1/*
2 * Copyright (c) 2012-2019 Etnaviv Project
3 * Copyright (c) 2019 Zodiac Inflight Innovations
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sub license,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the
13 * next paragraph) shall be included in all copies or substantial portions
14 * of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
22 * DEALINGS IN THE SOFTWARE.
23 *
24 * Authors:
25 *    Jonathan Marek <jonathan@marek.ca>
26 *    Wladimir J. van der Laan <laanwj@gmail.com>
27 */
28
29#include "etnaviv_compiler.h"
30#include "etnaviv_compiler_nir.h"
31#include "etnaviv_asm.h"
32#include "etnaviv_context.h"
33#include "etnaviv_debug.h"
34#include "etnaviv_nir.h"
35#include "etnaviv_uniforms.h"
36#include "etnaviv_util.h"
37
38#include <math.h>
39#include "util/u_memory.h"
40#include "util/register_allocate.h"
41#include "compiler/nir/nir_builder.h"
42
43#include "tgsi/tgsi_strings.h"
44#include "util/compiler.h"
45#include "util/half_float.h"
46
47static bool
48etna_alu_to_scalar_filter_cb(const nir_instr *instr, const void *data)
49{
50   const struct etna_specs *specs = data;
51
52   if (instr->type != nir_instr_type_alu)
53      return false;
54
55   nir_alu_instr *alu = nir_instr_as_alu(instr);
56   switch (alu->op) {
57   case nir_op_frsq:
58   case nir_op_frcp:
59   case nir_op_flog2:
60   case nir_op_fexp2:
61   case nir_op_fsqrt:
62   case nir_op_fcos:
63   case nir_op_fsin:
64   case nir_op_fdiv:
65   case nir_op_imul:
66      return true;
67   /* TODO: can do better than alu_to_scalar for vector compares */
68   case nir_op_b32all_fequal2:
69   case nir_op_b32all_fequal3:
70   case nir_op_b32all_fequal4:
71   case nir_op_b32any_fnequal2:
72   case nir_op_b32any_fnequal3:
73   case nir_op_b32any_fnequal4:
74   case nir_op_b32all_iequal2:
75   case nir_op_b32all_iequal3:
76   case nir_op_b32all_iequal4:
77   case nir_op_b32any_inequal2:
78   case nir_op_b32any_inequal3:
79   case nir_op_b32any_inequal4:
80      return true;
81   case nir_op_fdot2:
82      if (!specs->has_halti2_instructions)
83         return true;
84      break;
85   default:
86      break;
87   }
88
89   return false;
90}
91
92static void
93etna_emit_block_start(struct etna_compile *c, unsigned block)
94{
95   c->block_ptr[block] = c->inst_ptr;
96}
97
98static void
99etna_emit_output(struct etna_compile *c, nir_variable *var, struct etna_inst_src src)
100{
101   struct etna_shader_io_file *sf = &c->variant->outfile;
102
103   if (is_fs(c)) {
104      switch (var->data.location) {
105      case FRAG_RESULT_COLOR:
106      case FRAG_RESULT_DATA0: /* DATA0 is used by gallium shaders for color */
107         c->variant->ps_color_out_reg = src.reg;
108         break;
109      case FRAG_RESULT_DEPTH:
110         c->variant->ps_depth_out_reg = src.reg;
111         break;
112      default:
113         unreachable("Unsupported fs output");
114      }
115      return;
116   }
117
118   switch (var->data.location) {
119   case VARYING_SLOT_POS:
120      c->variant->vs_pos_out_reg = src.reg;
121      break;
122   case VARYING_SLOT_PSIZ:
123      c->variant->vs_pointsize_out_reg = src.reg;
124      break;
125   default:
126      sf->reg[sf->num_reg].reg = src.reg;
127      sf->reg[sf->num_reg].slot = var->data.location;
128      sf->reg[sf->num_reg].num_components = glsl_get_components(var->type);
129      sf->num_reg++;
130      break;
131   }
132}
133
134#define OPT(nir, pass, ...) ({                             \
135   bool this_progress = false;                             \
136   NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__);      \
137   this_progress;                                          \
138})
139
140static void
141etna_optimize_loop(nir_shader *s)
142{
143   bool progress;
144   do {
145      progress = false;
146
147      NIR_PASS_V(s, nir_lower_vars_to_ssa);
148      progress |= OPT(s, nir_opt_copy_prop_vars);
149      progress |= OPT(s, nir_opt_shrink_vectors, true);
150      progress |= OPT(s, nir_copy_prop);
151      progress |= OPT(s, nir_opt_dce);
152      progress |= OPT(s, nir_opt_cse);
153      progress |= OPT(s, nir_opt_peephole_select, 16, true, true);
154      progress |= OPT(s, nir_opt_intrinsics);
155      progress |= OPT(s, nir_opt_algebraic);
156      progress |= OPT(s, nir_opt_constant_folding);
157      progress |= OPT(s, nir_opt_dead_cf);
158      if (OPT(s, nir_opt_trivial_continues)) {
159         progress = true;
160         /* If nir_opt_trivial_continues makes progress, then we need to clean
161          * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll
162          * to make progress.
163          */
164         OPT(s, nir_copy_prop);
165         OPT(s, nir_opt_dce);
166      }
167      progress |= OPT(s, nir_opt_loop_unroll);
168      progress |= OPT(s, nir_opt_if, false);
169      progress |= OPT(s, nir_opt_remove_phis);
170      progress |= OPT(s, nir_opt_undef);
171   }
172   while (progress);
173}
174
175static int
176etna_glsl_type_size(const struct glsl_type *type, bool bindless)
177{
178   return glsl_count_attribute_slots(type, false);
179}
180
181static void
182copy_uniform_state_to_shader(struct etna_shader_variant *sobj, uint64_t *consts, unsigned count)
183{
184   struct etna_shader_uniform_info *uinfo = &sobj->uniforms;
185
186   uinfo->count = count * 4;
187   uinfo->data = MALLOC(uinfo->count * sizeof(*uinfo->data));
188   uinfo->contents = MALLOC(uinfo->count * sizeof(*uinfo->contents));
189
190   for (unsigned i = 0; i < uinfo->count; i++) {
191      uinfo->data[i] = consts[i];
192      uinfo->contents[i] = consts[i] >> 32;
193   }
194
195   etna_set_shader_uniforms_dirty_flags(sobj);
196}
197
198#define ALU_SWIZ(s) INST_SWIZ((s)->swizzle[0], (s)->swizzle[1], (s)->swizzle[2], (s)->swizzle[3])
199#define SRC_DISABLE ((hw_src){})
200#define SRC_CONST(idx, s) ((hw_src){.use=1, .rgroup = INST_RGROUP_UNIFORM_0, .reg=idx, .swiz=s})
201#define SRC_REG(idx, s) ((hw_src){.use=1, .rgroup = INST_RGROUP_TEMP, .reg=idx, .swiz=s})
202
203typedef struct etna_inst_dst hw_dst;
204typedef struct etna_inst_src hw_src;
205
206static inline hw_src
207src_swizzle(hw_src src, unsigned swizzle)
208{
209   if (src.rgroup != INST_RGROUP_IMMEDIATE)
210      src.swiz = inst_swiz_compose(src.swiz, swizzle);
211
212   return src;
213}
214
215/* constants are represented as 64-bit ints
216 * 32-bit for the value and 32-bit for the type (imm, uniform, etc)
217 */
218
219#define CONST_VAL(a, b) (nir_const_value) {.u64 = (uint64_t)(a) << 32 | (uint64_t)(b)}
220#define CONST(x) CONST_VAL(ETNA_UNIFORM_CONSTANT, x)
221#define UNIFORM(x) CONST_VAL(ETNA_UNIFORM_UNIFORM, x)
222#define TEXSCALE(x, i) CONST_VAL(ETNA_UNIFORM_TEXRECT_SCALE_X + (i), x)
223
224static int
225const_add(uint64_t *c, uint64_t value)
226{
227   for (unsigned i = 0; i < 4; i++) {
228      if (c[i] == value || !c[i]) {
229         c[i] = value;
230         return i;
231      }
232   }
233   return -1;
234}
235
236static hw_src
237const_src(struct etna_compile *c, nir_const_value *value, unsigned num_components)
238{
239   /* use inline immediates if possible */
240   if (c->specs->halti >= 2 && num_components == 1 &&
241       value[0].u64 >> 32 == ETNA_UNIFORM_CONSTANT) {
242      uint32_t bits = value[0].u32;
243
244      /* "float" - shifted by 12 */
245      if ((bits & 0xfff) == 0)
246         return etna_immediate_src(0, bits >> 12);
247
248      /* "unsigned" - raw 20 bit value */
249      if (bits < (1 << 20))
250         return etna_immediate_src(2, bits);
251
252      /* "signed" - sign extended 20-bit (sign included) value */
253      if (bits >= 0xfff80000)
254         return etna_immediate_src(1, bits);
255   }
256
257   unsigned i;
258   int swiz = -1;
259   for (i = 0; swiz < 0; i++) {
260      uint64_t *a = &c->consts[i*4];
261      uint64_t save[4];
262      memcpy(save, a, sizeof(save));
263      swiz = 0;
264      for (unsigned j = 0; j < num_components; j++) {
265         int c = const_add(a, value[j].u64);
266         if (c < 0) {
267            memcpy(a, save, sizeof(save));
268            swiz = -1;
269            break;
270         }
271         swiz |= c << j * 2;
272      }
273   }
274
275   assert(i <= ETNA_MAX_IMM / 4);
276   c->const_count = MAX2(c->const_count, i);
277
278   return SRC_CONST(i - 1, swiz);
279}
280
281/* how to swizzle when used as a src */
282static const uint8_t
283reg_swiz[NUM_REG_TYPES] = {
284   [REG_TYPE_VEC4] = INST_SWIZ_IDENTITY,
285   [REG_TYPE_VIRT_SCALAR_X] = INST_SWIZ_IDENTITY,
286   [REG_TYPE_VIRT_SCALAR_Y] = SWIZZLE(Y, Y, Y, Y),
287   [REG_TYPE_VIRT_VEC2_XY] = INST_SWIZ_IDENTITY,
288   [REG_TYPE_VIRT_VEC2T_XY] = INST_SWIZ_IDENTITY,
289   [REG_TYPE_VIRT_VEC2C_XY] = INST_SWIZ_IDENTITY,
290   [REG_TYPE_VIRT_SCALAR_Z] = SWIZZLE(Z, Z, Z, Z),
291   [REG_TYPE_VIRT_VEC2_XZ] = SWIZZLE(X, Z, X, Z),
292   [REG_TYPE_VIRT_VEC2_YZ] = SWIZZLE(Y, Z, Y, Z),
293   [REG_TYPE_VIRT_VEC2C_YZ] = SWIZZLE(Y, Z, Y, Z),
294   [REG_TYPE_VIRT_VEC3_XYZ] = INST_SWIZ_IDENTITY,
295   [REG_TYPE_VIRT_VEC3C_XYZ] = INST_SWIZ_IDENTITY,
296   [REG_TYPE_VIRT_SCALAR_W] = SWIZZLE(W, W, W, W),
297   [REG_TYPE_VIRT_VEC2_XW] = SWIZZLE(X, W, X, W),
298   [REG_TYPE_VIRT_VEC2_YW] = SWIZZLE(Y, W, Y, W),
299   [REG_TYPE_VIRT_VEC3_XYW] = SWIZZLE(X, Y, W, X),
300   [REG_TYPE_VIRT_VEC2_ZW] = SWIZZLE(Z, W, Z, W),
301   [REG_TYPE_VIRT_VEC2T_ZW] = SWIZZLE(Z, W, Z, W),
302   [REG_TYPE_VIRT_VEC2C_ZW] = SWIZZLE(Z, W, Z, W),
303   [REG_TYPE_VIRT_VEC3_XZW] = SWIZZLE(X, Z, W, X),
304   [REG_TYPE_VIRT_VEC3_YZW] = SWIZZLE(Y, Z, W, X),
305   [REG_TYPE_VIRT_VEC3C_YZW] = SWIZZLE(Y, Z, W, X),
306};
307
308/* how to swizzle when used as a dest */
309static const uint8_t
310reg_dst_swiz[NUM_REG_TYPES] = {
311   [REG_TYPE_VEC4] = INST_SWIZ_IDENTITY,
312   [REG_TYPE_VIRT_SCALAR_X] = INST_SWIZ_IDENTITY,
313   [REG_TYPE_VIRT_SCALAR_Y] = SWIZZLE(X, X, X, X),
314   [REG_TYPE_VIRT_VEC2_XY] = INST_SWIZ_IDENTITY,
315   [REG_TYPE_VIRT_VEC2T_XY] = INST_SWIZ_IDENTITY,
316   [REG_TYPE_VIRT_VEC2C_XY] = INST_SWIZ_IDENTITY,
317   [REG_TYPE_VIRT_SCALAR_Z] = SWIZZLE(X, X, X, X),
318   [REG_TYPE_VIRT_VEC2_XZ] = SWIZZLE(X, X, Y, Y),
319   [REG_TYPE_VIRT_VEC2_YZ] = SWIZZLE(X, X, Y, Y),
320   [REG_TYPE_VIRT_VEC2C_YZ] = SWIZZLE(X, X, Y, Y),
321   [REG_TYPE_VIRT_VEC3_XYZ] = INST_SWIZ_IDENTITY,
322   [REG_TYPE_VIRT_VEC3C_XYZ] = INST_SWIZ_IDENTITY,
323   [REG_TYPE_VIRT_SCALAR_W] = SWIZZLE(X, X, X, X),
324   [REG_TYPE_VIRT_VEC2_XW] = SWIZZLE(X, X, Y, Y),
325   [REG_TYPE_VIRT_VEC2_YW] = SWIZZLE(X, X, Y, Y),
326   [REG_TYPE_VIRT_VEC3_XYW] = SWIZZLE(X, Y, Z, Z),
327   [REG_TYPE_VIRT_VEC2_ZW] = SWIZZLE(X, X, X, Y),
328   [REG_TYPE_VIRT_VEC2T_ZW] = SWIZZLE(X, X, X, Y),
329   [REG_TYPE_VIRT_VEC2C_ZW] = SWIZZLE(X, X, X, Y),
330   [REG_TYPE_VIRT_VEC3_XZW] = SWIZZLE(X, Y, Y, Z),
331   [REG_TYPE_VIRT_VEC3_YZW] = SWIZZLE(X, X, Y, Z),
332   [REG_TYPE_VIRT_VEC3C_YZW] = SWIZZLE(X, X, Y, Z),
333};
334
335/* nir_src to allocated register */
336static hw_src
337ra_src(struct etna_compile *c, nir_src *src)
338{
339   unsigned reg = ra_get_node_reg(c->g, c->live_map[src_index(c->impl, src)]);
340   return SRC_REG(reg_get_base(c, reg), reg_swiz[reg_get_type(reg)]);
341}
342
343static hw_src
344get_src(struct etna_compile *c, nir_src *src)
345{
346   if (!src->is_ssa)
347      return ra_src(c, src);
348
349   nir_instr *instr = src->ssa->parent_instr;
350
351   if (instr->pass_flags & BYPASS_SRC) {
352      assert(instr->type == nir_instr_type_alu);
353      nir_alu_instr *alu = nir_instr_as_alu(instr);
354      assert(alu->op == nir_op_mov);
355      return src_swizzle(get_src(c, &alu->src[0].src), ALU_SWIZ(&alu->src[0]));
356   }
357
358   switch (instr->type) {
359   case nir_instr_type_load_const:
360      return const_src(c, nir_instr_as_load_const(instr)->value, src->ssa->num_components);
361   case nir_instr_type_intrinsic: {
362      nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
363      switch (intr->intrinsic) {
364      case nir_intrinsic_load_input:
365      case nir_intrinsic_load_instance_id:
366      case nir_intrinsic_load_uniform:
367      case nir_intrinsic_load_ubo:
368         return ra_src(c, src);
369      case nir_intrinsic_load_front_face:
370         return (hw_src) { .use = 1, .rgroup = INST_RGROUP_INTERNAL };
371      case nir_intrinsic_load_frag_coord:
372         return SRC_REG(0, INST_SWIZ_IDENTITY);
373      case nir_intrinsic_load_texture_rect_scaling: {
374         int sampler = nir_src_as_int(intr->src[0]);
375         nir_const_value values[] = {
376            TEXSCALE(sampler, 0),
377            TEXSCALE(sampler, 1),
378         };
379
380         return src_swizzle(const_src(c, values, 2), SWIZZLE(X,Y,X,X));
381      }
382      default:
383         compile_error(c, "Unhandled NIR intrinsic type: %s\n",
384                       nir_intrinsic_infos[intr->intrinsic].name);
385         break;
386      }
387   } break;
388   case nir_instr_type_alu:
389   case nir_instr_type_tex:
390      return ra_src(c, src);
391   case nir_instr_type_ssa_undef: {
392      /* return zero to deal with broken Blur demo */
393      nir_const_value value = CONST(0);
394      return src_swizzle(const_src(c, &value, 1), SWIZZLE(X,X,X,X));
395   }
396   default:
397      compile_error(c, "Unhandled NIR instruction type: %d\n", instr->type);
398      break;
399   }
400
401   return SRC_DISABLE;
402}
403
404static bool
405vec_dest_has_swizzle(nir_alu_instr *vec, nir_ssa_def *ssa)
406{
407   for (unsigned i = 0; i < 4; i++) {
408      if (!(vec->dest.write_mask & (1 << i)) || vec->src[i].src.ssa != ssa)
409         continue;
410
411      if (vec->src[i].swizzle[0] != i)
412         return true;
413   }
414
415   /* don't deal with possible bypassed vec/mov chain */
416   nir_foreach_use(use_src, ssa) {
417      nir_instr *instr = use_src->parent_instr;
418      if (instr->type != nir_instr_type_alu)
419         continue;
420
421      nir_alu_instr *alu = nir_instr_as_alu(instr);
422
423      switch (alu->op) {
424      case nir_op_mov:
425      case nir_op_vec2:
426      case nir_op_vec3:
427      case nir_op_vec4:
428         return true;
429      default:
430         break;
431      }
432   }
433   return false;
434}
435
436/* get allocated dest register for nir_dest
437 * *p_swiz tells how the components need to be placed into register
438 */
439static hw_dst
440ra_dest(struct etna_compile *c, nir_dest *dest, unsigned *p_swiz)
441{
442   unsigned swiz = INST_SWIZ_IDENTITY, mask = 0xf;
443   dest = real_dest(dest, &swiz, &mask);
444
445   unsigned r = ra_get_node_reg(c->g, c->live_map[dest_index(c->impl, dest)]);
446   unsigned t = reg_get_type(r);
447
448   *p_swiz = inst_swiz_compose(swiz, reg_dst_swiz[t]);
449
450   return (hw_dst) {
451      .use = 1,
452      .reg = reg_get_base(c, r),
453      .write_mask = inst_write_mask_compose(mask, reg_writemask[t]),
454   };
455}
456
457static void
458emit_alu(struct etna_compile *c, nir_alu_instr * alu)
459{
460   const nir_op_info *info = &nir_op_infos[alu->op];
461
462   /* marked as dead instruction (vecN and other bypassed instr) */
463   if (alu->instr.pass_flags)
464      return;
465
466   assert(!(alu->op >= nir_op_vec2 && alu->op <= nir_op_vec4));
467
468   unsigned dst_swiz;
469   hw_dst dst = ra_dest(c, &alu->dest.dest, &dst_swiz);
470
471   /* compose alu write_mask with RA write mask */
472   if (!alu->dest.dest.is_ssa)
473      dst.write_mask = inst_write_mask_compose(alu->dest.write_mask, dst.write_mask);
474
475   switch (alu->op) {
476   case nir_op_fdot2:
477   case nir_op_fdot3:
478   case nir_op_fdot4:
479      /* not per-component - don't compose dst_swiz */
480      dst_swiz = INST_SWIZ_IDENTITY;
481      break;
482   default:
483      break;
484   }
485
486   hw_src srcs[3];
487
488   for (int i = 0; i < info->num_inputs; i++) {
489      nir_alu_src *asrc = &alu->src[i];
490      hw_src src;
491
492      src = src_swizzle(get_src(c, &asrc->src), ALU_SWIZ(asrc));
493      src = src_swizzle(src, dst_swiz);
494
495      if (src.rgroup != INST_RGROUP_IMMEDIATE) {
496         src.neg = asrc->negate || (alu->op == nir_op_fneg);
497         src.abs = asrc->abs || (alu->op == nir_op_fabs);
498      } else {
499         assert(!asrc->negate && alu->op != nir_op_fneg);
500         assert(!asrc->abs && alu->op != nir_op_fabs);
501      }
502
503      srcs[i] = src;
504   }
505
506   etna_emit_alu(c, alu->op, dst, srcs, alu->dest.saturate || (alu->op == nir_op_fsat));
507}
508
509static void
510emit_tex(struct etna_compile *c, nir_tex_instr * tex)
511{
512   unsigned dst_swiz;
513   hw_dst dst = ra_dest(c, &tex->dest, &dst_swiz);
514   nir_src *coord = NULL, *lod_bias = NULL, *compare = NULL;
515
516   for (unsigned i = 0; i < tex->num_srcs; i++) {
517      switch (tex->src[i].src_type) {
518      case nir_tex_src_coord:
519         coord = &tex->src[i].src;
520         break;
521      case nir_tex_src_bias:
522      case nir_tex_src_lod:
523         assert(!lod_bias);
524         lod_bias = &tex->src[i].src;
525         break;
526      case nir_tex_src_comparator:
527         compare = &tex->src[i].src;
528         break;
529      default:
530         compile_error(c, "Unhandled NIR tex src type: %d\n",
531                       tex->src[i].src_type);
532         break;
533      }
534   }
535
536   etna_emit_tex(c, tex->op, tex->sampler_index, dst_swiz, dst, get_src(c, coord),
537                 lod_bias ? get_src(c, lod_bias) : SRC_DISABLE,
538                 compare ? get_src(c, compare) : SRC_DISABLE);
539}
540
541static void
542emit_intrinsic(struct etna_compile *c, nir_intrinsic_instr * intr)
543{
544   switch (intr->intrinsic) {
545   case nir_intrinsic_store_deref:
546      etna_emit_output(c, nir_src_as_deref(intr->src[0])->var, get_src(c, &intr->src[1]));
547      break;
548   case nir_intrinsic_discard_if:
549      etna_emit_discard(c, get_src(c, &intr->src[0]));
550      break;
551   case nir_intrinsic_discard:
552      etna_emit_discard(c, SRC_DISABLE);
553      break;
554   case nir_intrinsic_load_uniform: {
555      unsigned dst_swiz;
556      struct etna_inst_dst dst = ra_dest(c, &intr->dest, &dst_swiz);
557
558      /* TODO: rework so extra MOV isn't required, load up to 4 addresses at once */
559      emit_inst(c, &(struct etna_inst) {
560         .opcode = INST_OPCODE_MOVAR,
561         .dst.write_mask = 0x1,
562         .src[2] = get_src(c, &intr->src[0]),
563      });
564      emit_inst(c, &(struct etna_inst) {
565         .opcode = INST_OPCODE_MOV,
566         .dst = dst,
567         .src[2] = {
568            .use = 1,
569            .rgroup = INST_RGROUP_UNIFORM_0,
570            .reg = nir_intrinsic_base(intr),
571            .swiz = dst_swiz,
572            .amode = INST_AMODE_ADD_A_X,
573         },
574      });
575   } break;
576   case nir_intrinsic_load_ubo: {
577      /* TODO: if offset is of the form (x + C) then add C to the base instead */
578      unsigned idx = nir_src_as_const_value(intr->src[0])[0].u32;
579      unsigned dst_swiz;
580      emit_inst(c, &(struct etna_inst) {
581         .opcode = INST_OPCODE_LOAD,
582         .type = INST_TYPE_U32,
583         .dst = ra_dest(c, &intr->dest, &dst_swiz),
584         .src[0] = get_src(c, &intr->src[1]),
585         .src[1] = const_src(c, &CONST_VAL(ETNA_UNIFORM_UBO0_ADDR + idx, 0), 1),
586      });
587   } break;
588   case nir_intrinsic_load_front_face:
589   case nir_intrinsic_load_frag_coord:
590      assert(intr->dest.is_ssa); /* TODO - lower phis could cause this */
591      break;
592   case nir_intrinsic_load_input:
593   case nir_intrinsic_load_instance_id:
594   case nir_intrinsic_load_texture_rect_scaling:
595      break;
596   default:
597      compile_error(c, "Unhandled NIR intrinsic type: %s\n",
598                    nir_intrinsic_infos[intr->intrinsic].name);
599   }
600}
601
602static void
603emit_instr(struct etna_compile *c, nir_instr * instr)
604{
605   switch (instr->type) {
606   case nir_instr_type_alu:
607      emit_alu(c, nir_instr_as_alu(instr));
608      break;
609   case nir_instr_type_tex:
610      emit_tex(c, nir_instr_as_tex(instr));
611      break;
612   case nir_instr_type_intrinsic:
613      emit_intrinsic(c, nir_instr_as_intrinsic(instr));
614      break;
615   case nir_instr_type_jump:
616      assert(nir_instr_is_last(instr));
617      break;
618   case nir_instr_type_load_const:
619   case nir_instr_type_ssa_undef:
620   case nir_instr_type_deref:
621      break;
622   default:
623      compile_error(c, "Unhandled NIR instruction type: %d\n", instr->type);
624      break;
625   }
626}
627
628static void
629emit_block(struct etna_compile *c, nir_block * block)
630{
631   etna_emit_block_start(c, block->index);
632
633   nir_foreach_instr(instr, block)
634      emit_instr(c, instr);
635
636   /* succs->index < block->index is for the loop case  */
637   nir_block *succs = block->successors[0];
638   if (nir_block_ends_in_jump(block) || succs->index < block->index)
639      etna_emit_jump(c, succs->index, SRC_DISABLE);
640}
641
642static void
643emit_cf_list(struct etna_compile *c, struct exec_list *list);
644
645static void
646emit_if(struct etna_compile *c, nir_if * nif)
647{
648   etna_emit_jump(c, nir_if_first_else_block(nif)->index, get_src(c, &nif->condition));
649   emit_cf_list(c, &nif->then_list);
650
651   /* jump at end of then_list to skip else_list
652    * not needed if then_list already ends with a jump or else_list is empty
653    */
654   if (!nir_block_ends_in_jump(nir_if_last_then_block(nif)) &&
655       !nir_cf_list_is_empty_block(&nif->else_list))
656      etna_emit_jump(c, nir_if_last_else_block(nif)->successors[0]->index, SRC_DISABLE);
657
658   emit_cf_list(c, &nif->else_list);
659}
660
661static void
662emit_cf_list(struct etna_compile *c, struct exec_list *list)
663{
664   foreach_list_typed(nir_cf_node, node, node, list) {
665      switch (node->type) {
666      case nir_cf_node_block:
667         emit_block(c, nir_cf_node_as_block(node));
668         break;
669      case nir_cf_node_if:
670         emit_if(c, nir_cf_node_as_if(node));
671         break;
672      case nir_cf_node_loop:
673         emit_cf_list(c, &nir_cf_node_as_loop(node)->body);
674         break;
675      default:
676         compile_error(c, "Unknown NIR node type\n");
677         break;
678      }
679   }
680}
681
682/* based on nir_lower_vec_to_movs */
683static unsigned
684insert_vec_mov(nir_alu_instr *vec, unsigned start_idx, nir_shader *shader)
685{
686   assert(start_idx < nir_op_infos[vec->op].num_inputs);
687   unsigned write_mask = (1u << start_idx);
688
689   nir_alu_instr *mov = nir_alu_instr_create(shader, nir_op_mov);
690   nir_alu_src_copy(&mov->src[0], &vec->src[start_idx]);
691
692   mov->src[0].swizzle[0] = vec->src[start_idx].swizzle[0];
693   mov->src[0].negate = vec->src[start_idx].negate;
694   mov->src[0].abs = vec->src[start_idx].abs;
695
696   unsigned num_components = 1;
697
698   for (unsigned i = start_idx + 1; i < 4; i++) {
699      if (!(vec->dest.write_mask & (1 << i)))
700         continue;
701
702      if (nir_srcs_equal(vec->src[i].src, vec->src[start_idx].src) &&
703          vec->src[i].negate == vec->src[start_idx].negate &&
704          vec->src[i].abs == vec->src[start_idx].abs) {
705         write_mask |= (1 << i);
706         mov->src[0].swizzle[num_components] = vec->src[i].swizzle[0];
707         num_components++;
708      }
709   }
710
711   mov->dest.write_mask = (1 << num_components) - 1;
712   nir_ssa_dest_init(&mov->instr, &mov->dest.dest, num_components, 32, NULL);
713
714   /* replace vec srcs with inserted mov */
715   for (unsigned i = 0, j = 0; i < 4; i++) {
716      if (!(write_mask & (1 << i)))
717         continue;
718
719      nir_instr_rewrite_src(&vec->instr, &vec->src[i].src, nir_src_for_ssa(&mov->dest.dest.ssa));
720      vec->src[i].swizzle[0] = j++;
721   }
722
723   nir_instr_insert_before(&vec->instr, &mov->instr);
724
725   return write_mask;
726}
727
728/*
729 * for vecN instructions:
730 * -merge constant sources into a single src
731 * -insert movs (nir_lower_vec_to_movs equivalent)
732 * for non-vecN instructions:
733 * -try to merge constants as single constant
734 * -insert movs for multiple constants (pre-HALTI5)
735 */
736static void
737lower_alu(struct etna_compile *c, nir_alu_instr *alu)
738{
739   const nir_op_info *info = &nir_op_infos[alu->op];
740
741   nir_builder b;
742   nir_builder_init(&b, c->impl);
743   b.cursor = nir_before_instr(&alu->instr);
744
745   switch (alu->op) {
746   case nir_op_vec2:
747   case nir_op_vec3:
748   case nir_op_vec4:
749      break;
750   default:
751      /* pre-GC7000L can only have 1 uniform src per instruction */
752      if (c->specs->halti >= 5)
753         return;
754
755      nir_const_value value[4] = {};
756      uint8_t swizzle[4][4] = {};
757      unsigned swiz_max = 0, num_const = 0;
758
759      for (unsigned i = 0; i < info->num_inputs; i++) {
760         nir_const_value *cv = nir_src_as_const_value(alu->src[i].src);
761         if (!cv)
762            continue;
763
764         unsigned num_components = info->input_sizes[i] ?: alu->dest.dest.ssa.num_components;
765         for (unsigned j = 0; j < num_components; j++) {
766            int idx = const_add(&value[0].u64, cv[alu->src[i].swizzle[j]].u64);
767            swizzle[i][j] = idx;
768            swiz_max = MAX2(swiz_max, (unsigned) idx);
769         }
770         num_const++;
771      }
772
773      /* nothing to do */
774      if (num_const <= 1)
775         return;
776
777      /* resolve with single combined const src */
778      if (swiz_max < 4) {
779         nir_ssa_def *def = nir_build_imm(&b, swiz_max + 1, 32, value);
780
781         for (unsigned i = 0; i < info->num_inputs; i++) {
782            nir_const_value *cv = nir_src_as_const_value(alu->src[i].src);
783            if (!cv)
784               continue;
785
786            nir_instr_rewrite_src(&alu->instr, &alu->src[i].src, nir_src_for_ssa(def));
787
788            for (unsigned j = 0; j < 4; j++)
789               alu->src[i].swizzle[j] = swizzle[i][j];
790         }
791         return;
792      }
793
794      /* resolve with movs */
795      num_const = 0;
796      for (unsigned i = 0; i < info->num_inputs; i++) {
797         nir_const_value *cv = nir_src_as_const_value(alu->src[i].src);
798         if (!cv)
799            continue;
800
801         num_const++;
802         if (num_const == 1)
803            continue;
804
805         nir_ssa_def *mov = nir_mov(&b, alu->src[i].src.ssa);
806         nir_instr_rewrite_src(&alu->instr, &alu->src[i].src, nir_src_for_ssa(mov));
807      }
808      return;
809   }
810
811   nir_const_value value[4];
812   unsigned num_components = 0;
813
814   for (unsigned i = 0; i < info->num_inputs; i++) {
815      nir_const_value *cv = nir_src_as_const_value(alu->src[i].src);
816      if (cv)
817         value[num_components++] = cv[alu->src[i].swizzle[0]];
818   }
819
820   /* if there is more than one constant source to the vecN, combine them
821    * into a single load_const (removing the vecN completely if all components
822    * are constant)
823    */
824   if (num_components > 1) {
825      nir_ssa_def *def = nir_build_imm(&b, num_components, 32, value);
826
827      if (num_components == info->num_inputs) {
828         nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, def);
829         nir_instr_remove(&alu->instr);
830         return;
831      }
832
833      for (unsigned i = 0, j = 0; i < info->num_inputs; i++) {
834         nir_const_value *cv = nir_src_as_const_value(alu->src[i].src);
835         if (!cv)
836            continue;
837
838         nir_instr_rewrite_src(&alu->instr, &alu->src[i].src, nir_src_for_ssa(def));
839         alu->src[i].swizzle[0] = j++;
840      }
841   }
842
843   unsigned finished_write_mask = 0;
844   for (unsigned i = 0; i < 4; i++) {
845      if (!(alu->dest.write_mask & (1 << i)))
846            continue;
847
848      nir_ssa_def *ssa = alu->src[i].src.ssa;
849
850      /* check that vecN instruction is only user of this */
851      bool need_mov = list_length(&ssa->if_uses) != 0;
852      nir_foreach_use(use_src, ssa) {
853         if (use_src->parent_instr != &alu->instr)
854            need_mov = true;
855      }
856
857      nir_instr *instr = ssa->parent_instr;
858      switch (instr->type) {
859      case nir_instr_type_alu:
860      case nir_instr_type_tex:
861         break;
862      case nir_instr_type_intrinsic:
863         if (nir_instr_as_intrinsic(instr)->intrinsic == nir_intrinsic_load_input) {
864            need_mov = vec_dest_has_swizzle(alu, &nir_instr_as_intrinsic(instr)->dest.ssa);
865            break;
866         }
867         FALLTHROUGH;
868      default:
869         need_mov = true;
870      }
871
872      if (need_mov && !(finished_write_mask & (1 << i)))
873         finished_write_mask |= insert_vec_mov(alu, i, c->nir);
874   }
875}
876
877static bool
878emit_shader(struct etna_compile *c, unsigned *num_temps, unsigned *num_consts)
879{
880   nir_shader *shader = c->nir;
881   c->impl = nir_shader_get_entrypoint(shader);
882
883   bool have_indirect_uniform = false;
884   unsigned indirect_max = 0;
885
886   nir_builder b;
887   nir_builder_init(&b, c->impl);
888
889   /* convert non-dynamic uniform loads to constants, etc */
890   nir_foreach_block(block, c->impl) {
891      nir_foreach_instr_safe(instr, block) {
892         switch(instr->type) {
893         case nir_instr_type_alu:
894            /* deals with vecN and const srcs */
895            lower_alu(c, nir_instr_as_alu(instr));
896            break;
897         case nir_instr_type_load_const: {
898            nir_load_const_instr *load_const = nir_instr_as_load_const(instr);
899            for (unsigned  i = 0; i < load_const->def.num_components; i++)
900               load_const->value[i] = CONST(load_const->value[i].u32);
901         } break;
902         case nir_instr_type_intrinsic: {
903            nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
904            /* TODO: load_ubo can also become a constant in some cases
905             * (at the moment it can end up emitting a LOAD with two
906             *  uniform sources, which could be a problem on HALTI2)
907             */
908            if (intr->intrinsic != nir_intrinsic_load_uniform)
909               break;
910            nir_const_value *off = nir_src_as_const_value(intr->src[0]);
911            if (!off || off[0].u64 >> 32 != ETNA_UNIFORM_CONSTANT) {
912               have_indirect_uniform = true;
913               indirect_max = nir_intrinsic_base(intr) + nir_intrinsic_range(intr);
914               break;
915            }
916
917            unsigned base = nir_intrinsic_base(intr);
918            /* pre halti2 uniform offset will be float */
919            if (c->specs->halti < 2)
920               base += (unsigned) off[0].f32;
921            else
922               base += off[0].u32;
923            nir_const_value value[4];
924
925            for (unsigned i = 0; i < intr->dest.ssa.num_components; i++)
926               value[i] = UNIFORM(base * 4 + i);
927
928            b.cursor = nir_after_instr(instr);
929            nir_ssa_def *def = nir_build_imm(&b, intr->dest.ssa.num_components, 32, value);
930
931            nir_ssa_def_rewrite_uses(&intr->dest.ssa, def);
932            nir_instr_remove(instr);
933         } break;
934         default:
935            break;
936         }
937      }
938   }
939
940   /* TODO: only emit required indirect uniform ranges */
941   if (have_indirect_uniform) {
942      for (unsigned i = 0; i < indirect_max * 4; i++)
943         c->consts[i] = UNIFORM(i).u64;
944      c->const_count = indirect_max;
945   }
946
947   /* add mov for any store output using sysval/const and for depth stores from intrinsics */
948   nir_foreach_block(block, c->impl) {
949      nir_foreach_instr_safe(instr, block) {
950         if (instr->type != nir_instr_type_intrinsic)
951            continue;
952
953         nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
954
955         switch (intr->intrinsic) {
956         case nir_intrinsic_store_deref: {
957            nir_deref_instr *deref = nir_src_as_deref(intr->src[0]);
958            nir_src *src = &intr->src[1];
959            if (nir_src_is_const(*src) || is_sysval(src->ssa->parent_instr) ||
960                (shader->info.stage == MESA_SHADER_FRAGMENT &&
961                 deref->var->data.location == FRAG_RESULT_DEPTH &&
962                 src->is_ssa &&
963                 src->ssa->parent_instr->type != nir_instr_type_alu)) {
964               b.cursor = nir_before_instr(instr);
965               nir_instr_rewrite_src(instr, src, nir_src_for_ssa(nir_mov(&b, src->ssa)));
966            }
967         } break;
968         default:
969            break;
970         }
971      }
972   }
973
974   /* call directly to avoid validation (load_const don't pass validation at this point) */
975   nir_convert_from_ssa(shader, true);
976   nir_opt_dce(shader);
977
978   etna_ra_assign(c, shader);
979
980   emit_cf_list(c, &nir_shader_get_entrypoint(shader)->body);
981
982   *num_temps = etna_ra_finish(c);
983   *num_consts = c->const_count;
984   return true;
985}
986
987static bool
988etna_compile_check_limits(struct etna_shader_variant *v)
989{
990   const struct etna_specs *specs = v->shader->specs;
991   int max_uniforms = (v->stage == MESA_SHADER_VERTEX)
992                         ? specs->max_vs_uniforms
993                         : specs->max_ps_uniforms;
994
995   if (!specs->has_icache && v->needs_icache) {
996      DBG("Number of instructions (%d) exceeds maximum %d", v->code_size / 4,
997          specs->max_instructions);
998      return false;
999   }
1000
1001   if (v->num_temps > specs->max_registers) {
1002      DBG("Number of registers (%d) exceeds maximum %d", v->num_temps,
1003          specs->max_registers);
1004      return false;
1005   }
1006
1007   if (v->uniforms.count / 4 > max_uniforms) {
1008      DBG("Number of uniforms (%d) exceeds maximum %d",
1009          v->uniforms.count / 4, max_uniforms);
1010      return false;
1011   }
1012
1013   return true;
1014}
1015
1016static void
1017fill_vs_mystery(struct etna_shader_variant *v)
1018{
1019   const struct etna_specs *specs = v->shader->specs;
1020
1021   v->input_count_unk8 = DIV_ROUND_UP(v->infile.num_reg + 4, 16); /* XXX what is this */
1022
1023   /* fill in "mystery meat" load balancing value. This value determines how
1024    * work is scheduled between VS and PS
1025    * in the unified shader architecture. More precisely, it is determined from
1026    * the number of VS outputs, as well as chip-specific
1027    * vertex output buffer size, vertex cache size, and the number of shader
1028    * cores.
1029    *
1030    * XXX this is a conservative estimate, the "optimal" value is only known for
1031    * sure at link time because some
1032    * outputs may be unused and thus unmapped. Then again, in the general use
1033    * case with GLSL the vertex and fragment
1034    * shaders are linked already before submitting to Gallium, thus all outputs
1035    * are used.
1036    *
1037    * note: TGSI compiler counts all outputs (including position and pointsize), here
1038    * v->outfile.num_reg only counts varyings, +1 to compensate for the position output
1039    * TODO: might have a problem that we don't count pointsize when it is used
1040    */
1041
1042   int half_out = v->outfile.num_reg / 2 + 1;
1043   assert(half_out);
1044
1045   uint32_t b = ((20480 / (specs->vertex_output_buffer_size -
1046                           2 * half_out * specs->vertex_cache_size)) +
1047                 9) /
1048                10;
1049   uint32_t a = (b + 256 / (specs->shader_core_count * half_out)) / 2;
1050   v->vs_load_balancing = VIVS_VS_LOAD_BALANCING_A(MIN2(a, 255)) |
1051                             VIVS_VS_LOAD_BALANCING_B(MIN2(b, 255)) |
1052                             VIVS_VS_LOAD_BALANCING_C(0x3f) |
1053                             VIVS_VS_LOAD_BALANCING_D(0x0f);
1054}
1055
1056bool
1057etna_compile_shader_nir(struct etna_shader_variant *v)
1058{
1059   if (unlikely(!v))
1060      return false;
1061
1062   struct etna_compile *c = CALLOC_STRUCT(etna_compile);
1063   if (!c)
1064      return false;
1065
1066   c->variant = v;
1067   c->specs = v->shader->specs;
1068   c->nir = nir_shader_clone(NULL, v->shader->nir);
1069
1070   nir_shader *s = c->nir;
1071   const struct etna_specs *specs = c->specs;
1072
1073   v->stage = s->info.stage;
1074   v->uses_discard = s->info.fs.uses_discard;
1075   v->num_loops = 0; /* TODO */
1076   v->vs_id_in_reg = -1;
1077   v->vs_pos_out_reg = -1;
1078   v->vs_pointsize_out_reg = -1;
1079   v->ps_color_out_reg = 0; /* 0 for shader that doesn't write fragcolor.. */
1080   v->ps_depth_out_reg = -1;
1081
1082   /*
1083    * Lower glTexCoord, fixes e.g. neverball point sprite (exit cylinder stars)
1084    * and gl4es pointsprite.trace apitrace
1085    */
1086   if (s->info.stage == MESA_SHADER_FRAGMENT && v->key.sprite_coord_enable) {
1087      NIR_PASS_V(s, nir_lower_texcoord_replace, v->key.sprite_coord_enable,
1088                 false, v->key.sprite_coord_yinvert);
1089   }
1090
1091   /* setup input linking */
1092   struct etna_shader_io_file *sf = &v->infile;
1093   if (s->info.stage == MESA_SHADER_VERTEX) {
1094      nir_foreach_shader_in_variable(var, s) {
1095         unsigned idx = var->data.driver_location;
1096         sf->reg[idx].reg = idx;
1097         sf->reg[idx].slot = var->data.location;
1098         sf->reg[idx].num_components = glsl_get_components(var->type);
1099         sf->num_reg = MAX2(sf->num_reg, idx+1);
1100      }
1101   } else {
1102      unsigned count = 0;
1103      nir_foreach_shader_in_variable(var, s) {
1104         unsigned idx = var->data.driver_location;
1105         sf->reg[idx].reg = idx + 1;
1106         sf->reg[idx].slot = var->data.location;
1107         sf->reg[idx].num_components = glsl_get_components(var->type);
1108         sf->num_reg = MAX2(sf->num_reg, idx+1);
1109         count++;
1110      }
1111      assert(sf->num_reg == count);
1112   }
1113
1114   NIR_PASS_V(s, nir_lower_io, nir_var_shader_in | nir_var_uniform, etna_glsl_type_size,
1115            (nir_lower_io_options)0);
1116
1117   NIR_PASS_V(s, nir_lower_regs_to_ssa);
1118   NIR_PASS_V(s, nir_lower_vars_to_ssa);
1119   NIR_PASS_V(s, nir_lower_indirect_derefs, nir_var_all, UINT32_MAX);
1120   NIR_PASS_V(s, nir_lower_tex, &(struct nir_lower_tex_options) { .lower_txp = ~0u });
1121   NIR_PASS_V(s, nir_lower_alu_to_scalar, etna_alu_to_scalar_filter_cb, specs);
1122   nir_lower_idiv_options idiv_options = {
1123      .imprecise_32bit_lowering = true,
1124      .allow_fp16 = true,
1125   };
1126   NIR_PASS_V(s, nir_lower_idiv, &idiv_options);
1127
1128   etna_optimize_loop(s);
1129
1130   /* TODO: remove this extra run if nir_opt_peephole_select is able to handle ubo's. */
1131   if (OPT(s, etna_nir_lower_ubo_to_uniform))
1132      etna_optimize_loop(s);
1133
1134   NIR_PASS_V(s, etna_lower_io, v);
1135
1136   if (v->shader->specs->vs_need_z_div)
1137      NIR_PASS_V(s, nir_lower_clip_halfz);
1138
1139   /* lower pre-halti2 to float (halti0 has integers, but only scalar..) */
1140   if (c->specs->halti < 2) {
1141      /* use opt_algebraic between int_to_float and boot_to_float because
1142       * int_to_float emits ftrunc, and ftrunc lowering generates bool ops
1143       */
1144      NIR_PASS_V(s, nir_lower_int_to_float);
1145      NIR_PASS_V(s, nir_opt_algebraic);
1146      NIR_PASS_V(s, nir_lower_bool_to_float);
1147   } else {
1148      NIR_PASS_V(s, nir_lower_bool_to_int32);
1149   }
1150
1151   while( OPT(s, nir_opt_vectorize, NULL, NULL) );
1152   NIR_PASS_V(s, nir_lower_alu_to_scalar, etna_alu_to_scalar_filter_cb, specs);
1153
1154   NIR_PASS_V(s, nir_remove_dead_variables, nir_var_function_temp, NULL);
1155   NIR_PASS_V(s, nir_opt_algebraic_late);
1156
1157   NIR_PASS_V(s, nir_move_vec_src_uses_to_dest);
1158   NIR_PASS_V(s, nir_copy_prop);
1159   /* only HW supported integer source mod is ineg for iadd instruction (?) */
1160   NIR_PASS_V(s, nir_lower_to_source_mods, ~nir_lower_int_source_mods);
1161   /* need copy prop after uses_to_dest, and before src mods: see
1162    * dEQP-GLES2.functional.shaders.random.all_features.fragment.95
1163    */
1164
1165   NIR_PASS_V(s, nir_opt_dce);
1166
1167   NIR_PASS_V(s, nir_lower_bool_to_bitsize);
1168   NIR_PASS_V(s, etna_lower_alu, c->specs->has_new_transcendentals);
1169
1170   if (DBG_ENABLED(ETNA_DBG_DUMP_SHADERS))
1171      nir_print_shader(s, stdout);
1172
1173   unsigned block_ptr[nir_shader_get_entrypoint(s)->num_blocks];
1174   c->block_ptr = block_ptr;
1175
1176   unsigned num_consts;
1177   ASSERTED bool ok = emit_shader(c, &v->num_temps, &num_consts);
1178   assert(ok);
1179
1180   /* empty shader, emit NOP */
1181   if (!c->inst_ptr)
1182      emit_inst(c, &(struct etna_inst) { .opcode = INST_OPCODE_NOP });
1183
1184   /* assemble instructions, fixing up labels */
1185   uint32_t *code = MALLOC(c->inst_ptr * 16);
1186   for (unsigned i = 0; i < c->inst_ptr; i++) {
1187      struct etna_inst *inst = &c->code[i];
1188      if (inst->opcode == INST_OPCODE_BRANCH)
1189         inst->imm = block_ptr[inst->imm];
1190
1191      inst->halti5 = specs->halti >= 5;
1192      etna_assemble(&code[i * 4], inst);
1193   }
1194
1195   v->code_size = c->inst_ptr * 4;
1196   v->code = code;
1197   v->needs_icache = c->inst_ptr > specs->max_instructions;
1198
1199   copy_uniform_state_to_shader(v, c->consts, num_consts);
1200
1201   if (s->info.stage == MESA_SHADER_FRAGMENT) {
1202      v->input_count_unk8 = 31; /* XXX what is this */
1203      assert(v->ps_depth_out_reg <= 0);
1204   } else {
1205      fill_vs_mystery(v);
1206   }
1207
1208   bool result = etna_compile_check_limits(v);
1209   ralloc_free(c->nir);
1210   FREE(c);
1211   return result;
1212}
1213
1214static const struct etna_shader_inout *
1215etna_shader_vs_lookup(const struct etna_shader_variant *sobj,
1216                      const struct etna_shader_inout *in)
1217{
1218   for (int i = 0; i < sobj->outfile.num_reg; i++)
1219      if (sobj->outfile.reg[i].slot == in->slot)
1220         return &sobj->outfile.reg[i];
1221
1222   return NULL;
1223}
1224
1225bool
1226etna_link_shader_nir(struct etna_shader_link_info *info,
1227                     const struct etna_shader_variant *vs,
1228                     const struct etna_shader_variant *fs)
1229{
1230   int comp_ofs = 0;
1231   /* For each fragment input we need to find the associated vertex shader
1232    * output, which can be found by matching on semantic name and index. A
1233    * binary search could be used because the vs outputs are sorted by their
1234    * semantic index and grouped by semantic type by fill_in_vs_outputs.
1235    */
1236   assert(fs->infile.num_reg < ETNA_NUM_INPUTS);
1237   info->pcoord_varying_comp_ofs = -1;
1238
1239   for (int idx = 0; idx < fs->infile.num_reg; ++idx) {
1240      const struct etna_shader_inout *fsio = &fs->infile.reg[idx];
1241      const struct etna_shader_inout *vsio = etna_shader_vs_lookup(vs, fsio);
1242      struct etna_varying *varying;
1243      bool interpolate_always = true;
1244
1245      assert(fsio->reg > 0 && fsio->reg <= ARRAY_SIZE(info->varyings));
1246
1247      if (fsio->reg > info->num_varyings)
1248         info->num_varyings = fsio->reg;
1249
1250      varying = &info->varyings[fsio->reg - 1];
1251      varying->num_components = fsio->num_components;
1252
1253      if (!interpolate_always) /* colors affected by flat shading */
1254         varying->pa_attributes = 0x200;
1255      else /* texture coord or other bypasses flat shading */
1256         varying->pa_attributes = 0x2f1;
1257
1258      varying->use[0] = VARYING_COMPONENT_USE_UNUSED;
1259      varying->use[1] = VARYING_COMPONENT_USE_UNUSED;
1260      varying->use[2] = VARYING_COMPONENT_USE_UNUSED;
1261      varying->use[3] = VARYING_COMPONENT_USE_UNUSED;
1262
1263      /* point/tex coord is an input to the PS without matching VS output,
1264       * so it gets a varying slot without being assigned a VS register.
1265       */
1266      if (fsio->slot == VARYING_SLOT_PNTC) {
1267         varying->use[0] = VARYING_COMPONENT_USE_POINTCOORD_X;
1268         varying->use[1] = VARYING_COMPONENT_USE_POINTCOORD_Y;
1269
1270         info->pcoord_varying_comp_ofs = comp_ofs;
1271      } else if (util_varying_is_point_coord(fsio->slot, fs->key.sprite_coord_enable)) {
1272         /*
1273	  * Do nothing, TexCoord is lowered to PointCoord above
1274	  * and the TexCoord here is just a remnant. This needs
1275	  * to be removed with some nir_remove_dead_variables(),
1276	  * but that one removes all FS inputs ... why?
1277	  */
1278      } else {
1279         if (vsio == NULL) { /* not found -- link error */
1280            BUG("Semantic value not found in vertex shader outputs\n");
1281            return true;
1282         }
1283         varying->reg = vsio->reg;
1284      }
1285
1286      comp_ofs += varying->num_components;
1287   }
1288
1289   assert(info->num_varyings == fs->infile.num_reg);
1290
1291   return false;
1292}
1293