzw[i], TYPE_U32, TYPE_F32);
1206		}
1207
1208		ctx->frag_coord = ir3_create_collect(ctx, xyzw, 4);
1209
1210		add_sysval_input_compmask(ctx,
1211				SYSTEM_VALUE_FRAG_COORD,
1212				0xf, hw_frag_coord);
1213
1214		ctx->so->frag_coord = true;
1215	}
1216
1217	return ctx->frag_coord;
1218}
1219
1220static void
1221emit_intrinsic(struct ir3_context *ctx, nir_intrinsic_instr *intr)
1222{
1223	const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic];
1224	struct ir3_instruction **dst;
1225	struct ir3_instruction * const *src;
1226	struct ir3_block *b = ctx->block;
1227	int idx, comp;
1228
1229	if (info->has_dest) {
1230		unsigned n = nir_intrinsic_dest_components(intr);
1231		dst = ir3_get_dst(ctx, &intr->dest, n);
1232	} else {
1233		dst = NULL;
1234	}
1235
1236	switch (intr->intrinsic) {
1237	case nir_intrinsic_load_uniform:
1238		idx = nir_intrinsic_base(intr);
1239		if (nir_src_is_const(intr->src[0])) {
1240			idx += nir_src_as_uint(intr->src[0]);
1241			for (int i = 0; i < intr->num_components; i++) {
1242				dst[i] = create_uniform(b, idx + i);
1243			}
1244		} else {
1245			src = ir3_get_src(ctx, &intr->src[0]);
1246			for (int i = 0; i < intr->num_components; i++) {
1247				dst[i] = create_uniform_indirect(b, idx + i,
1248						ir3_get_addr(ctx, src[0], 1));
1249			}
1250			/* NOTE: if relative addressing is used, we set
1251			 * constlen in the compiler (to worst-case value)
1252			 * since we don't know in the assembler what the max
1253			 * addr reg value can be:
1254			 */
1255			ctx->so->constlen = ctx->s->num_uniforms;
1256		}
1257		break;
1258	case nir_intrinsic_load_ubo:
1259		emit_intrinsic_load_ubo(ctx, intr, dst);
1260		break;
1261	case nir_intrinsic_load_frag_coord:
1262		ir3_split_dest(b, dst, get_frag_coord(ctx), 0, 4);
1263		break;
1264	case nir_intrinsic_load_sample_pos_from_id: {
1265		/* NOTE: blob seems to always use TYPE_F16 and then cov.f16f32,
1266		 * but that doesn't seem necessary.
1267		 */
1268		struct ir3_instruction *offset =
1269			ir3_RGETPOS(b, ir3_get_src(ctx, &intr->src[0])[0], 0);
1270		offset->regs[0]->wrmask = 0x3;
1271		offset->cat5.type = TYPE_F32;
1272
1273		ir3_split_dest(b, dst, offset, 0, 2);
1274
1275		break;
1276	}
1277	case nir_intrinsic_load_size_ir3:
1278		if (!ctx->ij_size) {
1279			ctx->ij_size = create_input(ctx, 0);
1280
1281			add_sysval_input(ctx, SYSTEM_VALUE_BARYCENTRIC_SIZE,
1282					ctx->ij_size);
1283		}
1284		dst[0] = ctx->ij_size;
1285		break;
1286	case nir_intrinsic_load_barycentric_centroid:
1287		ir3_split_dest(b, dst, get_barycentric_centroid(ctx), 0, 2);
1288		break;
1289	case nir_intrinsic_load_barycentric_sample:
1290		if (ctx->so->key.msaa) {
1291			ir3_split_dest(b, dst, get_barycentric_sample(ctx), 0, 2);
1292		} else {
1293			ir3_split_dest(b, dst, get_barycentric_pixel(ctx), 0, 2);
1294		}
1295		break;
1296	case nir_intrinsic_load_barycentric_pixel:
1297		ir3_split_dest(b, dst, get_barycentric_pixel(ctx), 0, 2);
1298		break;
1299	case nir_intrinsic_load_interpolated_input:
1300		idx = nir_intrinsic_base(intr);
1301		comp = nir_intrinsic_component(intr);
1302		src = ir3_get_src(ctx, &intr->src[0]);
1303		if (nir_src_is_const(intr->src[1])) {
1304			struct ir3_instruction *coord = ir3_create_collect(ctx, src, 2);
1305			idx += nir_src_as_uint(intr->src[1]);
1306			for (int i = 0; i < intr->num_components; i++) {
1307				unsigned inloc = idx * 4 + i + comp;
1308				if (ctx->so->inputs[idx].bary &&
1309						!ctx->so->inputs[idx].use_ldlv) {
1310					dst[i] = ir3_BARY_F(b, create_immed(b, inloc), 0, coord, 0);
1311				} else {
1312					/* for non-varyings use the pre-setup input, since
1313					 * that is easier than mapping things back to a
1314					 * nir_variable to figure out what it is.
1315					 */
1316					dst[i] = ctx->ir->inputs[inloc];
1317				}
1318			}
1319		} else {
1320			ir3_context_error(ctx, "unhandled");
1321		}
1322		break;
1323	case nir_intrinsic_load_input:
1324		idx = nir_intrinsic_base(intr);
1325		comp = nir_intrinsic_component(intr);
1326		if (nir_src_is_const(intr->src[0])) {
1327			idx += nir_src_as_uint(intr->src[0]);
1328			for (int i = 0; i < intr->num_components; i++) {
1329				unsigned n = idx * 4 + i + comp;
1330				dst[i] = ctx->ir->inputs[n];
1331				compile_assert(ctx, ctx->ir->inputs[n]);
1332			}
1333		} else {
1334			src = ir3_get_src(ctx, &intr->src[0]);
1335			struct ir3_instruction *collect =
1336					ir3_create_collect(ctx, ctx->ir->inputs, ctx->ir->ninputs);
1337			struct ir3_instruction *addr = ir3_get_addr(ctx, src[0], 4);
1338			for (int i = 0; i < intr->num_components; i++) {
1339				unsigned n = idx * 4 + i + comp;
1340				dst[i] = create_indirect_load(ctx, ctx->ir->ninputs,
1341						n, addr, collect);
1342			}
1343		}
1344		break;
1345	/* All SSBO intrinsics should have been lowered by 'lower_io_offsets'
1346	 * pass and replaced by an ir3-specifc version that adds the
1347	 * dword-offset in the last source.
1348	 */
1349	case nir_intrinsic_load_ssbo_ir3:
1350		ctx->funcs->emit_intrinsic_load_ssbo(ctx, intr, dst);
1351		break;
1352	case nir_intrinsic_store_ssbo_ir3:
1353		if ((ctx->so->type == MESA_SHADER_FRAGMENT) &&
1354				!ctx->s->info.fs.early_fragment_tests)
1355			ctx->so->no_earlyz = true;
1356		ctx->funcs->emit_intrinsic_store_ssbo(ctx, intr);
1357		break;
1358	case nir_intrinsic_get_buffer_size:
1359		emit_intrinsic_ssbo_size(ctx, intr, dst);
1360		break;
1361	case nir_intrinsic_ssbo_atomic_add_ir3:
1362	case nir_intrinsic_ssbo_atomic_imin_ir3:
1363	case nir_intrinsic_ssbo_atomic_umin_ir3:
1364	case nir_intrinsic_ssbo_atomic_imax_ir3:
1365	case nir_intrinsic_ssbo_atomic_umax_ir3:
1366	case nir_intrinsic_ssbo_atomic_and_ir3:
1367	case nir_intrinsic_ssbo_atomic_or_ir3:
1368	case nir_intrinsic_ssbo_atomic_xor_ir3:
1369	case nir_intrinsic_ssbo_atomic_exchange_ir3:
1370	case nir_intrinsic_ssbo_atomic_comp_swap_ir3:
1371		if ((ctx->so->type == MESA_SHADER_FRAGMENT) &&
1372				!ctx->s->info.fs.early_fragment_tests)
1373			ctx->so->no_earlyz = true;
1374		dst[0] = ctx->funcs->emit_intrinsic_atomic_ssbo(ctx, intr);
1375		break;
1376	case nir_intrinsic_load_shared:
1377		emit_intrinsic_load_shared(ctx, intr, dst);
1378		break;
1379	case nir_intrinsic_store_shared:
1380		emit_intrinsic_store_shared(ctx, intr);
1381		break;
1382	case nir_intrinsic_shared_atomic_add:
1383	case nir_intrinsic_shared_atomic_imin:
1384	case nir_intrinsic_shared_atomic_umin:
1385	case nir_intrinsic_shared_atomic_imax:
1386	case nir_intrinsic_shared_atomic_umax:
1387	case nir_intrinsic_shared_atomic_and:
1388	case nir_intrinsic_shared_atomic_or:
1389	case nir_intrinsic_shared_atomic_xor:
1390	case nir_intrinsic_shared_atomic_exchange:
1391	case nir_intrinsic_shared_atomic_comp_swap:
1392		dst[0] = emit_intrinsic_atomic_shared(ctx, intr);
1393		break;
1394	case nir_intrinsic_image_deref_load:
1395		emit_intrinsic_load_image(ctx, intr, dst);
1396		break;
1397	case nir_intrinsic_image_deref_store:
1398		if ((ctx->so->type == MESA_SHADER_FRAGMENT) &&
1399				!ctx->s->info.fs.early_fragment_tests)
1400			ctx->so->no_earlyz = true;
1401		ctx->funcs->emit_intrinsic_store_image(ctx, intr);
1402		break;
1403	case nir_intrinsic_image_deref_size:
1404		emit_intrinsic_image_size(ctx, intr, dst);
1405		break;
1406	case nir_intrinsic_image_deref_atomic_add:
1407	case nir_intrinsic_image_deref_atomic_min:
1408	case nir_intrinsic_image_deref_atomic_max:
1409	case nir_intrinsic_image_deref_atomic_and:
1410	case nir_intrinsic_image_deref_atomic_or:
1411	case nir_intrinsic_image_deref_atomic_xor:
1412	case nir_intrinsic_image_deref_atomic_exchange:
1413	case nir_intrinsic_image_deref_atomic_comp_swap:
1414		if ((ctx->so->type == MESA_SHADER_FRAGMENT) &&
1415				!ctx->s->info.fs.early_fragment_tests)
1416			ctx->so->no_earlyz = true;
1417		dst[0] = ctx->funcs->emit_intrinsic_atomic_image(ctx, intr);
1418		break;
1419	case nir_intrinsic_barrier:
1420	case nir_intrinsic_memory_barrier:
1421	case nir_intrinsic_group_memory_barrier:
1422	case nir_intrinsic_memory_barrier_atomic_counter:
1423	case nir_intrinsic_memory_barrier_buffer:
1424	case nir_intrinsic_memory_barrier_image:
1425	case nir_intrinsic_memory_barrier_shared:
1426		emit_intrinsic_barrier(ctx, intr);
1427		/* note that blk ptr no longer valid, make that obvious: */
1428		b = NULL;
1429		break;
1430	case nir_intrinsic_store_output:
1431		idx = nir_intrinsic_base(intr);
1432		comp = nir_intrinsic_component(intr);
1433		compile_assert(ctx, nir_src_is_const(intr->src[1]));
1434		idx += nir_src_as_uint(intr->src[1]);
1435
1436		src = ir3_get_src(ctx, &intr->src[0]);
1437		for (int i = 0; i < intr->num_components; i++) {
1438			unsigned n = idx * 4 + i + comp;
1439			ctx->ir->outputs[n] = src[i];
1440		}
1441		break;
1442	case nir_intrinsic_load_base_vertex:
1443	case nir_intrinsic_load_first_vertex:
1444		if (!ctx->basevertex) {
1445			ctx->basevertex = create_driver_param(ctx, IR3_DP_VTXID_BASE);
1446			add_sysval_input(ctx, SYSTEM_VALUE_FIRST_VERTEX, ctx->basevertex);
1447		}
1448		dst[0] = ctx->basevertex;
1449		break;
1450	case nir_intrinsic_load_vertex_id_zero_base:
1451	case nir_intrinsic_load_vertex_id:
1452		if (!ctx->vertex_id) {
1453			gl_system_value sv = (intr->intrinsic == nir_intrinsic_load_vertex_id) ?
1454				SYSTEM_VALUE_VERTEX_ID : SYSTEM_VALUE_VERTEX_ID_ZERO_BASE;
1455			ctx->vertex_id = create_input(ctx, 0);
1456			add_sysval_input(ctx, sv, ctx->vertex_id);
1457		}
1458		dst[0] = ctx->vertex_id;
1459		break;
1460	case nir_intrinsic_load_instance_id:
1461		if (!ctx->instance_id) {
1462			ctx->instance_id = create_input(ctx, 0);
1463			add_sysval_input(ctx, SYSTEM_VALUE_INSTANCE_ID,
1464					ctx->instance_id);
1465		}
1466		dst[0] = ctx->instance_id;
1467		break;
1468	case nir_intrinsic_load_sample_id:
1469		ctx->so->per_samp = true;
1470		/* fall-thru */
1471	case nir_intrinsic_load_sample_id_no_per_sample:
1472		if (!ctx->samp_id) {
1473			ctx->samp_id = create_input(ctx, 0);
1474			ctx->samp_id->regs[0]->flags |= IR3_REG_HALF;
1475			add_sysval_input(ctx, SYSTEM_VALUE_SAMPLE_ID,
1476					ctx->samp_id);
1477		}
1478		dst[0] = ir3_COV(b, ctx->samp_id, TYPE_U16, TYPE_U32);
1479		break;
1480	case nir_intrinsic_load_sample_mask_in:
1481		if (!ctx->samp_mask_in) {
1482			ctx->samp_mask_in = create_input(ctx, 0);
1483			add_sysval_input(ctx, SYSTEM_VALUE_SAMPLE_MASK_IN,
1484					ctx->samp_mask_in);
1485		}
1486		dst[0] = ctx->samp_mask_in;
1487		break;
1488	case nir_intrinsic_load_user_clip_plane:
1489		idx = nir_intrinsic_ucp_id(intr);
1490		for (int i = 0; i < intr->num_components; i++) {
1491			unsigned n = idx * 4 + i;
1492			dst[i] = create_driver_param(ctx, IR3_DP_UCP0_X + n);
1493		}
1494		break;
1495	case nir_intrinsic_load_front_face:
1496		if (!ctx->frag_face) {
1497			ctx->so->frag_face = true;
1498			ctx->frag_face = create_input(ctx, 0);
1499			add_sysval_input(ctx, SYSTEM_VALUE_FRONT_FACE, ctx->frag_face);
1500			ctx->frag_face->regs[0]->flags |= IR3_REG_HALF;
1501		}
1502		/* for fragface, we get -1 for back and 0 for front. However this is
1503		 * the inverse of what nir expects (where ~0 is true).
1504		 */
1505		dst[0] = ir3_COV(b, ctx->frag_face, TYPE_S16, TYPE_S32);
1506		dst[0] = ir3_NOT_B(b, dst[0], 0);
1507		break;
1508	case nir_intrinsic_load_local_invocation_id:
1509		if (!ctx->local_invocation_id) {
1510			ctx->local_invocation_id = create_input_compmask(ctx, 0, 0x7);
1511			add_sysval_input_compmask(ctx, SYSTEM_VALUE_LOCAL_INVOCATION_ID,
1512					0x7, ctx->local_invocation_id);
1513		}
1514		ir3_split_dest(b, dst, ctx->local_invocation_id, 0, 3);
1515		break;
1516	case nir_intrinsic_load_work_group_id:
1517		if (!ctx->work_group_id) {
1518			ctx->work_group_id = create_input_compmask(ctx, 0, 0x7);
1519			add_sysval_input_compmask(ctx, SYSTEM_VALUE_WORK_GROUP_ID,
1520					0x7, ctx->work_group_id);
1521			ctx->work_group_id->regs[0]->flags |= IR3_REG_HIGH;
1522		}
1523		ir3_split_dest(b, dst, ctx->work_group_id, 0, 3);
1524		break;
1525	case nir_intrinsic_load_num_work_groups:
1526		for (int i = 0; i < intr->num_components; i++) {
1527			dst[i] = create_driver_param(ctx, IR3_DP_NUM_WORK_GROUPS_X + i);
1528		}
1529		break;
1530	case nir_intrinsic_load_local_group_size:
1531		for (int i = 0; i < intr->num_components; i++) {
1532			dst[i] = create_driver_param(ctx, IR3_DP_LOCAL_GROUP_SIZE_X + i);
1533		}
1534		break;
1535	case nir_intrinsic_discard_if:
1536	case nir_intrinsic_discard: {
1537		struct ir3_instruction *cond, *kill;
1538
1539		if (intr->intrinsic == nir_intrinsic_discard_if) {
1540			/* conditional discard: */
1541			src = ir3_get_src(ctx, &intr->src[0]);
1542			cond = ir3_b2n(b, src[0]);
1543		} else {
1544			/* unconditional discard: */
1545			cond = create_immed(b, 1);
1546		}
1547
1548		/* NOTE: only cmps.*.* can write p0.x: */
1549		cond = ir3_CMPS_S(b, cond, 0, create_immed(b, 0), 0);
1550		cond->cat2.condition = IR3_COND_NE;
1551
1552		/* condition always goes in predicate register: */
1553		cond->regs[0]->num = regid(REG_P0, 0);
1554
1555		kill = ir3_KILL(b, cond, 0);
1556		array_insert(ctx->ir, ctx->ir->predicates, kill);
1557
1558		array_insert(b, b->keeps, kill);
1559		ctx->so->no_earlyz = true;
1560
1561		break;
1562	}
1563	default:
1564		ir3_context_error(ctx, "Unhandled intrinsic type: %s\n",
1565				nir_intrinsic_infos[intr->intrinsic].name);
1566		break;
1567	}
1568
1569	if (info->has_dest)
1570		ir3_put_dst(ctx, &intr->dest);
1571}
1572
1573static void
1574emit_load_const(struct ir3_context *ctx, nir_load_const_instr *instr)
1575{
1576	struct ir3_instruction **dst = ir3_get_dst_ssa(ctx, &instr->def,
1577			instr->def.num_components);
1578	type_t type = (instr->def.bit_size < 32) ? TYPE_U16 : TYPE_U32;
1579
1580	for (int i = 0; i < instr->def.num_components; i++)
1581		dst[i] = create_immed_typed(ctx->block, instr->value[i].u32, type);
1582}
1583
1584static void
1585emit_undef(struct ir3_context *ctx, nir_ssa_undef_instr *undef)
1586{
1587	struct ir3_instruction **dst = ir3_get_dst_ssa(ctx, &undef->def,
1588			undef->def.num_components);
1589	type_t type = (undef->def.bit_size < 32) ? TYPE_U16 : TYPE_U32;
1590
1591	/* backend doesn't want undefined instructions, so just plug
1592	 * in 0.0..
1593	 */
1594	for (int i = 0; i < undef->def.num_components; i++)
1595		dst[i] = create_immed_typed(ctx->block, fui(0.0), type);
1596}
1597
1598/*
1599 * texture fetch/sample instructions:
1600 */
1601
1602static void
1603tex_info(nir_tex_instr *tex, unsigned *flagsp, unsigned *coordsp)
1604{
1605	unsigned coords, flags = 0;
1606
1607	/* note: would use tex->coord_components.. except txs.. also,
1608	 * since array index goes after shadow ref, we don't want to
1609	 * count it:
1610	 */
1611	switch (tex->sampler_dim) {
1612	case GLSL_SAMPLER_DIM_1D:
1613	case GLSL_SAMPLER_DIM_BUF:
1614		coords = 1;
1615		break;
1616	case GLSL_SAMPLER_DIM_2D:
1617	case GLSL_SAMPLER_DIM_RECT:
1618	case GLSL_SAMPLER_DIM_EXTERNAL:
1619	case GLSL_SAMPLER_DIM_MS:
1620		coords = 2;
1621		break;
1622	case GLSL_SAMPLER_DIM_3D:
1623	case GLSL_SAMPLER_DIM_CUBE:
1624		coords = 3;
1625		flags |= IR3_INSTR_3D;
1626		break;
1627	default:
1628		unreachable("bad sampler_dim");
1629	}
1630
1631	if (tex->is_shadow && tex->op != nir_texop_lod)
1632		flags |= IR3_INSTR_S;
1633
1634	if (tex->is_array && tex->op != nir_texop_lod)
1635		flags |= IR3_INSTR_A;
1636
1637	*flagsp = flags;
1638	*coordsp = coords;
1639}
1640
1641/* Gets the sampler/texture idx as a hvec2.  Which could either be dynamic
1642 * or immediate (in which case it will get lowered later to a non .s2en
1643 * version of the tex instruction which encode tex/samp as immediates:
1644 */
1645static struct ir3_instruction *
1646get_tex_samp_tex_src(struct ir3_context *ctx, nir_tex_instr *tex)
1647{
1648	int texture_idx = nir_tex_instr_src_index(tex, nir_tex_src_texture_offset);
1649	int sampler_idx = nir_tex_instr_src_index(tex, nir_tex_src_sampler_offset);
1650	struct ir3_instruction *texture, *sampler;
1651
1652	if (texture_idx >= 0) {
1653		texture = ir3_get_src(ctx, &tex->src[texture_idx].src)[0];
1654		texture = ir3_COV(ctx->block, texture, TYPE_U32, TYPE_U16);
1655	} else {
1656		/* TODO what to do for dynamic case? I guess we only need the
1657		 * max index for astc srgb workaround so maybe not a problem
1658		 * to worry about if we don't enable indirect samplers for
1659		 * a4xx?
1660		 */
1661		ctx->max_texture_index = MAX2(ctx->max_texture_index, tex->texture_index);
1662		texture = create_immed_typed(ctx->block, tex->texture_index, TYPE_U16);
1663	}
1664
1665	if (sampler_idx >= 0) {
1666		sampler = ir3_get_src(ctx, &tex->src[sampler_idx].src)[0];
1667		sampler = ir3_COV(ctx->block, sampler, TYPE_U32, TYPE_U16);
1668	} else {
1669		sampler = create_immed_typed(ctx->block, tex->sampler_index, TYPE_U16);
1670	}
1671
1672	return ir3_create_collect(ctx, (struct ir3_instruction*[]){
1673		sampler,
1674		texture,
1675	}, 2);
1676}
1677
1678static void
1679emit_tex(struct ir3_context *ctx, nir_tex_instr *tex)
1680{
1681	struct ir3_block *b = ctx->block;
1682	struct ir3_instruction **dst, *sam, *src0[12], *src1[4];
1683	struct ir3_instruction * const *coord, * const *off, * const *ddx, * const *ddy;
1684	struct ir3_instruction *lod, *compare, *proj, *sample_index;
1685	bool has_bias = false, has_lod = false, has_proj = false, has_off = false;
1686	unsigned i, coords, flags, ncomp;
1687	unsigned nsrc0 = 0, nsrc1 = 0;
1688	type_t type;
1689	opc_t opc = 0;
1690
1691	ncomp = nir_dest_num_components(tex->dest);
1692
1693	coord = off = ddx = ddy = NULL;
1694	lod = proj = compare = sample_index = NULL;
1695
1696	dst = ir3_get_dst(ctx, &tex->dest, ncomp);
1697
1698	for (unsigned i = 0; i < tex->num_srcs; i++) {
1699		switch (tex->src[i].src_type) {
1700		case nir_tex_src_coord:
1701			coord = ir3_get_src(ctx, &tex->src[i].src);
1702			break;
1703		case nir_tex_src_bias:
1704			lod = ir3_get_src(ctx, &tex->src[i].src)[0];
1705			has_bias = true;
1706			break;
1707		case nir_tex_src_lod:
1708			lod = ir3_get_src(ctx, &tex->src[i].src)[0];
1709			has_lod = true;
1710			break;
1711		case nir_tex_src_comparator: /* shadow comparator */
1712			compare = ir3_get_src(ctx, &tex->src[i].src)[0];
1713			break;
1714		case nir_tex_src_projector:
1715			proj = ir3_get_src(ctx, &tex->src[i].src)[0];
1716			has_proj = true;
1717			break;
1718		case nir_tex_src_offset:
1719			off = ir3_get_src(ctx, &tex->src[i].src);
1720			has_off = true;
1721			break;
1722		case nir_tex_src_ddx:
1723			ddx = ir3_get_src(ctx, &tex->src[i].src);
1724			break;
1725		case nir_tex_src_ddy:
1726			ddy = ir3_get_src(ctx, &tex->src[i].src);
1727			break;
1728		case nir_tex_src_ms_index:
1729			sample_index = ir3_get_src(ctx, &tex->src[i].src)[0];
1730			break;
1731		case nir_tex_src_texture_offset:
1732		case nir_tex_src_sampler_offset:
1733			/* handled in get_tex_samp_src() */
1734			break;
1735		default:
1736			ir3_context_error(ctx, "Unhandled NIR tex src type: %d\n",
1737					tex->src[i].src_type);
1738			return;
1739		}
1740	}
1741
1742	switch (tex->op) {
1743	case nir_texop_tex:      opc = has_lod ? OPC_SAML : OPC_SAM; break;
1744	case nir_texop_txb:      opc = OPC_SAMB;     break;
1745	case nir_texop_txl:      opc = OPC_SAML;     break;
1746	case nir_texop_txd:      opc = OPC_SAMGQ;    break;
1747	case nir_texop_txf:      opc = OPC_ISAML;    break;
1748	case nir_texop_lod:      opc = OPC_GETLOD;   break;
1749	case nir_texop_tg4:
1750		/* NOTE: a4xx might need to emulate gather w/ txf (this is
1751		 * what blob does, seems gather  is broken?), and a3xx did
1752		 * not support it (but probably could also emulate).
1753		 */
1754		switch (tex->component) {
1755		case 0:              opc = OPC_GATHER4R; break;
1756		case 1:              opc = OPC_GATHER4G; break;
1757		case 2:              opc = OPC_GATHER4B; break;
1758		case 3:              opc = OPC_GATHER4A; break;
1759		}
1760		break;
1761	case nir_texop_txf_ms_fb:
1762	case nir_texop_txf_ms:   opc = OPC_ISAMM;    break;
1763	default:
1764		ir3_context_error(ctx, "Unhandled NIR tex type: %d\n", tex->op);
1765		return;
1766	}
1767
1768	tex_info(tex, &flags, &coords);
1769
1770	/*
1771	 * lay out the first argument in the proper order:
1772	 *  - actual coordinates first
1773	 *  - shadow reference
1774	 *  - array index
1775	 *  - projection w
1776	 *  - starting at offset 4, dpdx.xy, dpdy.xy
1777	 *
1778	 * bias/lod go into the second arg
1779	 */
1780
1781	/* insert tex coords: */
1782	for (i = 0; i < coords; i++)
1783		src0[i] = coord[i];
1784
1785	nsrc0 = i;
1786
1787	/* scale up integer coords for TXF based on the LOD */
1788	if (ctx->compiler->unminify_coords && (opc == OPC_ISAML)) {
1789		assert(has_lod);
1790		for (i = 0; i < coords; i++)
1791			src0[i] = ir3_SHL_B(b, src0[i], 0, lod, 0);
1792	}
1793
1794	if (coords == 1) {
1795		/* hw doesn't do 1d, so we treat it as 2d with
1796		 * height of 1, and patch up the y coord.
1797		 */
1798		if (is_isam(opc)) {
1799			src0[nsrc0++] = create_immed(b, 0);
1800		} else {
1801			src0[nsrc0++] = create_immed(b, fui(0.5));
1802		}
1803	}
1804
1805	if (tex->is_shadow && tex->op != nir_texop_lod)
1806		src0[nsrc0++] = compare;
1807
1808	if (tex->is_array && tex->op != nir_texop_lod) {
1809		struct ir3_instruction *idx = coord[coords];
1810
1811		/* the array coord for cube arrays needs 0.5 added to it */
1812		if (ctx->compiler->array_index_add_half && !is_isam(opc))
1813			idx = ir3_ADD_F(b, idx, 0, create_immed(b, fui(0.5)), 0);
1814
1815		src0[nsrc0++] = idx;
1816	}
1817
1818	if (has_proj) {
1819		src0[nsrc0++] = proj;
1820		flags |= IR3_INSTR_P;
1821	}
1822
1823	/* pad to 4, then ddx/ddy: */
1824	if (tex->op == nir_texop_txd) {
1825		while (nsrc0 < 4)
1826			src0[nsrc0++] = create_immed(b, fui(0.0));
1827		for (i = 0; i < coords; i++)
1828			src0[nsrc0++] = ddx[i];
1829		if (coords < 2)
1830			src0[nsrc0++] = create_immed(b, fui(0.0));
1831		for (i = 0; i < coords; i++)
1832			src0[nsrc0++] = ddy[i];
1833		if (coords < 2)
1834			src0[nsrc0++] = create_immed(b, fui(0.0));
1835	}
1836
1837	/* NOTE a3xx (and possibly a4xx?) might be different, using isaml
1838	 * with scaled x coord according to requested sample:
1839	 */
1840	if (opc == OPC_ISAMM) {
1841		if (ctx->compiler->txf_ms_with_isaml) {
1842			/* the samples are laid out in x dimension as
1843			 *     0 1 2 3
1844			 * x_ms = (x << ms) + sample_index;
1845			 */
1846			struct ir3_instruction *ms;
1847			ms = create_immed(b, (ctx->samples >> (2 * tex->texture_index)) & 3);
1848
1849			src0[0] = ir3_SHL_B(b, src0[0], 0, ms, 0);
1850			src0[0] = ir3_ADD_U(b, src0[0], 0, sample_index, 0);
1851
1852			opc = OPC_ISAML;
1853		} else {
1854			src0[nsrc0++] = sample_index;
1855		}
1856	}
1857
1858	/*
1859	 * second argument (if applicable):
1860	 *  - offsets
1861	 *  - lod
1862	 *  - bias
1863	 */
1864	if (has_off | has_lod | has_bias) {
1865		if (has_off) {
1866			unsigned off_coords = coords;
1867			if (tex->sampler_dim == GLSL_SAMPLER_DIM_CUBE)
1868				off_coords--;
1869			for (i = 0; i < off_coords; i++)
1870				src1[nsrc1++] = off[i];
1871			if (off_coords < 2)
1872				src1[nsrc1++] = create_immed(b, fui(0.0));
1873			flags |= IR3_INSTR_O;
1874		}
1875
1876		if (has_lod | has_bias)
1877			src1[nsrc1++] = lod;
1878	}
1879
1880	switch (tex->dest_type) {
1881	case nir_type_invalid:
1882	case nir_type_float:
1883		type = TYPE_F32;
1884		break;
1885	case nir_type_int:
1886		type = TYPE_S32;
1887		break;
1888	case nir_type_uint:
1889	case nir_type_bool:
1890		type = TYPE_U32;
1891		break;
1892	default:
1893		unreachable("bad dest_type");
1894	}
1895
1896	if (opc == OPC_GETLOD)
1897		type = TYPE_U32;
1898
1899	struct ir3_instruction *samp_tex;
1900
1901	if (tex->op == nir_texop_txf_ms_fb) {
1902		/* only expect a single txf_ms_fb per shader: */
1903		compile_assert(ctx, !ctx->so->fb_read);
1904		compile_assert(ctx, ctx->so->type == MESA_SHADER_FRAGMENT);
1905
1906		ctx->so->fb_read = true;
1907		samp_tex = ir3_create_collect(ctx, (struct ir3_instruction*[]){
1908			create_immed_typed(ctx->block, ctx->so->num_samp, TYPE_U16),
1909			create_immed_typed(ctx->block, ctx->so->num_samp, TYPE_U16),
1910		}, 2);
1911
1912		ctx->so->num_samp++;
1913	} else {
1914		samp_tex = get_tex_samp_tex_src(ctx, tex);
1915	}
1916
1917	struct ir3_instruction *col0 = ir3_create_collect(ctx, src0, nsrc0);
1918	struct ir3_instruction *col1 = ir3_create_collect(ctx, src1, nsrc1);
1919
1920	sam = ir3_SAM(b, opc, type, MASK(ncomp), flags,
1921			samp_tex, col0, col1);
1922
1923	if ((ctx->astc_srgb & (1 << tex->texture_index)) && !nir_tex_instr_is_query(tex)) {
1924		/* only need first 3 components: */
1925		sam->regs[0]->wrmask = 0x7;
1926		ir3_split_dest(b, dst, sam, 0, 3);
1927
1928		/* we need to sample the alpha separately with a non-ASTC
1929		 * texture state:
1930		 */
1931		sam = ir3_SAM(b, opc, type, 0b1000, flags,
1932				samp_tex, col0, col1);
1933
1934		array_insert(ctx->ir, ctx->ir->astc_srgb, sam);
1935
1936		/* fixup .w component: */
1937		ir3_split_dest(b, &dst[3], sam, 3, 1);
1938	} else {
1939		/* normal (non-workaround) case: */
1940		ir3_split_dest(b, dst, sam, 0, ncomp);
1941	}
1942
1943	/* GETLOD returns results in 4.8 fixed point */
1944	if (opc == OPC_GETLOD) {
1945		struct ir3_instruction *factor = create_immed(b, fui(1.0 / 256));
1946
1947		compile_assert(ctx, tex->dest_type == nir_type_float);
1948		for (i = 0; i < 2; i++) {
1949			dst[i] = ir3_MUL_F(b, ir3_COV(b, dst[i], TYPE_U32, TYPE_F32), 0,
1950							   factor, 0);
1951		}
1952	}
1953
1954	ir3_put_dst(ctx, &tex->dest);
1955}
1956
1957static void
1958emit_tex_query_levels(struct ir3_context *ctx, nir_tex_instr *tex)
1959{
1960	struct ir3_block *b = ctx->block;
1961	struct ir3_instruction **dst, *sam;
1962
1963	dst = ir3_get_dst(ctx, &tex->dest, 1);
1964
1965	sam = ir3_SAM(b, OPC_GETINFO, TYPE_U32, 0b0100, 0,
1966			get_tex_samp_tex_src(ctx, tex), NULL, NULL);
1967
1968	/* even though there is only one component, since it ends
1969	 * up in .z rather than .x, we need a split_dest()
1970	 */
1971	ir3_split_dest(b, dst, sam, 0, 3);
1972
1973	/* The # of levels comes from getinfo.z. We need to add 1 to it, since
1974	 * the value in TEX_CONST_0 is zero-based.
1975	 */
1976	if (ctx->compiler->levels_add_one)
1977		dst[0] = ir3_ADD_U(b, dst[0], 0, create_immed(b, 1), 0);
1978
1979	ir3_put_dst(ctx, &tex->dest);
1980}
1981
1982static void
1983emit_tex_txs(struct ir3_context *ctx, nir_tex_instr *tex)
1984{
1985	struct ir3_block *b = ctx->block;
1986	struct ir3_instruction **dst, *sam;
1987	struct ir3_instruction *lod;
1988	unsigned flags, coords;
1989
1990	tex_info(tex, &flags, &coords);
1991
1992	/* Actually we want the number of dimensions, not coordinates. This
1993	 * distinction only matters for cubes.
1994	 */
1995	if (tex->sampler_dim == GLSL_SAMPLER_DIM_CUBE)
1996		coords = 2;
1997
1998	dst = ir3_get_dst(ctx, &tex->dest, 4);
1999
2000	compile_assert(ctx, tex->num_srcs == 1);
2001	compile_assert(ctx, tex->src[0].src_type == nir_tex_src_lod);
2002
2003	lod = ir3_get_src(ctx, &tex->src[0].src)[0];
2004
2005	sam = ir3_SAM(b, OPC_GETSIZE, TYPE_U32, 0b1111, flags,
2006			get_tex_samp_tex_src(ctx, tex), lod, NULL);
2007
2008	ir3_split_dest(b, dst, sam, 0, 4);
2009
2010	/* Array size actually ends up in .w rather than .z. This doesn't
2011	 * matter for miplevel 0, but for higher mips the value in z is
2012	 * minified whereas w stays. Also, the value in TEX_CONST_3_DEPTH is
2013	 * returned, which means that we have to add 1 to it for arrays.
2014	 */
2015	if (tex->is_array) {
2016		if (ctx->compiler->levels_add_one) {
2017			dst[coords] = ir3_ADD_U(b, dst[3], 0, create_immed(b, 1), 0);
2018		} else {
2019			dst[coords] = ir3_MOV(b, dst[3], TYPE_U32);
2020		}
2021	}
2022
2023	ir3_put_dst(ctx, &tex->dest);
2024}
2025
2026static void
2027emit_jump(struct ir3_context *ctx, nir_jump_instr *jump)
2028{
2029	switch (jump->type) {
2030	case nir_jump_break:
2031	case nir_jump_continue:
2032	case nir_jump_return:
2033		/* I *think* we can simply just ignore this, and use the
2034		 * successor block link to figure out where we need to
2035		 * jump to for break/continue
2036		 */
2037		break;
2038	default:
2039		ir3_context_error(ctx, "Unhandled NIR jump type: %d\n", jump->type);
2040		break;
2041	}
2042}
2043
2044static void
2045emit_instr(struct ir3_context *ctx, nir_instr *instr)
2046{
2047	switch (instr->type) {
2048	case nir_instr_type_alu:
2049		emit_alu(ctx, nir_instr_as_alu(instr));
2050		break;
2051	case nir_instr_type_deref:
2052		/* ignored, handled as part of the intrinsic they are src to */
2053		break;
2054	case nir_instr_type_intrinsic:
2055		emit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
2056		break;
2057	case nir_instr_type_load_const:
2058		emit_load_const(ctx, nir_instr_as_load_const(instr));
2059		break;
2060	case nir_instr_type_ssa_undef:
2061		emit_undef(ctx, nir_instr_as_ssa_undef(instr));
2062		break;
2063	case nir_instr_type_tex: {
2064		nir_tex_instr *tex = nir_instr_as_tex(instr);
2065		/* couple tex instructions get special-cased:
2066		 */
2067		switch (tex->op) {
2068		case nir_texop_txs:
2069			emit_tex_txs(ctx, tex);
2070			break;
2071		case nir_texop_query_levels:
2072			emit_tex_query_levels(ctx, tex);
2073			break;
2074		default:
2075			emit_tex(ctx, tex);
2076			break;
2077		}
2078		break;
2079	}
2080	case nir_instr_type_jump:
2081		emit_jump(ctx, nir_instr_as_jump(instr));
2082		break;
2083	case nir_instr_type_phi:
2084		/* we have converted phi webs to regs in NIR by now */
2085		ir3_context_error(ctx, "Unexpected NIR instruction type: %d\n", instr->type);
2086		break;
2087	case nir_instr_type_call:
2088	case nir_instr_type_parallel_copy:
2089		ir3_context_error(ctx, "Unhandled NIR instruction type: %d\n", instr->type);
2090		break;
2091	}
2092}
2093
2094static struct ir3_block *
2095get_block(struct ir3_context *ctx, const nir_block *nblock)
2096{
2097	struct ir3_block *block;
2098	struct hash_entry *hentry;
2099	unsigned i;
2100
2101	hentry = _mesa_hash_table_search(ctx->block_ht, nblock);
2102	if (hentry)
2103		return hentry->data;
2104
2105	block = ir3_block_create(ctx->ir);
2106	block->nblock = nblock;
2107	_mesa_hash_table_insert(ctx->block_ht, nblock, block);
2108
2109	block->predecessors_count = nblock->predecessors->entries;
2110	block->predecessors = ralloc_array_size(block,
2111		sizeof(block->predecessors[0]), block->predecessors_count);
2112	i = 0;
2113	set_foreach(nblock->predecessors, sentry) {
2114		block->predecessors[i++] = get_block(ctx, sentry->key);
2115	}
2116
2117	return block;
2118}
2119
2120static void
2121emit_block(struct ir3_context *ctx, nir_block *nblock)
2122{
2123	struct ir3_block *block = get_block(ctx, nblock);
2124
2125	for (int i = 0; i < ARRAY_SIZE(block->successors); i++) {
2126		if (nblock->successors[i]) {
2127			block->successors[i] =
2128				get_block(ctx, nblock->successors[i]);
2129		}
2130	}
2131
2132	ctx->block = block;
2133	list_addtail(&block->node, &ctx->ir->block_list);
2134
2135	/* re-emit addr register in each block if needed: */
2136	for (int i = 0; i < ARRAY_SIZE(ctx->addr_ht); i++) {
2137		_mesa_hash_table_destroy(ctx->addr_ht[i], NULL);
2138		ctx->addr_ht[i] = NULL;
2139	}
2140
2141	nir_foreach_instr(instr, nblock) {
2142		ctx->cur_instr = instr;
2143		emit_instr(ctx, instr);
2144		ctx->cur_instr = NULL;
2145		if (ctx->error)
2146			return;
2147	}
2148}
2149
2150static void emit_cf_list(struct ir3_context *ctx, struct exec_list *list);
2151
2152static void
2153emit_if(struct ir3_context *ctx, nir_if *nif)
2154{
2155	struct ir3_instruction *condition = ir3_get_src(ctx, &nif->condition)[0];
2156
2157	ctx->block->condition =
2158		ir3_get_predicate(ctx, ir3_b2n(condition->block, condition));
2159
2160	emit_cf_list(ctx, &nif->then_list);
2161	emit_cf_list(ctx, &nif->else_list);
2162}
2163
2164static void
2165emit_loop(struct ir3_context *ctx, nir_loop *nloop)
2166{
2167	emit_cf_list(ctx, &nloop->body);
2168}
2169
2170static void
2171stack_push(struct ir3_context *ctx)
2172{
2173	ctx->stack++;
2174	ctx->max_stack = MAX2(ctx->max_stack, ctx->stack);
2175}
2176
2177static void
2178stack_pop(struct ir3_context *ctx)
2179{
2180	compile_assert(ctx, ctx->stack > 0);
2181	ctx->stack--;
2182}
2183
2184static void
2185emit_cf_list(struct ir3_context *ctx, struct exec_list *list)
2186{
2187	foreach_list_typed(nir_cf_node, node, node, list) {
2188		switch (node->type) {
2189		case nir_cf_node_block:
2190			emit_block(ctx, nir_cf_node_as_block(node));
2191			break;
2192		case nir_cf_node_if:
2193			stack_push(ctx);
2194			emit_if(ctx, nir_cf_node_as_if(node));
2195			stack_pop(ctx);
2196			break;
2197		case nir_cf_node_loop:
2198			stack_push(ctx);
2199			emit_loop(ctx, nir_cf_node_as_loop(node));
2200			stack_pop(ctx);
2201			break;
2202		case nir_cf_node_function:
2203			ir3_context_error(ctx, "TODO\n");
2204			break;
2205		}
2206	}
2207}
2208
2209/* emit stream-out code.  At this point, the current block is the original
2210 * (nir) end block, and nir ensures that all flow control paths terminate
2211 * into the end block.  We re-purpose the original end block to generate
2212 * the 'if (vtxcnt < maxvtxcnt)' condition, then append the conditional
2213 * block holding stream-out write instructions, followed by the new end
2214 * block:
2215 *
2216 *   blockOrigEnd {
2217 *      p0.x = (vtxcnt < maxvtxcnt)
2218 *      // succs: blockStreamOut, blockNewEnd
2219 *   }
2220 *   blockStreamOut {
2221 *      ... stream-out instructions ...
2222 *      // succs: blockNewEnd
2223 *   }
2224 *   blockNewEnd {
2225 *   }
2226 */
2227static void
2228emit_stream_out(struct ir3_context *ctx)
2229{
2230	struct ir3_shader_variant *v = ctx->so;
2231	struct ir3 *ir = ctx->ir;
2232	struct ir3_stream_output_info *strmout =
2233			&ctx->so->shader->stream_output;
2234	struct ir3_block *orig_end_block, *stream_out_block, *new_end_block;
2235	struct ir3_instruction *vtxcnt, *maxvtxcnt, *cond;
2236	struct ir3_instruction *bases[IR3_MAX_SO_BUFFERS];
2237
2238	/* create vtxcnt input in input block at top of shader,
2239	 * so that it is seen as live over the entire duration
2240	 * of the shader:
2241	 */
2242	vtxcnt = create_input(ctx, 0);
2243	add_sysval_input(ctx, SYSTEM_VALUE_VERTEX_CNT, vtxcnt);
2244
2245	maxvtxcnt = create_driver_param(ctx, IR3_DP_VTXCNT_MAX);
2246
2247	/* at this point, we are at the original 'end' block,
2248	 * re-purpose this block to stream-out condition, then
2249	 * append stream-out block and new-end block
2250	 */
2251	orig_end_block = ctx->block;
2252
2253// TODO these blocks need to update predecessors..
2254// maybe w/ store_global intrinsic, we could do this
2255// stuff in nir->nir pass
2256
2257	stream_out_block = ir3_block_create(ir);
2258	list_addtail(&stream_out_block->node, &ir->block_list);
2259
2260	new_end_block = ir3_block_create(ir);
2261	list_addtail(&new_end_block->node, &ir->block_list);
2262
2263	orig_end_block->successors[0] = stream_out_block;
2264	orig_end_block->successors[1] = new_end_block;
2265	stream_out_block->successors[0] = new_end_block;
2266
2267	/* setup 'if (vtxcnt < maxvtxcnt)' condition: */
2268	cond = ir3_CMPS_S(ctx->block, vtxcnt, 0, maxvtxcnt, 0);
2269	cond->regs[0]->num = regid(REG_P0, 0);
2270	cond->cat2.condition = IR3_COND_LT;
2271
2272	/* condition goes on previous block to the conditional,
2273	 * since it is used to pick which of the two successor
2274	 * paths to take:
2275	 */
2276	orig_end_block->condition = cond;
2277
2278	/* switch to stream_out_block to generate the stream-out
2279	 * instructions:
2280	 */
2281	ctx->block = stream_out_block;
2282
2283	/* Calculate base addresses based on vtxcnt.  Instructions
2284	 * generated for bases not used in following loop will be
2285	 * stripped out in the backend.
2286	 */
2287	for (unsigned i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
2288		unsigned stride = strmout->stride[i];
2289		struct ir3_instruction *base, *off;
2290
2291		base = create_uniform(ctx->block, regid(v->constbase.tfbo, i));
2292
2293		/* 24-bit should be enough: */
2294		off = ir3_MUL_U(ctx->block, vtxcnt, 0,
2295				create_immed(ctx->block, stride * 4), 0);
2296
2297		bases[i] = ir3_ADD_S(ctx->block, off, 0, base, 0);
2298	}
2299
2300	/* Generate the per-output store instructions: */
2301	for (unsigned i = 0; i < strmout->num_outputs; i++) {
2302		for (unsigned j = 0; j < strmout->output[i].num_components; j++) {
2303			unsigned c = j + strmout->output[i].start_component;
2304			struct ir3_instruction *base, *out, *stg;
2305
2306			base = bases[strmout->output[i].output_buffer];
2307			out = ctx->ir->outputs[regid(strmout->output[i].register_index, c)];
2308
2309			stg = ir3_STG(ctx->block, base, 0, out, 0,
2310					create_immed(ctx->block, 1), 0);
2311			stg->cat6.type = TYPE_U32;
2312			stg->cat6.dst_offset = (strmout->output[i].dst_offset + j) * 4;
2313
2314			array_insert(ctx->block, ctx->block->keeps, stg);
2315		}
2316	}
2317
2318	/* and finally switch to the new_end_block: */
2319	ctx->block = new_end_block;
2320}
2321
2322static void
2323emit_function(struct ir3_context *ctx, nir_function_impl *impl)
2324{
2325	nir_metadata_require(impl, nir_metadata_block_index);
2326
2327	compile_assert(ctx, ctx->stack == 0);
2328
2329	emit_cf_list(ctx, &impl->body);
2330	emit_block(ctx, impl->end_block);
2331
2332	compile_assert(ctx, ctx->stack == 0);
2333
2334	/* at this point, we should have a single empty block,
2335	 * into which we emit the 'end' instruction.
2336	 */
2337	compile_assert(ctx, list_empty(&ctx->block->instr_list));
2338
2339	/* If stream-out (aka transform-feedback) enabled, emit the
2340	 * stream-out instructions, followed by a new empty block (into
2341	 * which the 'end' instruction lands).
2342	 *
2343	 * NOTE: it is done in this order, rather than inserting before
2344	 * we emit end_block, because NIR guarantees that all blocks
2345	 * flow into end_block, and that end_block has no successors.
2346	 * So by re-purposing end_block as the first block of stream-
2347	 * out, we guarantee that all exit paths flow into the stream-
2348	 * out instructions.
2349	 */
2350	if ((ctx->compiler->gpu_id < 500) &&
2351			(ctx->so->shader->stream_output.num_outputs > 0) &&
2352			!ctx->so->binning_pass) {
2353		debug_assert(ctx->so->type == MESA_SHADER_VERTEX);
2354		emit_stream_out(ctx);
2355	}
2356
2357	ir3_END(ctx->block);
2358}
2359
2360static void
2361setup_input(struct ir3_context *ctx, nir_variable *in)
2362{
2363	struct ir3_shader_variant *so = ctx->so;
2364	unsigned ncomp = glsl_get_components(in->type);
2365	unsigned n = in->data.driver_location;
2366	unsigned frac = in->data.location_frac;
2367	unsigned slot = in->data.location;
2368
2369	/* skip unread inputs, we could end up with (for example), unsplit
2370	 * matrix/etc inputs in the case they are not read, so just silently
2371	 * skip these.
2372	 */
2373	if (ncomp > 4)
2374		return;
2375
2376	so->inputs[n].slot = slot;
2377	so->inputs[n].compmask = (1 << (ncomp + frac)) - 1;
2378	so->inputs_count = MAX2(so->inputs_count, n + 1);
2379	so->inputs[n].interpolate = in->data.interpolation;
2380	so->inputs[n].ncomp = ncomp;
2381
2382	if (ctx->so->type == MESA_SHADER_FRAGMENT) {
2383
2384		/* if any varyings have 'sample' qualifer, that triggers us
2385		 * to run in per-sample mode:
2386		 */
2387		so->per_samp |= in->data.sample;
2388
2389		for (int i = 0; i < ncomp; i++) {
2390			struct ir3_instruction *instr = NULL;
2391			unsigned idx = (n * 4) + i + frac;
2392
2393			if (slot == VARYING_SLOT_POS) {
2394				ir3_context_error(ctx, "fragcoord should be a sysval!\n");
2395			} else if (slot == VARYING_SLOT_PNTC) {
2396				/* see for example st_nir_fixup_varying_slots().. this is
2397				 * maybe a bit mesa/st specific.  But we need things to line
2398				 * up for this in fdN_program:
2399				 *    unsigned texmask = 1 << (slot - VARYING_SLOT_VAR0);
2400				 *    if (emit->sprite_coord_enable & texmask) {
2401				 *       ...
2402				 *    }
2403				 */
2404				so->inputs[n].slot = VARYING_SLOT_VAR8;
2405				so->inputs[n].bary = true;
2406				instr = create_frag_input(ctx, false, idx);
2407			} else {
2408				/* detect the special case for front/back colors where
2409				 * we need to do flat vs smooth shading depending on
2410				 * rast state:
2411				 */
2412				if (in->data.interpolation == INTERP_MODE_NONE) {
2413					switch (slot) {
2414					case VARYING_SLOT_COL0:
2415					case VARYING_SLOT_COL1:
2416					case VARYING_SLOT_BFC0:
2417					case VARYING_SLOT_BFC1:
2418						so->inputs[n].rasterflat = true;
2419						break;
2420					default:
2421						break;
2422					}
2423				}
2424
2425				if (ctx->compiler->flat_bypass) {
2426					if ((so->inputs[n].interpolate == INTERP_MODE_FLAT) ||
2427							(so->inputs[n].rasterflat && ctx->so->key.rasterflat))
2428						so->inputs[n].use_ldlv = true;
2429				}
2430
2431				so->inputs[n].bary = true;
2432
2433				instr = create_frag_input(ctx, so->inputs[n].use_ldlv, idx);
2434			}
2435
2436			compile_assert(ctx, idx < ctx->ir->ninputs);
2437
2438			ctx->ir->inputs[idx] = instr;
2439		}
2440	} else if (ctx->so->type == MESA_SHADER_VERTEX) {
2441		for (int i = 0; i < ncomp; i++) {
2442			unsigned idx = (n * 4) + i + frac;
2443			compile_assert(ctx, idx < ctx->ir->ninputs);
2444			ctx->ir->inputs[idx] = create_input(ctx, idx);
2445		}
2446	} else {
2447		ir3_context_error(ctx, "unknown shader type: %d\n", ctx->so->type);
2448	}
2449
2450	if (so->inputs[n].bary || (ctx->so->type == MESA_SHADER_VERTEX)) {
2451		so->total_in += ncomp;
2452	}
2453}
2454
2455/* Initially we assign non-packed inloc's for varyings, as we don't really
2456 * know up-front which components will be unused.  After all the compilation
2457 * stages we scan the shader to see which components are actually used, and
2458 * re-pack the inlocs to eliminate unneeded varyings.
2459 */
2460static void
2461pack_inlocs(struct ir3_context *ctx)
2462{
2463	struct ir3_shader_variant *so = ctx->so;
2464	uint8_t used_components[so->inputs_count];
2465
2466	memset(used_components, 0, sizeof(used_components));
2467
2468	/*
2469	 * First Step: scan shader to find which bary.f/ldlv remain:
2470	 */
2471
2472	list_for_each_entry (struct ir3_block, block, &ctx->ir->block_list, node) {
2473		list_for_each_entry (struct ir3_instruction, instr, &block->instr_list, node) {
2474			if (is_input(instr)) {
2475				unsigned inloc = instr->regs[1]->iim_val;
2476				unsigned i = inloc / 4;
2477				unsigned j = inloc % 4;
2478
2479				compile_assert(ctx, instr->regs[1]->flags & IR3_REG_IMMED);
2480				compile_assert(ctx, i < so->inputs_count);
2481
2482				used_components[i] |= 1 << j;
2483			}
2484		}
2485	}
2486
2487	/*
2488	 * Second Step: reassign varying inloc/slots:
2489	 */
2490
2491	unsigned actual_in = 0;
2492	unsigned inloc = 0;
2493
2494	for (unsigned i = 0; i < so->inputs_count; i++) {
2495		unsigned compmask = 0, maxcomp = 0;
2496
2497		so->inputs[i].ncomp = 0;
2498		so->inputs[i].inloc = inloc;
2499		so->inputs[i].bary = false;
2500
2501		for (unsigned j = 0; j < 4; j++) {
2502			if (!(used_components[i] & (1 << j)))
2503				continue;
2504
2505			compmask |= (1 << j);
2506			actual_in++;
2507			so->inputs[i].ncomp++;
2508			maxcomp = j + 1;
2509
2510			/* at this point, since used_components[i] mask is only
2511			 * considering varyings (ie. not sysvals) we know this
2512			 * is a varying:
2513			 */
2514			so->inputs[i].bary = true;
2515		}
2516
2517		if (so->inputs[i].bary) {
2518			so->varying_in++;
2519			so->inputs[i].compmask = (1 << maxcomp) - 1;
2520			inloc += maxcomp;
2521		}
2522	}
2523
2524	/*
2525	 * Third Step: reassign packed inloc's:
2526	 */
2527
2528	list_for_each_entry (struct ir3_block, block, &ctx->ir->block_list, node) {
2529		list_for_each_entry (struct ir3_instruction, instr, &block->instr_list, node) {
2530			if (is_input(instr)) {
2531				unsigned inloc = instr->regs[1]->iim_val;
2532				unsigned i = inloc / 4;
2533				unsigned j = inloc % 4;
2534
2535				instr->regs[1]->iim_val = so->inputs[i].inloc + j;
2536			}
2537		}
2538	}
2539}
2540
2541static void
2542setup_output(struct ir3_context *ctx, nir_variable *out)
2543{
2544	struct ir3_shader_variant *so = ctx->so;
2545	unsigned ncomp = glsl_get_components(out->type);
2546	unsigned n = out->data.driver_location;
2547	unsigned frac = out->data.location_frac;
2548	unsigned slot = out->data.location;
2549	unsigned comp = 0;
2550
2551	if (ctx->so->type == MESA_SHADER_FRAGMENT) {
2552		switch (slot) {
2553		case FRAG_RESULT_DEPTH:
2554			comp = 2;  /* tgsi will write to .z component */
2555			so->writes_pos = true;
2556			break;
2557		case FRAG_RESULT_COLOR:
2558			so->color0_mrt = 1;
2559			break;
2560		case FRAG_RESULT_SAMPLE_MASK:
2561			so->writes_smask = true;
2562			break;
2563		default:
2564			if (slot >= FRAG_RESULT_DATA0)
2565				break;
2566			ir3_context_error(ctx, "unknown FS output name: %s\n",
2567					gl_frag_result_name(slot));
2568		}
2569	} else if (ctx->so->type == MESA_SHADER_VERTEX) {
2570		switch (slot) {
2571		case VARYING_SLOT_POS:
2572			so->writes_pos = true;
2573			break;
2574		case VARYING_SLOT_PSIZ:
2575			so->writes_psize = true;
2576			break;
2577		case VARYING_SLOT_COL0:
2578		case VARYING_SLOT_COL1:
2579		case VARYING_SLOT_BFC0:
2580		case VARYING_SLOT_BFC1:
2581		case VARYING_SLOT_FOGC:
2582		case VARYING_SLOT_CLIP_DIST0:
2583		case VARYING_SLOT_CLIP_DIST1:
2584		case VARYING_SLOT_CLIP_VERTEX:
2585			break;
2586		default:
2587			if (slot >= VARYING_SLOT_VAR0)
2588				break;
2589			if ((VARYING_SLOT_TEX0 <= slot) && (slot <= VARYING_SLOT_TEX7))
2590				break;
2591			ir3_context_error(ctx, "unknown VS output name: %s\n",
2592					gl_varying_slot_name(slot));
2593		}
2594	} else {
2595		ir3_context_error(ctx, "unknown shader type: %d\n", ctx->so->type);
2596	}
2597
2598	compile_assert(ctx, n < ARRAY_SIZE(so->outputs));
2599
2600	so->outputs[n].slot = slot;
2601	so->outputs[n].regid = regid(n, comp);
2602	so->outputs_count = MAX2(so->outputs_count, n + 1);
2603
2604	for (int i = 0; i < ncomp; i++) {
2605		unsigned idx = (n * 4) + i + frac;
2606		compile_assert(ctx, idx < ctx->ir->noutputs);
2607		ctx->ir->outputs[idx] = create_immed(ctx->block, fui(0.0));
2608	}
2609
2610	/* if varying packing doesn't happen, we could end up in a situation
2611	 * with "holes" in the output, and since the per-generation code that
2612	 * sets up varying linkage registers doesn't expect to have more than
2613	 * one varying per vec4 slot, pad the holes.
2614	 *
2615	 * Note that this should probably generate a performance warning of
2616	 * some sort.
2617	 */
2618	for (int i = 0; i < frac; i++) {
2619		unsigned idx = (n * 4) + i;
2620		if (!ctx->ir->outputs[idx]) {
2621			ctx->ir->outputs[idx] = create_immed(ctx->block, fui(0.0));
2622		}
2623	}
2624}
2625
2626static int
2627max_drvloc(struct exec_list *vars)
2628{
2629	int drvloc = -1;
2630	nir_foreach_variable(var, vars) {
2631		drvloc = MAX2(drvloc, (int)var->data.driver_location);
2632	}
2633	return drvloc;
2634}
2635
2636static const unsigned max_sysvals[] = {
2637	[MESA_SHADER_FRAGMENT] = 24,  // TODO
2638	[MESA_SHADER_VERTEX]  = 16,
2639	[MESA_SHADER_COMPUTE] = 16, // TODO how many do we actually need?
2640	[MESA_SHADER_KERNEL]  = 16, // TODO how many do we actually need?
2641};
2642
2643static void
2644emit_instructions(struct ir3_context *ctx)
2645{
2646	unsigned ninputs, noutputs;
2647	nir_function_impl *fxn = nir_shader_get_entrypoint(ctx->s);
2648
2649	ninputs  = (max_drvloc(&ctx->s->inputs) + 1) * 4;
2650	noutputs = (max_drvloc(&ctx->s->outputs) + 1) * 4;
2651
2652	/* we need to leave room for sysvals:
2653	 */
2654	ninputs += max_sysvals[ctx->so->type];
2655
2656	ctx->ir = ir3_create(ctx->compiler, ctx->so->type, ninputs, noutputs);
2657
2658	/* Create inputs in first block: */
2659	ctx->block = get_block(ctx, nir_start_block(fxn));
2660	ctx->in_block = ctx->block;
2661	list_addtail(&ctx->block->node, &ctx->ir->block_list);
2662
2663	ninputs -= max_sysvals[ctx->so->type];
2664
2665	/* for fragment shader, the vcoord input register is used as the
2666	 * base for bary.f varying fetch instrs:
2667	 *
2668	 * TODO defer creating ctx->ij_pixel and corresponding sysvals
2669	 * until emit_intrinsic when we know they are actually needed.
2670	 * For now, we defer creating ctx->ij_centroid, etc, since we
2671	 * only need ij_pixel for "old style" varying inputs (ie.
2672	 * tgsi_to_nir)
2673	 */
2674	struct ir3_instruction *vcoord = NULL;
2675	if (ctx->so->type == MESA_SHADER_FRAGMENT) {
2676		struct ir3_instruction *xy[2];
2677
2678		vcoord = create_input_compmask(ctx, 0, 0x3);
2679		ir3_split_dest(ctx->block, xy, vcoord, 0, 2);
2680
2681		ctx->ij_pixel = ir3_create_collect(ctx, xy, 2);
2682	}
2683
2684	/* Setup inputs: */
2685	nir_foreach_variable(var, &ctx->s->inputs) {
2686		setup_input(ctx, var);
2687	}
2688
2689	/* Defer add_sysval_input() stuff until after setup_inputs(),
2690	 * because sysvals need to be appended after varyings:
2691	 */
2692	if (vcoord) {
2693		add_sysval_input_compmask(ctx, SYSTEM_VALUE_BARYCENTRIC_PIXEL,
2694				0x3, vcoord);
2695	}
2696
2697	/* Setup outputs: */
2698	nir_foreach_variable(var, &ctx->s->outputs) {
2699		setup_output(ctx, var);
2700	}
2701
2702	/* Find # of samplers: */
2703	nir_foreach_variable(var, &ctx->s->uniforms) {
2704		ctx->so->num_samp += glsl_type_get_sampler_count(var->type);
2705		/* just assume that we'll be reading from images.. if it
2706		 * is write-only we don't have to count it, but not sure
2707		 * if there is a good way to know?
2708		 */
2709		ctx->so->num_samp += glsl_type_get_image_count(var->type);
2710	}
2711
2712	/* NOTE: need to do something more clever when we support >1 fxn */
2713	nir_foreach_register(reg, &fxn->registers) {
2714		ir3_declare_array(ctx, reg);
2715	}
2716	/* And emit the body: */
2717	ctx->impl = fxn;
2718	emit_function(ctx, fxn);
2719}
2720
2721/* from NIR perspective, we actually have varying inputs.  But the varying
2722 * inputs, from an IR standpoint, are just bary.f/ldlv instructions.  The
2723 * only actual inputs are the sysvals.
2724 */
2725static void
2726fixup_frag_inputs(struct ir3_context *ctx)
2727{
2728	struct ir3_shader_variant *so = ctx->so;
2729	struct ir3 *ir = ctx->ir;
2730	unsigned i = 0;
2731
2732	/* sysvals should appear at the end of the inputs, drop everything else: */
2733	while ((i < so->inputs_count) && !so->inputs[i].sysval)
2734		i++;
2735
2736	/* at IR level, inputs are always blocks of 4 scalars: */
2737	i *= 4;
2738
2739	ir->inputs = &ir->inputs[i];
2740	ir->ninputs -= i;
2741}
2742
2743/* Fixup tex sampler state for astc/srgb workaround instructions.  We
2744 * need to assign the tex state indexes for these after we know the
2745 * max tex index.
2746 */
2747static void
2748fixup_astc_srgb(struct ir3_context *ctx)
2749{
2750	struct ir3_shader_variant *so = ctx->so;
2751	/* indexed by original tex idx, value is newly assigned alpha sampler
2752	 * state tex idx.  Zero is invalid since there is at least one sampler
2753	 * if we get here.
2754	 */
2755	unsigned alt_tex_state[16] = {0};
2756	unsigned tex_idx = ctx->max_texture_index + 1;
2757	unsigned idx = 0;
2758
2759	so->astc_srgb.base = tex_idx;
2760
2761	for (unsigned i = 0; i < ctx->ir->astc_srgb_count; i++) {
2762		struct ir3_instruction *sam = ctx->ir->astc_srgb[i];
2763
2764		compile_assert(ctx, sam->cat5.tex < ARRAY_SIZE(alt_tex_state));
2765
2766		if (alt_tex_state[sam->cat5.tex] == 0) {
2767			/* assign new alternate/alpha tex state slot: */
2768			alt_tex_state[sam->cat5.tex] = tex_idx++;
2769			so->astc_srgb.orig_idx[idx++] = sam->cat5.tex;
2770			so->astc_srgb.count++;
2771		}
2772
2773		sam->cat5.tex = alt_tex_state[sam->cat5.tex];
2774	}
2775}
2776
2777static void
2778fixup_binning_pass(struct ir3_context *ctx)
2779{
2780	struct ir3_shader_variant *so = ctx->so;
2781	struct ir3 *ir = ctx->ir;
2782	unsigned i, j;
2783
2784	for (i = 0, j = 0; i < so->outputs_count; i++) {
2785		unsigned slot = so->outputs[i].slot;
2786
2787		/* throw away everything but first position/psize */
2788		if ((slot == VARYING_SLOT_POS) || (slot == VARYING_SLOT_PSIZ)) {
2789			if (i != j) {
2790				so->outputs[j] = so->outputs[i];
2791				ir->outputs[(j*4)+0] = ir->outputs[(i*4)+0];
2792				ir->outputs[(j*4)+1] = ir->outputs[(i*4)+1];
2793				ir->outputs[(j*4)+2] = ir->outputs[(i*4)+2];
2794				ir->outputs[(j*4)+3] = ir->outputs[(i*4)+3];
2795			}
2796			j++;
2797		}
2798	}
2799	so->outputs_count = j;
2800	ir->noutputs = j * 4;
2801}
2802
2803int
2804ir3_compile_shader_nir(struct ir3_compiler *compiler,
2805		struct ir3_shader_variant *so)
2806{
2807	struct ir3_context *ctx;
2808	struct ir3 *ir;
2809	struct ir3_instruction **inputs;
2810	unsigned i;
2811	int ret = 0, max_bary;
2812
2813	assert(!so->ir);
2814
2815	ctx = ir3_context_init(compiler, so);
2816	if (!ctx) {
2817		DBG("INIT failed!");
2818		ret = -1;
2819		goto out;
2820	}
2821
2822	emit_instructions(ctx);
2823
2824	if (ctx->error) {
2825		DBG("EMIT failed!");
2826		ret = -1;
2827		goto out;
2828	}
2829
2830	ir = so->ir = ctx->ir;
2831
2832	/* keep track of the inputs from TGSI perspective.. */
2833	inputs = ir->inputs;
2834
2835	/* but fixup actual inputs for frag shader: */
2836	if (so->type == MESA_SHADER_FRAGMENT)
2837		fixup_frag_inputs(ctx);
2838
2839	/* at this point, for binning pass, throw away unneeded outputs: */
2840	if (so->binning_pass && (ctx->compiler->gpu_id < 600))
2841		fixup_binning_pass(ctx);
2842
2843	/* if we want half-precision outputs, mark the output registers
2844	 * as half:
2845	 */
2846	if (so->key.half_precision) {
2847		for (i = 0; i < ir->noutputs; i++) {
2848			struct ir3_instruction *out = ir->outputs[i];
2849
2850			if (!out)
2851				continue;
2852
2853			/* if frag shader writes z, that needs to be full precision: */
2854			if (so->outputs[i/4].slot == FRAG_RESULT_DEPTH)
2855				continue;
2856
2857			out->regs[0]->flags |= IR3_REG_HALF;
2858			/* output could be a fanout (ie. texture fetch output)
2859			 * in which case we need to propagate the half-reg flag
2860			 * up to the definer so that RA sees it:
2861			 */
2862			if (out->opc == OPC_META_FO) {
2863				out = out->regs[1]->instr;
2864				out->regs[0]->flags |= IR3_REG_HALF;
2865			}
2866
2867			if (out->opc == OPC_MOV) {
2868				out->cat1.dst_type = half_type(out->cat1.dst_type);
2869			}
2870		}
2871	}
2872
2873	if (ir3_shader_debug & IR3_DBG_OPTMSGS) {
2874		printf("BEFORE CP:\n");
2875		ir3_print(ir);
2876	}
2877
2878	ir3_cp(ir, so);
2879
2880	/* at this point, for binning pass, throw away unneeded outputs:
2881	 * Note that for a6xx and later, we do this after ir3_cp to ensure
2882	 * that the uniform/constant layout for BS and VS matches, so that
2883	 * we can re-use same VS_CONST state group.
2884	 */
2885	if (so->binning_pass && (ctx->compiler->gpu_id >= 600))
2886		fixup_binning_pass(ctx);
2887
2888	/* Insert mov if there's same instruction for each output.
2889	 * eg. dEQP-GLES31.functional.shaders.opaque_type_indexing.sampler.const_expression.vertex.sampler2dshadow
2890	 */
2891	for (int i = ir->noutputs - 1; i >= 0; i--) {
2892		if (!ir->outputs[i])
2893			continue;
2894		for (unsigned j = 0; j < i; j++) {
2895			if (ir->outputs[i] == ir->outputs[j]) {
2896				ir->outputs[i] =
2897					ir3_MOV(ir->outputs[i]->block, ir->outputs[i], TYPE_F32);
2898			}
2899		}
2900	}
2901
2902	if (ir3_shader_debug & IR3_DBG_OPTMSGS) {
2903		printf("BEFORE GROUPING:\n");
2904		ir3_print(ir);
2905	}
2906
2907	ir3_sched_add_deps(ir);
2908
2909	/* Group left/right neighbors, inserting mov's where needed to
2910	 * solve conflicts:
2911	 */
2912	ir3_group(ir);
2913
2914	if (ir3_shader_debug & IR3_DBG_OPTMSGS) {
2915		printf("AFTER GROUPING:\n");
2916		ir3_print(ir);
2917	}
2918
2919	ir3_depth(ir);
2920
2921	if (ir3_shader_debug & IR3_DBG_OPTMSGS) {
2922		printf("AFTER DEPTH:\n");
2923		ir3_print(ir);
2924	}
2925
2926	/* do Sethi–Ullman numbering before scheduling: */
2927	ir3_sun(ir);
2928
2929	ret = ir3_sched(ir);
2930	if (ret) {
2931		DBG("SCHED failed!");
2932		goto out;
2933	}
2934
2935	if (compiler->gpu_id >= 600) {
2936		ir3_a6xx_fixup_atomic_dests(ir, so);
2937	}
2938
2939	if (ir3_shader_debug & IR3_DBG_OPTMSGS) {
2940		printf("AFTER SCHED:\n");
2941		ir3_print(ir);
2942	}
2943
2944	ret = ir3_ra(ir, so->type, so->frag_coord, so->frag_face);
2945	if (ret) {
2946		DBG("RA failed!");
2947		goto out;
2948	}
2949
2950	if (ir3_shader_debug & IR3_DBG_OPTMSGS) {
2951		printf("AFTER RA:\n");
2952		ir3_print(ir);
2953	}
2954
2955	if (so->type == MESA_SHADER_FRAGMENT)
2956		pack_inlocs(ctx);
2957
2958	/* fixup input/outputs: */
2959	for (i = 0; i < so->outputs_count; i++) {
2960		/* sometimes we get outputs that don't write the .x coord, like:
2961		 *
2962		 *   decl_var shader_out INTERP_MODE_NONE float Color (VARYING_SLOT_VAR9.z, 1, 0)
2963		 *
2964		 * Presumably the result of varying packing and then eliminating
2965		 * some unneeded varyings?  Just skip head to the first valid
2966		 * component of the output.
2967		 */
2968		for (unsigned j = 0; j < 4; j++) {
2969			struct ir3_instruction *instr = ir->outputs[(i*4) + j];
2970			if (instr) {
2971				so->outputs[i].regid = instr->regs[0]->num;
2972				so->outputs[i].half  = !!(instr->regs[0]->flags & IR3_REG_HALF);
2973				break;
2974			}
2975		}
2976	}
2977
2978	/* Note that some or all channels of an input may be unused: */
2979	for (i = 0; i < so->inputs_count; i++) {
2980		unsigned j, reg = regid(63,0);
2981		bool half = false;
2982		for (j = 0; j < 4; j++) {
2983			struct ir3_instruction *in = inputs[(i*4) + j];
2984
2985			if (in && !(in->flags & IR3_INSTR_UNUSED)) {
2986				reg = in->regs[0]->num - j;
2987				if (half) {
2988					compile_assert(ctx, in->regs[0]->flags & IR3_REG_HALF);
2989				} else {
2990					half = !!(in->regs[0]->flags & IR3_REG_HALF);
2991				}
2992			}
2993		}
2994		so->inputs[i].regid = reg;
2995		so->inputs[i].half  = half;
2996	}
2997
2998	if (ctx->astc_srgb)
2999		fixup_astc_srgb(ctx);
3000
3001	/* We need to do legalize after (for frag shader's) the "bary.f"
3002	 * offsets (inloc) have been assigned.
3003	 */
3004	ir3_legalize(ir, &so->has_ssbo, &so->need_pixlod, &max_bary);
3005
3006	if (ir3_shader_debug & IR3_DBG_OPTMSGS) {
3007		printf("AFTER LEGALIZE:\n");
3008		ir3_print(ir);
3009	}
3010
3011	so->branchstack = ctx->max_stack;
3012
3013	/* Note that actual_in counts inputs that are not bary.f'd for FS: */
3014	if (so->type == MESA_SHADER_FRAGMENT)
3015		so->total_in = max_bary + 1;
3016
3017	so->max_sun = ir->max_sun;
3018
3019out:
3020	if (ret) {
3021		if (so->ir)
3022			ir3_destroy(so->ir);
3023		so->ir = NULL;
3024	}
3025	ir3_context_free(ctx);
3026
3027	return ret;
3028}
3029