• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
47 static bool
etna_alu_to_scalar_filter_cb(const nir_instr * instr,const void * data)48 etna_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 
92 static void
etna_emit_block_start(struct etna_compile * c,unsigned block)93 etna_emit_block_start(struct etna_compile *c, unsigned block)
94 {
95    c->block_ptr[block] = c->inst_ptr;
96 }
97 
98 static void
etna_emit_output(struct etna_compile * c,nir_variable * var,struct etna_inst_src src)99 etna_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       assert(sf->num_reg < ETNA_NUM_INPUTS);
127       sf->reg[sf->num_reg].reg = src.reg;
128       sf->reg[sf->num_reg].slot = var->data.location;
129       sf->reg[sf->num_reg].num_components = glsl_get_components(var->type);
130       sf->num_reg++;
131       break;
132    }
133 }
134 
135 #define OPT(nir, pass, ...) ({                             \
136    bool this_progress = false;                             \
137    NIR_PASS(this_progress, nir, pass, ##__VA_ARGS__);      \
138    this_progress;                                          \
139 })
140 
141 static void
etna_optimize_loop(nir_shader * s)142 etna_optimize_loop(nir_shader *s)
143 {
144    bool progress;
145    do {
146       progress = false;
147 
148       NIR_PASS_V(s, nir_lower_vars_to_ssa);
149       progress |= OPT(s, nir_opt_copy_prop_vars);
150       progress |= OPT(s, nir_opt_shrink_stores, true);
151       progress |= OPT(s, nir_opt_shrink_vectors);
152       progress |= OPT(s, nir_copy_prop);
153       progress |= OPT(s, nir_opt_dce);
154       progress |= OPT(s, nir_opt_cse);
155       progress |= OPT(s, nir_opt_peephole_select, 16, true, true);
156       progress |= OPT(s, nir_opt_intrinsics);
157       progress |= OPT(s, nir_opt_algebraic);
158       progress |= OPT(s, nir_opt_constant_folding);
159       progress |= OPT(s, nir_opt_dead_cf);
160       if (OPT(s, nir_opt_trivial_continues)) {
161          progress = true;
162          /* If nir_opt_trivial_continues makes progress, then we need to clean
163           * things up if we want any hope of nir_opt_if or nir_opt_loop_unroll
164           * to make progress.
165           */
166          OPT(s, nir_copy_prop);
167          OPT(s, nir_opt_dce);
168       }
169       progress |= OPT(s, nir_opt_loop_unroll);
170       progress |= OPT(s, nir_opt_if, nir_opt_if_optimize_phi_true_false);
171       progress |= OPT(s, nir_opt_remove_phis);
172       progress |= OPT(s, nir_opt_undef);
173    }
174    while (progress);
175 }
176 
177 static int
etna_glsl_type_size(const struct glsl_type * type,bool bindless)178 etna_glsl_type_size(const struct glsl_type *type, bool bindless)
179 {
180    return glsl_count_attribute_slots(type, false);
181 }
182 
183 static void
copy_uniform_state_to_shader(struct etna_shader_variant * sobj,uint64_t * consts,unsigned count)184 copy_uniform_state_to_shader(struct etna_shader_variant *sobj, uint64_t *consts, unsigned count)
185 {
186    struct etna_shader_uniform_info *uinfo = &sobj->uniforms;
187 
188    uinfo->count = count * 4;
189    uinfo->data = MALLOC(uinfo->count * sizeof(*uinfo->data));
190    uinfo->contents = MALLOC(uinfo->count * sizeof(*uinfo->contents));
191 
192    for (unsigned i = 0; i < uinfo->count; i++) {
193       uinfo->data[i] = consts[i];
194       uinfo->contents[i] = consts[i] >> 32;
195    }
196 
197    etna_set_shader_uniforms_dirty_flags(sobj);
198 }
199 
200 #define ALU_SWIZ(s) INST_SWIZ((s)->swizzle[0], (s)->swizzle[1], (s)->swizzle[2], (s)->swizzle[3])
201 #define SRC_DISABLE ((hw_src){})
202 #define SRC_CONST(idx, s) ((hw_src){.use=1, .rgroup = INST_RGROUP_UNIFORM_0, .reg=idx, .swiz=s})
203 #define SRC_REG(idx, s) ((hw_src){.use=1, .rgroup = INST_RGROUP_TEMP, .reg=idx, .swiz=s})
204 
205 typedef struct etna_inst_dst hw_dst;
206 typedef struct etna_inst_src hw_src;
207 
208 static inline hw_src
src_swizzle(hw_src src,unsigned swizzle)209 src_swizzle(hw_src src, unsigned swizzle)
210 {
211    if (src.rgroup != INST_RGROUP_IMMEDIATE)
212       src.swiz = inst_swiz_compose(src.swiz, swizzle);
213 
214    return src;
215 }
216 
217 /* constants are represented as 64-bit ints
218  * 32-bit for the value and 32-bit for the type (imm, uniform, etc)
219  */
220 
221 #define CONST_VAL(a, b) (nir_const_value) {.u64 = (uint64_t)(a) << 32 | (uint64_t)(b)}
222 #define CONST(x) CONST_VAL(ETNA_UNIFORM_CONSTANT, x)
223 #define UNIFORM(x) CONST_VAL(ETNA_UNIFORM_UNIFORM, x)
224 #define TEXSCALE(x, i) CONST_VAL(ETNA_UNIFORM_TEXRECT_SCALE_X + (i), x)
225 
226 static int
const_add(uint64_t * c,uint64_t value)227 const_add(uint64_t *c, uint64_t value)
228 {
229    for (unsigned i = 0; i < 4; i++) {
230       if (c[i] == value || !c[i]) {
231          c[i] = value;
232          return i;
233       }
234    }
235    return -1;
236 }
237 
238 static hw_src
const_src(struct etna_compile * c,nir_const_value * value,unsigned num_components)239 const_src(struct etna_compile *c, nir_const_value *value, unsigned num_components)
240 {
241    /* use inline immediates if possible */
242    if (c->specs->halti >= 2 && num_components == 1 &&
243        value[0].u64 >> 32 == ETNA_UNIFORM_CONSTANT) {
244       uint32_t bits = value[0].u32;
245 
246       /* "float" - shifted by 12 */
247       if ((bits & 0xfff) == 0)
248          return etna_immediate_src(0, bits >> 12);
249 
250       /* "unsigned" - raw 20 bit value */
251       if (bits < (1 << 20))
252          return etna_immediate_src(2, bits);
253 
254       /* "signed" - sign extended 20-bit (sign included) value */
255       if (bits >= 0xfff80000)
256          return etna_immediate_src(1, bits);
257    }
258 
259    unsigned i;
260    int swiz = -1;
261    for (i = 0; swiz < 0; i++) {
262       uint64_t *a = &c->consts[i*4];
263       uint64_t save[4];
264       memcpy(save, a, sizeof(save));
265       swiz = 0;
266       for (unsigned j = 0; j < num_components; j++) {
267          int c = const_add(a, value[j].u64);
268          if (c < 0) {
269             memcpy(a, save, sizeof(save));
270             swiz = -1;
271             break;
272          }
273          swiz |= c << j * 2;
274       }
275    }
276 
277    assert(i <= ETNA_MAX_IMM / 4);
278    c->const_count = MAX2(c->const_count, i);
279 
280    return SRC_CONST(i - 1, swiz);
281 }
282 
283 /* how to swizzle when used as a src */
284 static const uint8_t
285 reg_swiz[NUM_REG_TYPES] = {
286    [REG_TYPE_VEC4] = INST_SWIZ_IDENTITY,
287    [REG_TYPE_VIRT_SCALAR_X] = INST_SWIZ_IDENTITY,
288    [REG_TYPE_VIRT_SCALAR_Y] = SWIZZLE(Y, Y, Y, Y),
289    [REG_TYPE_VIRT_VEC2_XY] = INST_SWIZ_IDENTITY,
290    [REG_TYPE_VIRT_VEC2T_XY] = INST_SWIZ_IDENTITY,
291    [REG_TYPE_VIRT_VEC2C_XY] = INST_SWIZ_IDENTITY,
292    [REG_TYPE_VIRT_SCALAR_Z] = SWIZZLE(Z, Z, Z, Z),
293    [REG_TYPE_VIRT_VEC2_XZ] = SWIZZLE(X, Z, X, Z),
294    [REG_TYPE_VIRT_VEC2_YZ] = SWIZZLE(Y, Z, Y, Z),
295    [REG_TYPE_VIRT_VEC2C_YZ] = SWIZZLE(Y, Z, Y, Z),
296    [REG_TYPE_VIRT_VEC3_XYZ] = INST_SWIZ_IDENTITY,
297    [REG_TYPE_VIRT_VEC3C_XYZ] = INST_SWIZ_IDENTITY,
298    [REG_TYPE_VIRT_SCALAR_W] = SWIZZLE(W, W, W, W),
299    [REG_TYPE_VIRT_VEC2_XW] = SWIZZLE(X, W, X, W),
300    [REG_TYPE_VIRT_VEC2_YW] = SWIZZLE(Y, W, Y, W),
301    [REG_TYPE_VIRT_VEC3_XYW] = SWIZZLE(X, Y, W, X),
302    [REG_TYPE_VIRT_VEC2_ZW] = SWIZZLE(Z, W, Z, W),
303    [REG_TYPE_VIRT_VEC2T_ZW] = SWIZZLE(Z, W, Z, W),
304    [REG_TYPE_VIRT_VEC2C_ZW] = SWIZZLE(Z, W, Z, W),
305    [REG_TYPE_VIRT_VEC3_XZW] = SWIZZLE(X, Z, W, X),
306    [REG_TYPE_VIRT_VEC3_YZW] = SWIZZLE(Y, Z, W, X),
307    [REG_TYPE_VIRT_VEC3C_YZW] = SWIZZLE(Y, Z, W, X),
308 };
309 
310 /* how to swizzle when used as a dest */
311 static const uint8_t
312 reg_dst_swiz[NUM_REG_TYPES] = {
313    [REG_TYPE_VEC4] = INST_SWIZ_IDENTITY,
314    [REG_TYPE_VIRT_SCALAR_X] = INST_SWIZ_IDENTITY,
315    [REG_TYPE_VIRT_SCALAR_Y] = SWIZZLE(X, X, X, X),
316    [REG_TYPE_VIRT_VEC2_XY] = INST_SWIZ_IDENTITY,
317    [REG_TYPE_VIRT_VEC2T_XY] = INST_SWIZ_IDENTITY,
318    [REG_TYPE_VIRT_VEC2C_XY] = INST_SWIZ_IDENTITY,
319    [REG_TYPE_VIRT_SCALAR_Z] = SWIZZLE(X, X, X, X),
320    [REG_TYPE_VIRT_VEC2_XZ] = SWIZZLE(X, X, Y, Y),
321    [REG_TYPE_VIRT_VEC2_YZ] = SWIZZLE(X, X, Y, Y),
322    [REG_TYPE_VIRT_VEC2C_YZ] = SWIZZLE(X, X, Y, Y),
323    [REG_TYPE_VIRT_VEC3_XYZ] = INST_SWIZ_IDENTITY,
324    [REG_TYPE_VIRT_VEC3C_XYZ] = INST_SWIZ_IDENTITY,
325    [REG_TYPE_VIRT_SCALAR_W] = SWIZZLE(X, X, X, X),
326    [REG_TYPE_VIRT_VEC2_XW] = SWIZZLE(X, X, Y, Y),
327    [REG_TYPE_VIRT_VEC2_YW] = SWIZZLE(X, X, Y, Y),
328    [REG_TYPE_VIRT_VEC3_XYW] = SWIZZLE(X, Y, Z, Z),
329    [REG_TYPE_VIRT_VEC2_ZW] = SWIZZLE(X, X, X, Y),
330    [REG_TYPE_VIRT_VEC2T_ZW] = SWIZZLE(X, X, X, Y),
331    [REG_TYPE_VIRT_VEC2C_ZW] = SWIZZLE(X, X, X, Y),
332    [REG_TYPE_VIRT_VEC3_XZW] = SWIZZLE(X, Y, Y, Z),
333    [REG_TYPE_VIRT_VEC3_YZW] = SWIZZLE(X, X, Y, Z),
334    [REG_TYPE_VIRT_VEC3C_YZW] = SWIZZLE(X, X, Y, Z),
335 };
336 
337 /* nir_src to allocated register */
338 static hw_src
ra_src(struct etna_compile * c,nir_src * src)339 ra_src(struct etna_compile *c, nir_src *src)
340 {
341    unsigned reg = ra_get_node_reg(c->g, c->live_map[src_index(c->impl, src)]);
342    return SRC_REG(reg_get_base(c, reg), reg_swiz[reg_get_type(reg)]);
343 }
344 
345 static hw_src
get_src(struct etna_compile * c,nir_src * src)346 get_src(struct etna_compile *c, nir_src *src)
347 {
348    if (!src->is_ssa)
349       return ra_src(c, src);
350 
351    nir_instr *instr = src->ssa->parent_instr;
352 
353    if (instr->pass_flags & BYPASS_SRC) {
354       assert(instr->type == nir_instr_type_alu);
355       nir_alu_instr *alu = nir_instr_as_alu(instr);
356       assert(alu->op == nir_op_mov);
357       return src_swizzle(get_src(c, &alu->src[0].src), ALU_SWIZ(&alu->src[0]));
358    }
359 
360    switch (instr->type) {
361    case nir_instr_type_load_const:
362       return const_src(c, nir_instr_as_load_const(instr)->value, src->ssa->num_components);
363    case nir_instr_type_intrinsic: {
364       nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
365       switch (intr->intrinsic) {
366       case nir_intrinsic_load_input:
367       case nir_intrinsic_load_instance_id:
368       case nir_intrinsic_load_uniform:
369       case nir_intrinsic_load_ubo:
370          return ra_src(c, src);
371       case nir_intrinsic_load_front_face:
372          return (hw_src) { .use = 1, .rgroup = INST_RGROUP_INTERNAL };
373       case nir_intrinsic_load_frag_coord:
374          return SRC_REG(0, INST_SWIZ_IDENTITY);
375       case nir_intrinsic_load_texture_rect_scaling: {
376          int sampler = nir_src_as_int(intr->src[0]);
377          nir_const_value values[] = {
378             TEXSCALE(sampler, 0),
379             TEXSCALE(sampler, 1),
380          };
381 
382          return src_swizzle(const_src(c, values, 2), SWIZZLE(X,Y,X,X));
383       }
384       default:
385          compile_error(c, "Unhandled NIR intrinsic type: %s\n",
386                        nir_intrinsic_infos[intr->intrinsic].name);
387          break;
388       }
389    } break;
390    case nir_instr_type_alu:
391    case nir_instr_type_tex:
392       return ra_src(c, src);
393    case nir_instr_type_ssa_undef: {
394       /* return zero to deal with broken Blur demo */
395       nir_const_value value = CONST(0);
396       return src_swizzle(const_src(c, &value, 1), SWIZZLE(X,X,X,X));
397    }
398    default:
399       compile_error(c, "Unhandled NIR instruction type: %d\n", instr->type);
400       break;
401    }
402 
403    return SRC_DISABLE;
404 }
405 
406 static bool
vec_dest_has_swizzle(nir_alu_instr * vec,nir_ssa_def * ssa)407 vec_dest_has_swizzle(nir_alu_instr *vec, nir_ssa_def *ssa)
408 {
409    for (unsigned i = 0; i < 4; i++) {
410       if (!(vec->dest.write_mask & (1 << i)) || vec->src[i].src.ssa != ssa)
411          continue;
412 
413       if (vec->src[i].swizzle[0] != i)
414          return true;
415    }
416 
417    /* don't deal with possible bypassed vec/mov chain */
418    nir_foreach_use(use_src, ssa) {
419       nir_instr *instr = use_src->parent_instr;
420       if (instr->type != nir_instr_type_alu)
421          continue;
422 
423       nir_alu_instr *alu = nir_instr_as_alu(instr);
424 
425       switch (alu->op) {
426       case nir_op_mov:
427       case nir_op_vec2:
428       case nir_op_vec3:
429       case nir_op_vec4:
430          return true;
431       default:
432          break;
433       }
434    }
435    return false;
436 }
437 
438 /* get allocated dest register for nir_dest
439  * *p_swiz tells how the components need to be placed into register
440  */
441 static hw_dst
ra_dest(struct etna_compile * c,nir_dest * dest,unsigned * p_swiz)442 ra_dest(struct etna_compile *c, nir_dest *dest, unsigned *p_swiz)
443 {
444    unsigned swiz = INST_SWIZ_IDENTITY, mask = 0xf;
445    dest = real_dest(dest, &swiz, &mask);
446 
447    unsigned r = ra_get_node_reg(c->g, c->live_map[dest_index(c->impl, dest)]);
448    unsigned t = reg_get_type(r);
449 
450    *p_swiz = inst_swiz_compose(swiz, reg_dst_swiz[t]);
451 
452    return (hw_dst) {
453       .use = 1,
454       .reg = reg_get_base(c, r),
455       .write_mask = inst_write_mask_compose(mask, reg_writemask[t]),
456    };
457 }
458 
459 static void
emit_alu(struct etna_compile * c,nir_alu_instr * alu)460 emit_alu(struct etna_compile *c, nir_alu_instr * alu)
461 {
462    const nir_op_info *info = &nir_op_infos[alu->op];
463 
464    /* marked as dead instruction (vecN and other bypassed instr) */
465    if (alu->instr.pass_flags)
466       return;
467 
468    assert(!(alu->op >= nir_op_vec2 && alu->op <= nir_op_vec4));
469 
470    unsigned dst_swiz;
471    hw_dst dst = ra_dest(c, &alu->dest.dest, &dst_swiz);
472 
473    /* compose alu write_mask with RA write mask */
474    if (!alu->dest.dest.is_ssa)
475       dst.write_mask = inst_write_mask_compose(alu->dest.write_mask, dst.write_mask);
476 
477    switch (alu->op) {
478    case nir_op_fdot2:
479    case nir_op_fdot3:
480    case nir_op_fdot4:
481       /* not per-component - don't compose dst_swiz */
482       dst_swiz = INST_SWIZ_IDENTITY;
483       break;
484    default:
485       break;
486    }
487 
488    hw_src srcs[3];
489 
490    for (int i = 0; i < info->num_inputs; i++) {
491       nir_alu_src *asrc = &alu->src[i];
492       hw_src src;
493 
494       src = src_swizzle(get_src(c, &asrc->src), ALU_SWIZ(asrc));
495       src = src_swizzle(src, dst_swiz);
496 
497       if (src.rgroup != INST_RGROUP_IMMEDIATE) {
498          src.neg = asrc->negate || (alu->op == nir_op_fneg);
499          src.abs = asrc->abs || (alu->op == nir_op_fabs);
500       } else {
501          assert(!asrc->negate && alu->op != nir_op_fneg);
502          assert(!asrc->abs && alu->op != nir_op_fabs);
503       }
504 
505       srcs[i] = src;
506    }
507 
508    etna_emit_alu(c, alu->op, dst, srcs, alu->dest.saturate || (alu->op == nir_op_fsat));
509 }
510 
511 static void
emit_tex(struct etna_compile * c,nir_tex_instr * tex)512 emit_tex(struct etna_compile *c, nir_tex_instr * tex)
513 {
514    unsigned dst_swiz;
515    hw_dst dst = ra_dest(c, &tex->dest, &dst_swiz);
516    nir_src *coord = NULL, *src1 = NULL, *src2 = NULL;
517 
518    for (unsigned i = 0; i < tex->num_srcs; i++) {
519       switch (tex->src[i].src_type) {
520       case nir_tex_src_coord:
521          coord = &tex->src[i].src;
522          break;
523       case nir_tex_src_bias:
524       case nir_tex_src_lod:
525       case nir_tex_src_ddx:
526          assert(!src1);
527          src1 = &tex->src[i].src;
528          break;
529       case nir_tex_src_comparator:
530       case nir_tex_src_ddy:
531          src2 = &tex->src[i].src;
532          break;
533       default:
534          compile_error(c, "Unhandled NIR tex src type: %d\n",
535                        tex->src[i].src_type);
536          break;
537       }
538    }
539 
540    etna_emit_tex(c, tex->op, tex->sampler_index, dst_swiz, dst, get_src(c, coord),
541                  src1 ? get_src(c, src1) : SRC_DISABLE,
542                  src2 ? get_src(c, src2) : SRC_DISABLE);
543 }
544 
545 static void
emit_intrinsic(struct etna_compile * c,nir_intrinsic_instr * intr)546 emit_intrinsic(struct etna_compile *c, nir_intrinsic_instr * intr)
547 {
548    switch (intr->intrinsic) {
549    case nir_intrinsic_store_deref:
550       etna_emit_output(c, nir_src_as_deref(intr->src[0])->var, get_src(c, &intr->src[1]));
551       break;
552    case nir_intrinsic_discard_if:
553       etna_emit_discard(c, get_src(c, &intr->src[0]));
554       break;
555    case nir_intrinsic_discard:
556       etna_emit_discard(c, SRC_DISABLE);
557       break;
558    case nir_intrinsic_load_uniform: {
559       unsigned dst_swiz;
560       struct etna_inst_dst dst = ra_dest(c, &intr->dest, &dst_swiz);
561 
562       /* TODO: rework so extra MOV isn't required, load up to 4 addresses at once */
563       emit_inst(c, &(struct etna_inst) {
564          .opcode = INST_OPCODE_MOVAR,
565          .dst.write_mask = 0x1,
566          .src[2] = get_src(c, &intr->src[0]),
567       });
568       emit_inst(c, &(struct etna_inst) {
569          .opcode = INST_OPCODE_MOV,
570          .dst = dst,
571          .src[2] = {
572             .use = 1,
573             .rgroup = INST_RGROUP_UNIFORM_0,
574             .reg = nir_intrinsic_base(intr),
575             .swiz = dst_swiz,
576             .amode = INST_AMODE_ADD_A_X,
577          },
578       });
579    } break;
580    case nir_intrinsic_load_ubo: {
581       /* TODO: if offset is of the form (x + C) then add C to the base instead */
582       unsigned idx = nir_src_as_const_value(intr->src[0])[0].u32;
583       unsigned dst_swiz;
584       emit_inst(c, &(struct etna_inst) {
585          .opcode = INST_OPCODE_LOAD,
586          .type = INST_TYPE_U32,
587          .dst = ra_dest(c, &intr->dest, &dst_swiz),
588          .src[0] = get_src(c, &intr->src[1]),
589          .src[1] = const_src(c, &CONST_VAL(ETNA_UNIFORM_UBO0_ADDR + idx, 0), 1),
590       });
591    } break;
592    case nir_intrinsic_load_front_face:
593    case nir_intrinsic_load_frag_coord:
594       assert(intr->dest.is_ssa); /* TODO - lower phis could cause this */
595       break;
596    case nir_intrinsic_load_input:
597    case nir_intrinsic_load_instance_id:
598    case nir_intrinsic_load_texture_rect_scaling:
599       break;
600    default:
601       compile_error(c, "Unhandled NIR intrinsic type: %s\n",
602                     nir_intrinsic_infos[intr->intrinsic].name);
603    }
604 }
605 
606 static void
emit_instr(struct etna_compile * c,nir_instr * instr)607 emit_instr(struct etna_compile *c, nir_instr * instr)
608 {
609    switch (instr->type) {
610    case nir_instr_type_alu:
611       emit_alu(c, nir_instr_as_alu(instr));
612       break;
613    case nir_instr_type_tex:
614       emit_tex(c, nir_instr_as_tex(instr));
615       break;
616    case nir_instr_type_intrinsic:
617       emit_intrinsic(c, nir_instr_as_intrinsic(instr));
618       break;
619    case nir_instr_type_jump:
620       assert(nir_instr_is_last(instr));
621       break;
622    case nir_instr_type_load_const:
623    case nir_instr_type_ssa_undef:
624    case nir_instr_type_deref:
625       break;
626    default:
627       compile_error(c, "Unhandled NIR instruction type: %d\n", instr->type);
628       break;
629    }
630 }
631 
632 static void
emit_block(struct etna_compile * c,nir_block * block)633 emit_block(struct etna_compile *c, nir_block * block)
634 {
635    etna_emit_block_start(c, block->index);
636 
637    nir_foreach_instr(instr, block)
638       emit_instr(c, instr);
639 
640    /* succs->index < block->index is for the loop case  */
641    nir_block *succs = block->successors[0];
642    if (nir_block_ends_in_jump(block) || succs->index < block->index)
643       etna_emit_jump(c, succs->index, SRC_DISABLE);
644 }
645 
646 static void
647 emit_cf_list(struct etna_compile *c, struct exec_list *list);
648 
649 static void
emit_if(struct etna_compile * c,nir_if * nif)650 emit_if(struct etna_compile *c, nir_if * nif)
651 {
652    etna_emit_jump(c, nir_if_first_else_block(nif)->index, get_src(c, &nif->condition));
653    emit_cf_list(c, &nif->then_list);
654 
655    /* jump at end of then_list to skip else_list
656     * not needed if then_list already ends with a jump or else_list is empty
657     */
658    if (!nir_block_ends_in_jump(nir_if_last_then_block(nif)) &&
659        !nir_cf_list_is_empty_block(&nif->else_list))
660       etna_emit_jump(c, nir_if_last_then_block(nif)->successors[0]->index, SRC_DISABLE);
661 
662    emit_cf_list(c, &nif->else_list);
663 }
664 
665 static void
emit_cf_list(struct etna_compile * c,struct exec_list * list)666 emit_cf_list(struct etna_compile *c, struct exec_list *list)
667 {
668    foreach_list_typed(nir_cf_node, node, node, list) {
669       switch (node->type) {
670       case nir_cf_node_block:
671          emit_block(c, nir_cf_node_as_block(node));
672          break;
673       case nir_cf_node_if:
674          emit_if(c, nir_cf_node_as_if(node));
675          break;
676       case nir_cf_node_loop:
677          emit_cf_list(c, &nir_cf_node_as_loop(node)->body);
678          break;
679       default:
680          compile_error(c, "Unknown NIR node type\n");
681          break;
682       }
683    }
684 }
685 
686 /* based on nir_lower_vec_to_movs */
687 static unsigned
insert_vec_mov(nir_alu_instr * vec,unsigned start_idx,nir_shader * shader)688 insert_vec_mov(nir_alu_instr *vec, unsigned start_idx, nir_shader *shader)
689 {
690    assert(start_idx < nir_op_infos[vec->op].num_inputs);
691    unsigned write_mask = (1u << start_idx);
692 
693    nir_alu_instr *mov = nir_alu_instr_create(shader, nir_op_mov);
694    nir_alu_src_copy(&mov->src[0], &vec->src[start_idx]);
695 
696    mov->src[0].swizzle[0] = vec->src[start_idx].swizzle[0];
697    mov->src[0].negate = vec->src[start_idx].negate;
698    mov->src[0].abs = vec->src[start_idx].abs;
699 
700    unsigned num_components = 1;
701 
702    for (unsigned i = start_idx + 1; i < 4; i++) {
703       if (!(vec->dest.write_mask & (1 << i)))
704          continue;
705 
706       if (nir_srcs_equal(vec->src[i].src, vec->src[start_idx].src) &&
707           vec->src[i].negate == vec->src[start_idx].negate &&
708           vec->src[i].abs == vec->src[start_idx].abs) {
709          write_mask |= (1 << i);
710          mov->src[0].swizzle[num_components] = vec->src[i].swizzle[0];
711          num_components++;
712       }
713    }
714 
715    mov->dest.write_mask = (1 << num_components) - 1;
716    nir_ssa_dest_init(&mov->instr, &mov->dest.dest, num_components, 32, NULL);
717 
718    /* replace vec srcs with inserted mov */
719    for (unsigned i = 0, j = 0; i < 4; i++) {
720       if (!(write_mask & (1 << i)))
721          continue;
722 
723       nir_instr_rewrite_src(&vec->instr, &vec->src[i].src, nir_src_for_ssa(&mov->dest.dest.ssa));
724       vec->src[i].swizzle[0] = j++;
725    }
726 
727    nir_instr_insert_before(&vec->instr, &mov->instr);
728 
729    return write_mask;
730 }
731 
732 /*
733  * for vecN instructions:
734  * -merge constant sources into a single src
735  * -insert movs (nir_lower_vec_to_movs equivalent)
736  * for non-vecN instructions:
737  * -try to merge constants as single constant
738  * -insert movs for multiple constants if required
739  */
740 static void
lower_alu(struct etna_compile * c,nir_alu_instr * alu)741 lower_alu(struct etna_compile *c, nir_alu_instr *alu)
742 {
743    const nir_op_info *info = &nir_op_infos[alu->op];
744 
745    nir_builder b;
746    nir_builder_init(&b, c->impl);
747    b.cursor = nir_before_instr(&alu->instr);
748 
749    switch (alu->op) {
750    case nir_op_vec2:
751    case nir_op_vec3:
752    case nir_op_vec4:
753       break;
754    default:
755       if (c->specs->has_no_oneconst_limit)
756          return;
757 
758       nir_const_value value[4] = {};
759       uint8_t swizzle[4][4] = {};
760       unsigned swiz_max = 0, num_const = 0;
761 
762       for (unsigned i = 0; i < info->num_inputs; i++) {
763          nir_const_value *cv = nir_src_as_const_value(alu->src[i].src);
764          if (!cv)
765             continue;
766 
767          unsigned num_components = info->input_sizes[i] ?: alu->dest.dest.ssa.num_components;
768          for (unsigned j = 0; j < num_components; j++) {
769             int idx = const_add(&value[0].u64, cv[alu->src[i].swizzle[j]].u64);
770             swizzle[i][j] = idx;
771             swiz_max = MAX2(swiz_max, (unsigned) idx);
772          }
773          num_const++;
774       }
775 
776       /* nothing to do */
777       if (num_const <= 1)
778          return;
779 
780       /* resolve with single combined const src */
781       if (swiz_max < 4) {
782          nir_ssa_def *def = nir_build_imm(&b, swiz_max + 1, 32, value);
783 
784          for (unsigned i = 0; i < info->num_inputs; i++) {
785             nir_const_value *cv = nir_src_as_const_value(alu->src[i].src);
786             if (!cv)
787                continue;
788 
789             nir_instr_rewrite_src(&alu->instr, &alu->src[i].src, nir_src_for_ssa(def));
790 
791             for (unsigned j = 0; j < 4; j++)
792                alu->src[i].swizzle[j] = swizzle[i][j];
793          }
794          return;
795       }
796 
797       /* resolve with movs */
798       num_const = 0;
799       for (unsigned i = 0; i < info->num_inputs; i++) {
800          nir_const_value *cv = nir_src_as_const_value(alu->src[i].src);
801          if (!cv)
802             continue;
803 
804          num_const++;
805          if (num_const == 1)
806             continue;
807 
808          nir_ssa_def *mov = nir_mov(&b, alu->src[i].src.ssa);
809          nir_instr_rewrite_src(&alu->instr, &alu->src[i].src, nir_src_for_ssa(mov));
810       }
811       return;
812    }
813 
814    nir_const_value value[4];
815    unsigned num_components = 0;
816 
817    for (unsigned i = 0; i < info->num_inputs; i++) {
818       nir_const_value *cv = nir_src_as_const_value(alu->src[i].src);
819       if (cv)
820          value[num_components++] = cv[alu->src[i].swizzle[0]];
821    }
822 
823    /* if there is more than one constant source to the vecN, combine them
824     * into a single load_const (removing the vecN completely if all components
825     * are constant)
826     */
827    if (num_components > 1) {
828       nir_ssa_def *def = nir_build_imm(&b, num_components, 32, value);
829 
830       if (num_components == info->num_inputs) {
831          nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, def);
832          nir_instr_remove(&alu->instr);
833          return;
834       }
835 
836       for (unsigned i = 0, j = 0; i < info->num_inputs; i++) {
837          nir_const_value *cv = nir_src_as_const_value(alu->src[i].src);
838          if (!cv)
839             continue;
840 
841          nir_instr_rewrite_src(&alu->instr, &alu->src[i].src, nir_src_for_ssa(def));
842          alu->src[i].swizzle[0] = j++;
843       }
844    }
845 
846    unsigned finished_write_mask = 0;
847    for (unsigned i = 0; i < 4; i++) {
848       if (!(alu->dest.write_mask & (1 << i)))
849             continue;
850 
851       nir_ssa_def *ssa = alu->src[i].src.ssa;
852 
853       /* check that vecN instruction is only user of this */
854       bool need_mov = list_length(&ssa->if_uses) != 0;
855       nir_foreach_use(use_src, ssa) {
856          if (use_src->parent_instr != &alu->instr)
857             need_mov = true;
858       }
859 
860       nir_instr *instr = ssa->parent_instr;
861       switch (instr->type) {
862       case nir_instr_type_alu:
863       case nir_instr_type_tex:
864          break;
865       case nir_instr_type_intrinsic:
866          if (nir_instr_as_intrinsic(instr)->intrinsic == nir_intrinsic_load_input) {
867             need_mov = vec_dest_has_swizzle(alu, &nir_instr_as_intrinsic(instr)->dest.ssa);
868             break;
869          }
870          FALLTHROUGH;
871       default:
872          need_mov = true;
873       }
874 
875       if (need_mov && !(finished_write_mask & (1 << i)))
876          finished_write_mask |= insert_vec_mov(alu, i, c->nir);
877    }
878 }
879 
880 static bool
emit_shader(struct etna_compile * c,unsigned * num_temps,unsigned * num_consts)881 emit_shader(struct etna_compile *c, unsigned *num_temps, unsigned *num_consts)
882 {
883    nir_shader *shader = c->nir;
884    c->impl = nir_shader_get_entrypoint(shader);
885 
886    bool have_indirect_uniform = false;
887    unsigned indirect_max = 0;
888 
889    nir_builder b;
890    nir_builder_init(&b, c->impl);
891 
892    /* convert non-dynamic uniform loads to constants, etc */
893    nir_foreach_block(block, c->impl) {
894       nir_foreach_instr_safe(instr, block) {
895          switch(instr->type) {
896          case nir_instr_type_alu:
897             /* deals with vecN and const srcs */
898             lower_alu(c, nir_instr_as_alu(instr));
899             break;
900          case nir_instr_type_load_const: {
901             nir_load_const_instr *load_const = nir_instr_as_load_const(instr);
902             for (unsigned  i = 0; i < load_const->def.num_components; i++)
903                load_const->value[i] = CONST(load_const->value[i].u32);
904          } break;
905          case nir_instr_type_intrinsic: {
906             nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
907             /* TODO: load_ubo can also become a constant in some cases
908              * (at the moment it can end up emitting a LOAD with two
909              *  uniform sources, which could be a problem on HALTI2)
910              */
911             if (intr->intrinsic != nir_intrinsic_load_uniform)
912                break;
913             nir_const_value *off = nir_src_as_const_value(intr->src[0]);
914             if (!off || off[0].u64 >> 32 != ETNA_UNIFORM_CONSTANT) {
915                have_indirect_uniform = true;
916                indirect_max = nir_intrinsic_base(intr) + nir_intrinsic_range(intr);
917                break;
918             }
919 
920             unsigned base = nir_intrinsic_base(intr);
921             /* pre halti2 uniform offset will be float */
922             if (c->specs->halti < 2)
923                base += (unsigned) off[0].f32;
924             else
925                base += off[0].u32;
926             nir_const_value value[4];
927 
928             for (unsigned i = 0; i < intr->dest.ssa.num_components; i++)
929                value[i] = UNIFORM(base * 4 + i);
930 
931             b.cursor = nir_after_instr(instr);
932             nir_ssa_def *def = nir_build_imm(&b, intr->dest.ssa.num_components, 32, value);
933 
934             nir_ssa_def_rewrite_uses(&intr->dest.ssa, def);
935             nir_instr_remove(instr);
936          } break;
937          default:
938             break;
939          }
940       }
941    }
942 
943    /* TODO: only emit required indirect uniform ranges */
944    if (have_indirect_uniform) {
945       for (unsigned i = 0; i < indirect_max * 4; i++)
946          c->consts[i] = UNIFORM(i).u64;
947       c->const_count = indirect_max;
948    }
949 
950    /* add mov for any store output using sysval/const and for depth stores from intrinsics */
951    nir_foreach_block(block, c->impl) {
952       nir_foreach_instr_safe(instr, block) {
953          if (instr->type != nir_instr_type_intrinsic)
954             continue;
955 
956          nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
957 
958          switch (intr->intrinsic) {
959          case nir_intrinsic_store_deref: {
960             nir_deref_instr *deref = nir_src_as_deref(intr->src[0]);
961             nir_src *src = &intr->src[1];
962             if (nir_src_is_const(*src) || is_sysval(src->ssa->parent_instr) ||
963                 (shader->info.stage == MESA_SHADER_FRAGMENT &&
964                  deref->var->data.location == FRAG_RESULT_DEPTH &&
965                  src->is_ssa &&
966                  src->ssa->parent_instr->type != nir_instr_type_alu)) {
967                b.cursor = nir_before_instr(instr);
968                nir_instr_rewrite_src(instr, src, nir_src_for_ssa(nir_mov(&b, src->ssa)));
969             }
970          } break;
971          default:
972             break;
973          }
974       }
975    }
976 
977    /* call directly to avoid validation (load_const don't pass validation at this point) */
978    nir_convert_from_ssa(shader, true);
979    nir_opt_dce(shader);
980 
981    etna_ra_assign(c, shader);
982 
983    emit_cf_list(c, &nir_shader_get_entrypoint(shader)->body);
984 
985    *num_temps = etna_ra_finish(c);
986    *num_consts = c->const_count;
987    return true;
988 }
989 
990 static bool
etna_compile_check_limits(struct etna_shader_variant * v)991 etna_compile_check_limits(struct etna_shader_variant *v)
992 {
993    const struct etna_specs *specs = v->shader->specs;
994    int max_uniforms = (v->stage == MESA_SHADER_VERTEX)
995                          ? specs->max_vs_uniforms
996                          : specs->max_ps_uniforms;
997 
998    if (!specs->has_icache && v->needs_icache) {
999       DBG("Number of instructions (%d) exceeds maximum %d", v->code_size / 4,
1000           specs->max_instructions);
1001       return false;
1002    }
1003 
1004    if (v->num_temps > specs->max_registers) {
1005       DBG("Number of registers (%d) exceeds maximum %d", v->num_temps,
1006           specs->max_registers);
1007       return false;
1008    }
1009 
1010    if (v->uniforms.count / 4 > max_uniforms) {
1011       DBG("Number of uniforms (%d) exceeds maximum %d",
1012           v->uniforms.count / 4, max_uniforms);
1013       return false;
1014    }
1015 
1016    return true;
1017 }
1018 
1019 static void
fill_vs_mystery(struct etna_shader_variant * v)1020 fill_vs_mystery(struct etna_shader_variant *v)
1021 {
1022    const struct etna_specs *specs = v->shader->specs;
1023 
1024    v->input_count_unk8 = DIV_ROUND_UP(v->infile.num_reg + 4, 16); /* XXX what is this */
1025 
1026    /* fill in "mystery meat" load balancing value. This value determines how
1027     * work is scheduled between VS and PS
1028     * in the unified shader architecture. More precisely, it is determined from
1029     * the number of VS outputs, as well as chip-specific
1030     * vertex output buffer size, vertex cache size, and the number of shader
1031     * cores.
1032     *
1033     * XXX this is a conservative estimate, the "optimal" value is only known for
1034     * sure at link time because some
1035     * outputs may be unused and thus unmapped. Then again, in the general use
1036     * case with GLSL the vertex and fragment
1037     * shaders are linked already before submitting to Gallium, thus all outputs
1038     * are used.
1039     *
1040     * note: TGSI compiler counts all outputs (including position and pointsize), here
1041     * v->outfile.num_reg only counts varyings, +1 to compensate for the position output
1042     * TODO: might have a problem that we don't count pointsize when it is used
1043     */
1044 
1045    int half_out = v->outfile.num_reg / 2 + 1;
1046    assert(half_out);
1047 
1048    uint32_t b = ((20480 / (specs->vertex_output_buffer_size -
1049                            2 * half_out * specs->vertex_cache_size)) +
1050                  9) /
1051                 10;
1052    uint32_t a = (b + 256 / (specs->shader_core_count * half_out)) / 2;
1053    v->vs_load_balancing = VIVS_VS_LOAD_BALANCING_A(MIN2(a, 255)) |
1054                              VIVS_VS_LOAD_BALANCING_B(MIN2(b, 255)) |
1055                              VIVS_VS_LOAD_BALANCING_C(0x3f) |
1056                              VIVS_VS_LOAD_BALANCING_D(0x0f);
1057 }
1058 
1059 bool
etna_compile_shader(struct etna_shader_variant * v)1060 etna_compile_shader(struct etna_shader_variant *v)
1061 {
1062    if (unlikely(!v))
1063       return false;
1064 
1065    struct etna_compile *c = CALLOC_STRUCT(etna_compile);
1066    if (!c)
1067       return false;
1068 
1069    c->variant = v;
1070    c->specs = v->shader->specs;
1071    c->nir = nir_shader_clone(NULL, v->shader->nir);
1072 
1073    nir_shader *s = c->nir;
1074    const struct etna_specs *specs = c->specs;
1075 
1076    v->stage = s->info.stage;
1077    v->uses_discard = s->info.fs.uses_discard;
1078    v->num_loops = 0; /* TODO */
1079    v->vs_id_in_reg = -1;
1080    v->vs_pos_out_reg = -1;
1081    v->vs_pointsize_out_reg = -1;
1082    v->ps_color_out_reg = 0; /* 0 for shader that doesn't write fragcolor.. */
1083    v->ps_depth_out_reg = -1;
1084 
1085    /*
1086     * Lower glTexCoord, fixes e.g. neverball point sprite (exit cylinder stars)
1087     * and gl4es pointsprite.trace apitrace
1088     */
1089    if (s->info.stage == MESA_SHADER_FRAGMENT && v->key.sprite_coord_enable) {
1090       NIR_PASS_V(s, nir_lower_texcoord_replace, v->key.sprite_coord_enable,
1091                  false, v->key.sprite_coord_yinvert);
1092    }
1093 
1094    /* setup input linking */
1095    struct etna_shader_io_file *sf = &v->infile;
1096    if (s->info.stage == MESA_SHADER_VERTEX) {
1097       nir_foreach_shader_in_variable(var, s) {
1098          unsigned idx = var->data.driver_location;
1099          sf->reg[idx].reg = idx;
1100          sf->reg[idx].slot = var->data.location;
1101          sf->reg[idx].num_components = glsl_get_components(var->type);
1102          sf->num_reg = MAX2(sf->num_reg, idx+1);
1103       }
1104    } else {
1105       unsigned count = 0;
1106       nir_foreach_shader_in_variable(var, s) {
1107          unsigned idx = var->data.driver_location;
1108          sf->reg[idx].reg = idx + 1;
1109          sf->reg[idx].slot = var->data.location;
1110          sf->reg[idx].num_components = glsl_get_components(var->type);
1111          sf->num_reg = MAX2(sf->num_reg, idx+1);
1112          count++;
1113       }
1114       assert(sf->num_reg == count);
1115    }
1116 
1117    NIR_PASS_V(s, nir_lower_io, nir_var_shader_in | nir_var_uniform, etna_glsl_type_size,
1118             (nir_lower_io_options)0);
1119 
1120    NIR_PASS_V(s, nir_lower_regs_to_ssa);
1121    NIR_PASS_V(s, nir_lower_vars_to_ssa);
1122    NIR_PASS_V(s, nir_lower_indirect_derefs, nir_var_all, UINT32_MAX);
1123    NIR_PASS_V(s, nir_lower_tex, &(struct nir_lower_tex_options) { .lower_txp = ~0u, .lower_invalid_implicit_lod = true, });
1124 
1125    if (v->key.has_sample_tex_compare)
1126       NIR_PASS_V(s, nir_lower_tex_shadow, v->key.num_texture_states,
1127                                           v->key.tex_compare_func,
1128                                           v->key.tex_swizzle);
1129 
1130    NIR_PASS_V(s, nir_lower_alu_to_scalar, etna_alu_to_scalar_filter_cb, specs);
1131    nir_lower_idiv_options idiv_options = {
1132       .imprecise_32bit_lowering = true,
1133       .allow_fp16 = true,
1134    };
1135    NIR_PASS_V(s, nir_lower_idiv, &idiv_options);
1136 
1137    etna_optimize_loop(s);
1138 
1139    /* TODO: remove this extra run if nir_opt_peephole_select is able to handle ubo's. */
1140    if (OPT(s, etna_nir_lower_ubo_to_uniform))
1141       etna_optimize_loop(s);
1142 
1143    NIR_PASS_V(s, etna_lower_io, v);
1144 
1145    if (v->shader->specs->vs_need_z_div)
1146       NIR_PASS_V(s, nir_lower_clip_halfz);
1147 
1148    /* lower pre-halti2 to float (halti0 has integers, but only scalar..) */
1149    if (c->specs->halti < 2) {
1150       /* use opt_algebraic between int_to_float and boot_to_float because
1151        * int_to_float emits ftrunc, and ftrunc lowering generates bool ops
1152        */
1153       NIR_PASS_V(s, nir_lower_int_to_float);
1154       NIR_PASS_V(s, nir_opt_algebraic);
1155       NIR_PASS_V(s, nir_lower_bool_to_float);
1156    } else {
1157       NIR_PASS_V(s, nir_lower_bool_to_int32);
1158    }
1159 
1160    while( OPT(s, nir_opt_vectorize, NULL, NULL) );
1161    NIR_PASS_V(s, nir_lower_alu_to_scalar, etna_alu_to_scalar_filter_cb, specs);
1162 
1163    NIR_PASS_V(s, nir_remove_dead_variables, nir_var_function_temp, NULL);
1164    NIR_PASS_V(s, nir_opt_algebraic_late);
1165 
1166    NIR_PASS_V(s, nir_move_vec_src_uses_to_dest);
1167    NIR_PASS_V(s, nir_copy_prop);
1168    /* only HW supported integer source mod is ineg for iadd instruction (?) */
1169    NIR_PASS_V(s, nir_lower_to_source_mods, ~nir_lower_int_source_mods);
1170    /* need copy prop after uses_to_dest, and before src mods: see
1171     * dEQP-GLES2.functional.shaders.random.all_features.fragment.95
1172     */
1173 
1174    NIR_PASS_V(s, nir_opt_dce);
1175 
1176    NIR_PASS_V(s, nir_lower_bool_to_bitsize);
1177    NIR_PASS_V(s, etna_lower_alu, c->specs->has_new_transcendentals);
1178 
1179    if (DBG_ENABLED(ETNA_DBG_DUMP_SHADERS))
1180       nir_print_shader(s, stdout);
1181 
1182    unsigned block_ptr[nir_shader_get_entrypoint(s)->num_blocks];
1183    c->block_ptr = block_ptr;
1184 
1185    unsigned num_consts;
1186    ASSERTED bool ok = emit_shader(c, &v->num_temps, &num_consts);
1187    assert(ok);
1188 
1189    /* empty shader, emit NOP */
1190    if (!c->inst_ptr)
1191       emit_inst(c, &(struct etna_inst) { .opcode = INST_OPCODE_NOP });
1192 
1193    /* assemble instructions, fixing up labels */
1194    uint32_t *code = MALLOC(c->inst_ptr * 16);
1195    for (unsigned i = 0; i < c->inst_ptr; i++) {
1196       struct etna_inst *inst = &c->code[i];
1197       if (inst->opcode == INST_OPCODE_BRANCH)
1198          inst->imm = block_ptr[inst->imm];
1199 
1200       inst->no_oneconst_limit = specs->has_no_oneconst_limit;
1201       etna_assemble(&code[i * 4], inst);
1202    }
1203 
1204    v->code_size = c->inst_ptr * 4;
1205    v->code = code;
1206    v->needs_icache = c->inst_ptr > specs->max_instructions;
1207 
1208    copy_uniform_state_to_shader(v, c->consts, num_consts);
1209 
1210    if (s->info.stage == MESA_SHADER_FRAGMENT) {
1211       v->input_count_unk8 = 31; /* XXX what is this */
1212       assert(v->ps_depth_out_reg <= 0);
1213    } else {
1214       fill_vs_mystery(v);
1215    }
1216 
1217    bool result = etna_compile_check_limits(v);
1218    ralloc_free(c->nir);
1219    FREE(c);
1220    return result;
1221 }
1222 
1223 static const struct etna_shader_inout *
etna_shader_vs_lookup(const struct etna_shader_variant * sobj,const struct etna_shader_inout * in)1224 etna_shader_vs_lookup(const struct etna_shader_variant *sobj,
1225                       const struct etna_shader_inout *in)
1226 {
1227    for (int i = 0; i < sobj->outfile.num_reg; i++)
1228       if (sobj->outfile.reg[i].slot == in->slot)
1229          return &sobj->outfile.reg[i];
1230 
1231    return NULL;
1232 }
1233 
1234 bool
etna_link_shader(struct etna_shader_link_info * info,const struct etna_shader_variant * vs,const struct etna_shader_variant * fs)1235 etna_link_shader(struct etna_shader_link_info *info,
1236                  const struct etna_shader_variant *vs,
1237                  const struct etna_shader_variant *fs)
1238 {
1239    int comp_ofs = 0;
1240    /* For each fragment input we need to find the associated vertex shader
1241     * output, which can be found by matching on semantic name and index. A
1242     * binary search could be used because the vs outputs are sorted by their
1243     * semantic index and grouped by semantic type by fill_in_vs_outputs.
1244     */
1245    assert(fs->infile.num_reg < ETNA_NUM_INPUTS);
1246    info->pcoord_varying_comp_ofs = -1;
1247 
1248    for (int idx = 0; idx < fs->infile.num_reg; ++idx) {
1249       const struct etna_shader_inout *fsio = &fs->infile.reg[idx];
1250       const struct etna_shader_inout *vsio = etna_shader_vs_lookup(vs, fsio);
1251       struct etna_varying *varying;
1252       bool interpolate_always = true;
1253 
1254       assert(fsio->reg > 0 && fsio->reg <= ARRAY_SIZE(info->varyings));
1255 
1256       if (fsio->reg > info->num_varyings)
1257          info->num_varyings = fsio->reg;
1258 
1259       varying = &info->varyings[fsio->reg - 1];
1260       varying->num_components = fsio->num_components;
1261 
1262       if (!interpolate_always) /* colors affected by flat shading */
1263          varying->pa_attributes = 0x200;
1264       else /* texture coord or other bypasses flat shading */
1265          varying->pa_attributes = 0x2f1;
1266 
1267       varying->use[0] = VARYING_COMPONENT_USE_UNUSED;
1268       varying->use[1] = VARYING_COMPONENT_USE_UNUSED;
1269       varying->use[2] = VARYING_COMPONENT_USE_UNUSED;
1270       varying->use[3] = VARYING_COMPONENT_USE_UNUSED;
1271 
1272       /* point/tex coord is an input to the PS without matching VS output,
1273        * so it gets a varying slot without being assigned a VS register.
1274        */
1275       if (fsio->slot == VARYING_SLOT_PNTC) {
1276          varying->use[0] = VARYING_COMPONENT_USE_POINTCOORD_X;
1277          varying->use[1] = VARYING_COMPONENT_USE_POINTCOORD_Y;
1278 
1279          info->pcoord_varying_comp_ofs = comp_ofs;
1280       } else if (util_varying_is_point_coord(fsio->slot, fs->key.sprite_coord_enable)) {
1281          /*
1282 	  * Do nothing, TexCoord is lowered to PointCoord above
1283 	  * and the TexCoord here is just a remnant. This needs
1284 	  * to be removed with some nir_remove_dead_variables(),
1285 	  * but that one removes all FS inputs ... why?
1286 	  */
1287       } else {
1288          if (vsio == NULL) { /* not found -- link error */
1289             BUG("Semantic value not found in vertex shader outputs\n");
1290             return true;
1291          }
1292          varying->reg = vsio->reg;
1293       }
1294 
1295       comp_ofs += varying->num_components;
1296    }
1297 
1298    assert(info->num_varyings == fs->infile.num_reg);
1299 
1300    return false;
1301 }
1302