nir_to_vir.c revision b8e80941
1/*
2 * Copyright © 2016 Broadcom
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24#include <inttypes.h>
25#include "util/u_format.h"
26#include "util/u_math.h"
27#include "util/u_memory.h"
28#include "util/ralloc.h"
29#include "util/hash_table.h"
30#include "compiler/nir/nir.h"
31#include "compiler/nir/nir_builder.h"
32#include "common/v3d_device_info.h"
33#include "v3d_compiler.h"
34
35#define GENERAL_TMU_LOOKUP_PER_QUAD                 (0 << 7)
36#define GENERAL_TMU_LOOKUP_PER_PIXEL                (1 << 7)
37#define GENERAL_TMU_READ_OP_PREFETCH                (0 << 3)
38#define GENERAL_TMU_READ_OP_CACHE_CLEAR             (1 << 3)
39#define GENERAL_TMU_READ_OP_CACHE_FLUSH             (3 << 3)
40#define GENERAL_TMU_READ_OP_CACHE_CLEAN             (3 << 3)
41#define GENERAL_TMU_READ_OP_CACHE_L1T_CLEAR         (4 << 3)
42#define GENERAL_TMU_READ_OP_CACHE_L1T_FLUSH_AGGREGATION (5 << 3)
43#define GENERAL_TMU_READ_OP_ATOMIC_INC              (8 << 3)
44#define GENERAL_TMU_READ_OP_ATOMIC_DEC              (9 << 3)
45#define GENERAL_TMU_READ_OP_ATOMIC_NOT              (10 << 3)
46#define GENERAL_TMU_READ_OP_READ                    (15 << 3)
47#define GENERAL_TMU_LOOKUP_TYPE_8BIT_I              (0 << 0)
48#define GENERAL_TMU_LOOKUP_TYPE_16BIT_I             (1 << 0)
49#define GENERAL_TMU_LOOKUP_TYPE_VEC2                (2 << 0)
50#define GENERAL_TMU_LOOKUP_TYPE_VEC3                (3 << 0)
51#define GENERAL_TMU_LOOKUP_TYPE_VEC4                (4 << 0)
52#define GENERAL_TMU_LOOKUP_TYPE_8BIT_UI             (5 << 0)
53#define GENERAL_TMU_LOOKUP_TYPE_16BIT_UI            (6 << 0)
54#define GENERAL_TMU_LOOKUP_TYPE_32BIT_UI            (7 << 0)
55
56#define GENERAL_TMU_WRITE_OP_ATOMIC_ADD_WRAP         (0 << 3)
57#define GENERAL_TMU_WRITE_OP_ATOMIC_SUB_WRAP         (1 << 3)
58#define GENERAL_TMU_WRITE_OP_ATOMIC_XCHG             (2 << 3)
59#define GENERAL_TMU_WRITE_OP_ATOMIC_CMPXCHG          (3 << 3)
60#define GENERAL_TMU_WRITE_OP_ATOMIC_UMIN             (4 << 3)
61#define GENERAL_TMU_WRITE_OP_ATOMIC_UMAX             (5 << 3)
62#define GENERAL_TMU_WRITE_OP_ATOMIC_SMIN             (6 << 3)
63#define GENERAL_TMU_WRITE_OP_ATOMIC_SMAX             (7 << 3)
64#define GENERAL_TMU_WRITE_OP_ATOMIC_AND              (8 << 3)
65#define GENERAL_TMU_WRITE_OP_ATOMIC_OR               (9 << 3)
66#define GENERAL_TMU_WRITE_OP_ATOMIC_XOR              (10 << 3)
67#define GENERAL_TMU_WRITE_OP_WRITE                   (15 << 3)
68
69#define V3D_TSY_SET_QUORUM          0
70#define V3D_TSY_INC_WAITERS         1
71#define V3D_TSY_DEC_WAITERS         2
72#define V3D_TSY_INC_QUORUM          3
73#define V3D_TSY_DEC_QUORUM          4
74#define V3D_TSY_FREE_ALL            5
75#define V3D_TSY_RELEASE             6
76#define V3D_TSY_ACQUIRE             7
77#define V3D_TSY_WAIT                8
78#define V3D_TSY_WAIT_INC            9
79#define V3D_TSY_WAIT_CHECK          10
80#define V3D_TSY_WAIT_INC_CHECK      11
81#define V3D_TSY_WAIT_CV             12
82#define V3D_TSY_INC_SEMAPHORE       13
83#define V3D_TSY_DEC_SEMAPHORE       14
84#define V3D_TSY_SET_QUORUM_FREE_ALL 15
85
86static void
87ntq_emit_cf_list(struct v3d_compile *c, struct exec_list *list);
88
89static void
90resize_qreg_array(struct v3d_compile *c,
91                  struct qreg **regs,
92                  uint32_t *size,
93                  uint32_t decl_size)
94{
95        if (*size >= decl_size)
96                return;
97
98        uint32_t old_size = *size;
99        *size = MAX2(*size * 2, decl_size);
100        *regs = reralloc(c, *regs, struct qreg, *size);
101        if (!*regs) {
102                fprintf(stderr, "Malloc failure\n");
103                abort();
104        }
105
106        for (uint32_t i = old_size; i < *size; i++)
107                (*regs)[i] = c->undef;
108}
109
110void
111vir_emit_thrsw(struct v3d_compile *c)
112{
113        if (c->threads == 1)
114                return;
115
116        /* Always thread switch after each texture operation for now.
117         *
118         * We could do better by batching a bunch of texture fetches up and
119         * then doing one thread switch and collecting all their results
120         * afterward.
121         */
122        c->last_thrsw = vir_NOP(c);
123        c->last_thrsw->qpu.sig.thrsw = true;
124        c->last_thrsw_at_top_level = !c->in_control_flow;
125}
126
127static uint32_t
128v3d_general_tmu_op(nir_intrinsic_instr *instr)
129{
130        switch (instr->intrinsic) {
131        case nir_intrinsic_load_ssbo:
132        case nir_intrinsic_load_ubo:
133        case nir_intrinsic_load_uniform:
134        case nir_intrinsic_load_shared:
135        case nir_intrinsic_load_scratch:
136                return GENERAL_TMU_READ_OP_READ;
137        case nir_intrinsic_store_ssbo:
138        case nir_intrinsic_store_shared:
139        case nir_intrinsic_store_scratch:
140                return GENERAL_TMU_WRITE_OP_WRITE;
141        case nir_intrinsic_ssbo_atomic_add:
142        case nir_intrinsic_shared_atomic_add:
143                return GENERAL_TMU_WRITE_OP_ATOMIC_ADD_WRAP;
144        case nir_intrinsic_ssbo_atomic_imin:
145        case nir_intrinsic_shared_atomic_imin:
146                return GENERAL_TMU_WRITE_OP_ATOMIC_SMIN;
147        case nir_intrinsic_ssbo_atomic_umin:
148        case nir_intrinsic_shared_atomic_umin:
149                return GENERAL_TMU_WRITE_OP_ATOMIC_UMIN;
150        case nir_intrinsic_ssbo_atomic_imax:
151        case nir_intrinsic_shared_atomic_imax:
152                return GENERAL_TMU_WRITE_OP_ATOMIC_SMAX;
153        case nir_intrinsic_ssbo_atomic_umax:
154        case nir_intrinsic_shared_atomic_umax:
155                return GENERAL_TMU_WRITE_OP_ATOMIC_UMAX;
156        case nir_intrinsic_ssbo_atomic_and:
157        case nir_intrinsic_shared_atomic_and:
158                return GENERAL_TMU_WRITE_OP_ATOMIC_AND;
159        case nir_intrinsic_ssbo_atomic_or:
160        case nir_intrinsic_shared_atomic_or:
161                return GENERAL_TMU_WRITE_OP_ATOMIC_OR;
162        case nir_intrinsic_ssbo_atomic_xor:
163        case nir_intrinsic_shared_atomic_xor:
164                return GENERAL_TMU_WRITE_OP_ATOMIC_XOR;
165        case nir_intrinsic_ssbo_atomic_exchange:
166        case nir_intrinsic_shared_atomic_exchange:
167                return GENERAL_TMU_WRITE_OP_ATOMIC_XCHG;
168        case nir_intrinsic_ssbo_atomic_comp_swap:
169        case nir_intrinsic_shared_atomic_comp_swap:
170                return GENERAL_TMU_WRITE_OP_ATOMIC_CMPXCHG;
171        default:
172                unreachable("unknown intrinsic op");
173        }
174}
175
176/**
177 * Implements indirect uniform loads and SSBO accesses through the TMU general
178 * memory access interface.
179 */
180static void
181ntq_emit_tmu_general(struct v3d_compile *c, nir_intrinsic_instr *instr,
182                     bool is_shared_or_scratch)
183{
184        /* XXX perf: We should turn add/sub of 1 to inc/dec.  Perhaps NIR
185         * wants to have support for inc/dec?
186         */
187
188        uint32_t tmu_op = v3d_general_tmu_op(instr);
189        bool is_store = (instr->intrinsic == nir_intrinsic_store_ssbo ||
190                         instr->intrinsic == nir_intrinsic_store_scratch ||
191                         instr->intrinsic == nir_intrinsic_store_shared);
192        bool has_index = !is_shared_or_scratch;
193
194        int offset_src;
195        int tmu_writes = 1; /* address */
196        if (instr->intrinsic == nir_intrinsic_load_uniform) {
197                offset_src = 0;
198        } else if (instr->intrinsic == nir_intrinsic_load_ssbo ||
199                   instr->intrinsic == nir_intrinsic_load_ubo ||
200                   instr->intrinsic == nir_intrinsic_load_scratch ||
201                   instr->intrinsic == nir_intrinsic_load_shared) {
202                offset_src = 0 + has_index;
203        } else if (is_store) {
204                offset_src = 1 + has_index;
205                for (int i = 0; i < instr->num_components; i++) {
206                        vir_MOV_dest(c,
207                                     vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUD),
208                                     ntq_get_src(c, instr->src[0], i));
209                        tmu_writes++;
210                }
211        } else {
212                offset_src = 0 + has_index;
213                vir_MOV_dest(c,
214                             vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUD),
215                             ntq_get_src(c, instr->src[1 + has_index], 0));
216                tmu_writes++;
217                if (tmu_op == GENERAL_TMU_WRITE_OP_ATOMIC_CMPXCHG) {
218                        vir_MOV_dest(c,
219                                     vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUD),
220                                     ntq_get_src(c, instr->src[2 + has_index],
221                                                 0));
222                        tmu_writes++;
223                }
224        }
225
226        bool dynamic_src = !nir_src_is_const(instr->src[offset_src]);
227        uint32_t const_offset = 0;
228        if (!dynamic_src)
229                const_offset = nir_src_as_uint(instr->src[offset_src]);
230
231        /* Make sure we won't exceed the 16-entry TMU fifo if each thread is
232         * storing at the same time.
233         */
234        while (tmu_writes > 16 / c->threads)
235                c->threads /= 2;
236
237        struct qreg offset;
238        if (instr->intrinsic == nir_intrinsic_load_uniform) {
239                const_offset += nir_intrinsic_base(instr);
240                offset = vir_uniform(c, QUNIFORM_UBO_ADDR,
241                                     v3d_unit_data_create(0, const_offset));
242                const_offset = 0;
243        } else if (instr->intrinsic == nir_intrinsic_load_ubo) {
244                uint32_t index = nir_src_as_uint(instr->src[0]) + 1;
245                /* Note that QUNIFORM_UBO_ADDR takes a UBO index shifted up by
246                 * 1 (0 is gallium's constant buffer 0).
247                 */
248                offset = vir_uniform(c, QUNIFORM_UBO_ADDR,
249                                     v3d_unit_data_create(index, const_offset));
250                const_offset = 0;
251        } else if (is_shared_or_scratch) {
252                /* Shared and scratch variables have no buffer index, and all
253                 * start from a common base that we set up at the start of
254                 * dispatch.
255                 */
256                if (instr->intrinsic == nir_intrinsic_load_scratch ||
257                    instr->intrinsic == nir_intrinsic_store_scratch) {
258                        offset = c->spill_base;
259                } else {
260                        offset = c->cs_shared_offset;
261                        const_offset += nir_intrinsic_base(instr);
262                }
263        } else {
264                offset = vir_uniform(c, QUNIFORM_SSBO_OFFSET,
265                                     nir_src_as_uint(instr->src[is_store ?
266                                                                1 : 0]));
267        }
268
269        /* The spec says that for atomics, the TYPE field is ignored, but that
270         * doesn't seem to be the case for CMPXCHG.  Just use the number of
271         * tmud writes we did to decide the type (or choose "32bit" for atomic
272         * reads, which has been fine).
273         */
274        int num_components;
275        if (tmu_op == GENERAL_TMU_WRITE_OP_ATOMIC_CMPXCHG)
276                num_components = 2;
277        else
278                num_components = instr->num_components;
279
280        uint32_t config = (0xffffff00 |
281                           tmu_op |
282                           GENERAL_TMU_LOOKUP_PER_PIXEL);
283        if (num_components == 1) {
284                config |= GENERAL_TMU_LOOKUP_TYPE_32BIT_UI;
285        } else {
286                config |= GENERAL_TMU_LOOKUP_TYPE_VEC2 + num_components - 2;
287        }
288
289        if (vir_in_nonuniform_control_flow(c)) {
290                vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), c->execute),
291                           V3D_QPU_PF_PUSHZ);
292        }
293
294        struct qreg tmua;
295        if (config == ~0)
296                tmua = vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUA);
297        else
298                tmua = vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_TMUAU);
299
300        struct qinst *tmu;
301        if (dynamic_src) {
302                if (const_offset != 0) {
303                        offset = vir_ADD(c, offset,
304                                         vir_uniform_ui(c, const_offset));
305                }
306                tmu = vir_ADD_dest(c, tmua, offset,
307                                   ntq_get_src(c, instr->src[offset_src], 0));
308        } else {
309                if (const_offset != 0) {
310                        tmu = vir_ADD_dest(c, tmua, offset,
311                                           vir_uniform_ui(c, const_offset));
312                } else {
313                        tmu = vir_MOV_dest(c, tmua, offset);
314                }
315        }
316
317        if (config != ~0) {
318                tmu->uniform = vir_get_uniform_index(c, QUNIFORM_CONSTANT,
319                                                     config);
320        }
321
322        if (vir_in_nonuniform_control_flow(c))
323                vir_set_cond(tmu, V3D_QPU_COND_IFA);
324
325        vir_emit_thrsw(c);
326
327        /* Read the result, or wait for the TMU op to complete. */
328        for (int i = 0; i < nir_intrinsic_dest_components(instr); i++)
329                ntq_store_dest(c, &instr->dest, i, vir_MOV(c, vir_LDTMU(c)));
330
331        if (nir_intrinsic_dest_components(instr) == 0)
332                vir_TMUWT(c);
333}
334
335static struct qreg *
336ntq_init_ssa_def(struct v3d_compile *c, nir_ssa_def *def)
337{
338        struct qreg *qregs = ralloc_array(c->def_ht, struct qreg,
339                                          def->num_components);
340        _mesa_hash_table_insert(c->def_ht, def, qregs);
341        return qregs;
342}
343
344/**
345 * This function is responsible for getting VIR results into the associated
346 * storage for a NIR instruction.
347 *
348 * If it's a NIR SSA def, then we just set the associated hash table entry to
349 * the new result.
350 *
351 * If it's a NIR reg, then we need to update the existing qreg assigned to the
352 * NIR destination with the incoming value.  To do that without introducing
353 * new MOVs, we require that the incoming qreg either be a uniform, or be
354 * SSA-defined by the previous VIR instruction in the block and rewritable by
355 * this function.  That lets us sneak ahead and insert the SF flag beforehand
356 * (knowing that the previous instruction doesn't depend on flags) and rewrite
357 * its destination to be the NIR reg's destination
358 */
359void
360ntq_store_dest(struct v3d_compile *c, nir_dest *dest, int chan,
361               struct qreg result)
362{
363        struct qinst *last_inst = NULL;
364        if (!list_empty(&c->cur_block->instructions))
365                last_inst = (struct qinst *)c->cur_block->instructions.prev;
366
367        assert((result.file == QFILE_TEMP &&
368                last_inst && last_inst == c->defs[result.index]));
369
370        if (dest->is_ssa) {
371                assert(chan < dest->ssa.num_components);
372
373                struct qreg *qregs;
374                struct hash_entry *entry =
375                        _mesa_hash_table_search(c->def_ht, &dest->ssa);
376
377                if (entry)
378                        qregs = entry->data;
379                else
380                        qregs = ntq_init_ssa_def(c, &dest->ssa);
381
382                qregs[chan] = result;
383        } else {
384                nir_register *reg = dest->reg.reg;
385                assert(dest->reg.base_offset == 0);
386                assert(reg->num_array_elems == 0);
387                struct hash_entry *entry =
388                        _mesa_hash_table_search(c->def_ht, reg);
389                struct qreg *qregs = entry->data;
390
391                /* Insert a MOV if the source wasn't an SSA def in the
392                 * previous instruction.
393                 */
394                if ((vir_in_nonuniform_control_flow(c) &&
395                     c->defs[last_inst->dst.index]->qpu.sig.ldunif)) {
396                        result = vir_MOV(c, result);
397                        last_inst = c->defs[result.index];
398                }
399
400                /* We know they're both temps, so just rewrite index. */
401                c->defs[last_inst->dst.index] = NULL;
402                last_inst->dst.index = qregs[chan].index;
403
404                /* If we're in control flow, then make this update of the reg
405                 * conditional on the execution mask.
406                 */
407                if (vir_in_nonuniform_control_flow(c)) {
408                        last_inst->dst.index = qregs[chan].index;
409
410                        /* Set the flags to the current exec mask.
411                         */
412                        c->cursor = vir_before_inst(last_inst);
413                        vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), c->execute),
414                                   V3D_QPU_PF_PUSHZ);
415                        c->cursor = vir_after_inst(last_inst);
416
417                        vir_set_cond(last_inst, V3D_QPU_COND_IFA);
418                }
419        }
420}
421
422struct qreg
423ntq_get_src(struct v3d_compile *c, nir_src src, int i)
424{
425        struct hash_entry *entry;
426        if (src.is_ssa) {
427                entry = _mesa_hash_table_search(c->def_ht, src.ssa);
428                assert(i < src.ssa->num_components);
429        } else {
430                nir_register *reg = src.reg.reg;
431                entry = _mesa_hash_table_search(c->def_ht, reg);
432                assert(reg->num_array_elems == 0);
433                assert(src.reg.base_offset == 0);
434                assert(i < reg->num_components);
435        }
436
437        struct qreg *qregs = entry->data;
438        return qregs[i];
439}
440
441static struct qreg
442ntq_get_alu_src(struct v3d_compile *c, nir_alu_instr *instr,
443                unsigned src)
444{
445        assert(util_is_power_of_two_or_zero(instr->dest.write_mask));
446        unsigned chan = ffs(instr->dest.write_mask) - 1;
447        struct qreg r = ntq_get_src(c, instr->src[src].src,
448                                    instr->src[src].swizzle[chan]);
449
450        assert(!instr->src[src].abs);
451        assert(!instr->src[src].negate);
452
453        return r;
454};
455
456static struct qreg
457ntq_minify(struct v3d_compile *c, struct qreg size, struct qreg level)
458{
459        return vir_MAX(c, vir_SHR(c, size, level), vir_uniform_ui(c, 1));
460}
461
462static void
463ntq_emit_txs(struct v3d_compile *c, nir_tex_instr *instr)
464{
465        unsigned unit = instr->texture_index;
466        int lod_index = nir_tex_instr_src_index(instr, nir_tex_src_lod);
467        int dest_size = nir_tex_instr_dest_size(instr);
468
469        struct qreg lod = c->undef;
470        if (lod_index != -1)
471                lod = ntq_get_src(c, instr->src[lod_index].src, 0);
472
473        for (int i = 0; i < dest_size; i++) {
474                assert(i < 3);
475                enum quniform_contents contents;
476
477                if (instr->is_array && i == dest_size - 1)
478                        contents = QUNIFORM_TEXTURE_ARRAY_SIZE;
479                else
480                        contents = QUNIFORM_TEXTURE_WIDTH + i;
481
482                struct qreg size = vir_uniform(c, contents, unit);
483
484                switch (instr->sampler_dim) {
485                case GLSL_SAMPLER_DIM_1D:
486                case GLSL_SAMPLER_DIM_2D:
487                case GLSL_SAMPLER_DIM_MS:
488                case GLSL_SAMPLER_DIM_3D:
489                case GLSL_SAMPLER_DIM_CUBE:
490                        /* Don't minify the array size. */
491                        if (!(instr->is_array && i == dest_size - 1)) {
492                                size = ntq_minify(c, size, lod);
493                        }
494                        break;
495
496                case GLSL_SAMPLER_DIM_RECT:
497                        /* There's no LOD field for rects */
498                        break;
499
500                default:
501                        unreachable("Bad sampler type");
502                }
503
504                ntq_store_dest(c, &instr->dest, i, size);
505        }
506}
507
508static void
509ntq_emit_tex(struct v3d_compile *c, nir_tex_instr *instr)
510{
511        unsigned unit = instr->texture_index;
512
513        /* Since each texture sampling op requires uploading uniforms to
514         * reference the texture, there's no HW support for texture size and
515         * you just upload uniforms containing the size.
516         */
517        switch (instr->op) {
518        case nir_texop_query_levels:
519                ntq_store_dest(c, &instr->dest, 0,
520                               vir_uniform(c, QUNIFORM_TEXTURE_LEVELS, unit));
521                return;
522        case nir_texop_txs:
523                ntq_emit_txs(c, instr);
524                return;
525        default:
526                break;
527        }
528
529        if (c->devinfo->ver >= 40)
530                v3d40_vir_emit_tex(c, instr);
531        else
532                v3d33_vir_emit_tex(c, instr);
533}
534
535static struct qreg
536ntq_fsincos(struct v3d_compile *c, struct qreg src, bool is_cos)
537{
538        struct qreg input = vir_FMUL(c, src, vir_uniform_f(c, 1.0f / M_PI));
539        if (is_cos)
540                input = vir_FADD(c, input, vir_uniform_f(c, 0.5));
541
542        struct qreg periods = vir_FROUND(c, input);
543        struct qreg sin_output = vir_SIN(c, vir_FSUB(c, input, periods));
544        return vir_XOR(c, sin_output, vir_SHL(c,
545                                              vir_FTOIN(c, periods),
546                                              vir_uniform_ui(c, -1)));
547}
548
549static struct qreg
550ntq_fsign(struct v3d_compile *c, struct qreg src)
551{
552        struct qreg t = vir_get_temp(c);
553
554        vir_MOV_dest(c, t, vir_uniform_f(c, 0.0));
555        vir_set_pf(vir_FMOV_dest(c, vir_nop_reg(), src), V3D_QPU_PF_PUSHZ);
556        vir_MOV_cond(c, V3D_QPU_COND_IFNA, t, vir_uniform_f(c, 1.0));
557        vir_set_pf(vir_FMOV_dest(c, vir_nop_reg(), src), V3D_QPU_PF_PUSHN);
558        vir_MOV_cond(c, V3D_QPU_COND_IFA, t, vir_uniform_f(c, -1.0));
559        return vir_MOV(c, t);
560}
561
562static void
563emit_fragcoord_input(struct v3d_compile *c, int attr)
564{
565        c->inputs[attr * 4 + 0] = vir_FXCD(c);
566        c->inputs[attr * 4 + 1] = vir_FYCD(c);
567        c->inputs[attr * 4 + 2] = c->payload_z;
568        c->inputs[attr * 4 + 3] = vir_RECIP(c, c->payload_w);
569}
570
571static struct qreg
572emit_fragment_varying(struct v3d_compile *c, nir_variable *var,
573                      uint8_t swizzle, int array_index)
574{
575        struct qreg r3 = vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R3);
576        struct qreg r5 = vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R5);
577
578        struct qreg vary;
579        if (c->devinfo->ver >= 41) {
580                struct qinst *ldvary = vir_add_inst(V3D_QPU_A_NOP, c->undef,
581                                                    c->undef, c->undef);
582                ldvary->qpu.sig.ldvary = true;
583                vary = vir_emit_def(c, ldvary);
584        } else {
585                vir_NOP(c)->qpu.sig.ldvary = true;
586                vary = r3;
587        }
588
589        /* For gl_PointCoord input or distance along a line, we'll be called
590         * with no nir_variable, and we don't count toward VPM size so we
591         * don't track an input slot.
592         */
593        if (!var) {
594                return vir_FADD(c, vir_FMUL(c, vary, c->payload_w), r5);
595        }
596
597        int i = c->num_inputs++;
598        c->input_slots[i] =
599                v3d_slot_from_slot_and_component(var->data.location +
600                                                 array_index, swizzle);
601
602        switch (var->data.interpolation) {
603        case INTERP_MODE_NONE:
604                /* If a gl_FrontColor or gl_BackColor input has no interp
605                 * qualifier, then if we're using glShadeModel(GL_FLAT) it
606                 * needs to be flat shaded.
607                 */
608                switch (var->data.location + array_index) {
609                case VARYING_SLOT_COL0:
610                case VARYING_SLOT_COL1:
611                case VARYING_SLOT_BFC0:
612                case VARYING_SLOT_BFC1:
613                        if (c->fs_key->shade_model_flat) {
614                                BITSET_SET(c->flat_shade_flags, i);
615                                vir_MOV_dest(c, c->undef, vary);
616                                return vir_MOV(c, r5);
617                        } else {
618                                return vir_FADD(c, vir_FMUL(c, vary,
619                                                            c->payload_w), r5);
620                        }
621                default:
622                        break;
623                }
624                /* FALLTHROUGH */
625        case INTERP_MODE_SMOOTH:
626                if (var->data.centroid) {
627                        BITSET_SET(c->centroid_flags, i);
628                        return vir_FADD(c, vir_FMUL(c, vary,
629                                                    c->payload_w_centroid), r5);
630                } else {
631                        return vir_FADD(c, vir_FMUL(c, vary, c->payload_w), r5);
632                }
633        case INTERP_MODE_NOPERSPECTIVE:
634                BITSET_SET(c->noperspective_flags, i);
635                return vir_FADD(c, vir_MOV(c, vary), r5);
636        case INTERP_MODE_FLAT:
637                BITSET_SET(c->flat_shade_flags, i);
638                vir_MOV_dest(c, c->undef, vary);
639                return vir_MOV(c, r5);
640        default:
641                unreachable("Bad interp mode");
642        }
643}
644
645static void
646emit_fragment_input(struct v3d_compile *c, int attr, nir_variable *var,
647                    int array_index)
648{
649        for (int i = 0; i < glsl_get_vector_elements(var->type); i++) {
650                int chan = var->data.location_frac + i;
651                c->inputs[attr * 4 + chan] =
652                        emit_fragment_varying(c, var, chan, array_index);
653        }
654}
655
656static void
657add_output(struct v3d_compile *c,
658           uint32_t decl_offset,
659           uint8_t slot,
660           uint8_t swizzle)
661{
662        uint32_t old_array_size = c->outputs_array_size;
663        resize_qreg_array(c, &c->outputs, &c->outputs_array_size,
664                          decl_offset + 1);
665
666        if (old_array_size != c->outputs_array_size) {
667                c->output_slots = reralloc(c,
668                                           c->output_slots,
669                                           struct v3d_varying_slot,
670                                           c->outputs_array_size);
671        }
672
673        c->output_slots[decl_offset] =
674                v3d_slot_from_slot_and_component(slot, swizzle);
675}
676
677/**
678 * If compare_instr is a valid comparison instruction, emits the
679 * compare_instr's comparison and returns the sel_instr's return value based
680 * on the compare_instr's result.
681 */
682static bool
683ntq_emit_comparison(struct v3d_compile *c,
684                    nir_alu_instr *compare_instr,
685                    enum v3d_qpu_cond *out_cond)
686{
687        struct qreg src0 = ntq_get_alu_src(c, compare_instr, 0);
688        struct qreg src1;
689        if (nir_op_infos[compare_instr->op].num_inputs > 1)
690                src1 = ntq_get_alu_src(c, compare_instr, 1);
691        bool cond_invert = false;
692        struct qreg nop = vir_nop_reg();
693
694        switch (compare_instr->op) {
695        case nir_op_feq32:
696        case nir_op_seq:
697                vir_set_pf(vir_FCMP_dest(c, nop, src0, src1), V3D_QPU_PF_PUSHZ);
698                break;
699        case nir_op_ieq32:
700                vir_set_pf(vir_XOR_dest(c, nop, src0, src1), V3D_QPU_PF_PUSHZ);
701                break;
702
703        case nir_op_fne32:
704        case nir_op_sne:
705                vir_set_pf(vir_FCMP_dest(c, nop, src0, src1), V3D_QPU_PF_PUSHZ);
706                cond_invert = true;
707                break;
708        case nir_op_ine32:
709                vir_set_pf(vir_XOR_dest(c, nop, src0, src1), V3D_QPU_PF_PUSHZ);
710                cond_invert = true;
711                break;
712
713        case nir_op_fge32:
714        case nir_op_sge:
715                vir_set_pf(vir_FCMP_dest(c, nop, src1, src0), V3D_QPU_PF_PUSHC);
716                break;
717        case nir_op_ige32:
718                vir_set_pf(vir_MIN_dest(c, nop, src1, src0), V3D_QPU_PF_PUSHC);
719                cond_invert = true;
720                break;
721        case nir_op_uge32:
722                vir_set_pf(vir_SUB_dest(c, nop, src0, src1), V3D_QPU_PF_PUSHC);
723                cond_invert = true;
724                break;
725
726        case nir_op_slt:
727        case nir_op_flt32:
728                vir_set_pf(vir_FCMP_dest(c, nop, src0, src1), V3D_QPU_PF_PUSHN);
729                break;
730        case nir_op_ilt32:
731                vir_set_pf(vir_MIN_dest(c, nop, src1, src0), V3D_QPU_PF_PUSHC);
732                break;
733        case nir_op_ult32:
734                vir_set_pf(vir_SUB_dest(c, nop, src0, src1), V3D_QPU_PF_PUSHC);
735                break;
736
737        case nir_op_i2b32:
738                vir_set_pf(vir_MOV_dest(c, nop, src0), V3D_QPU_PF_PUSHZ);
739                cond_invert = true;
740                break;
741
742        case nir_op_f2b32:
743                vir_set_pf(vir_FMOV_dest(c, nop, src0), V3D_QPU_PF_PUSHZ);
744                cond_invert = true;
745                break;
746
747        default:
748                return false;
749        }
750
751        *out_cond = cond_invert ? V3D_QPU_COND_IFNA : V3D_QPU_COND_IFA;
752
753        return true;
754}
755
756/* Finds an ALU instruction that generates our src value that could
757 * (potentially) be greedily emitted in the consuming instruction.
758 */
759static struct nir_alu_instr *
760ntq_get_alu_parent(nir_src src)
761{
762        if (!src.is_ssa || src.ssa->parent_instr->type != nir_instr_type_alu)
763                return NULL;
764        nir_alu_instr *instr = nir_instr_as_alu(src.ssa->parent_instr);
765        if (!instr)
766                return NULL;
767
768        /* If the ALU instr's srcs are non-SSA, then we would have to avoid
769         * moving emission of the ALU instr down past another write of the
770         * src.
771         */
772        for (int i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
773                if (!instr->src[i].src.is_ssa)
774                        return NULL;
775        }
776
777        return instr;
778}
779
780/* Turns a NIR bool into a condition code to predicate on. */
781static enum v3d_qpu_cond
782ntq_emit_bool_to_cond(struct v3d_compile *c, nir_src src)
783{
784        nir_alu_instr *compare = ntq_get_alu_parent(src);
785        if (!compare)
786                goto out;
787
788        enum v3d_qpu_cond cond;
789        if (ntq_emit_comparison(c, compare, &cond))
790                return cond;
791
792out:
793        vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), ntq_get_src(c, src, 0)),
794                   V3D_QPU_PF_PUSHZ);
795        return V3D_QPU_COND_IFNA;
796}
797
798static void
799ntq_emit_alu(struct v3d_compile *c, nir_alu_instr *instr)
800{
801        /* This should always be lowered to ALU operations for V3D. */
802        assert(!instr->dest.saturate);
803
804        /* Vectors are special in that they have non-scalarized writemasks,
805         * and just take the first swizzle channel for each argument in order
806         * into each writemask channel.
807         */
808        if (instr->op == nir_op_vec2 ||
809            instr->op == nir_op_vec3 ||
810            instr->op == nir_op_vec4) {
811                struct qreg srcs[4];
812                for (int i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
813                        srcs[i] = ntq_get_src(c, instr->src[i].src,
814                                              instr->src[i].swizzle[0]);
815                for (int i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
816                        ntq_store_dest(c, &instr->dest.dest, i,
817                                       vir_MOV(c, srcs[i]));
818                return;
819        }
820
821        /* General case: We can just grab the one used channel per src. */
822        struct qreg src[nir_op_infos[instr->op].num_inputs];
823        for (int i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
824                src[i] = ntq_get_alu_src(c, instr, i);
825        }
826
827        struct qreg result;
828
829        switch (instr->op) {
830        case nir_op_fmov:
831        case nir_op_imov:
832                result = vir_MOV(c, src[0]);
833                break;
834
835        case nir_op_fneg:
836                result = vir_XOR(c, src[0], vir_uniform_ui(c, 1 << 31));
837                break;
838        case nir_op_ineg:
839                result = vir_NEG(c, src[0]);
840                break;
841
842        case nir_op_fmul:
843                result = vir_FMUL(c, src[0], src[1]);
844                break;
845        case nir_op_fadd:
846                result = vir_FADD(c, src[0], src[1]);
847                break;
848        case nir_op_fsub:
849                result = vir_FSUB(c, src[0], src[1]);
850                break;
851        case nir_op_fmin:
852                result = vir_FMIN(c, src[0], src[1]);
853                break;
854        case nir_op_fmax:
855                result = vir_FMAX(c, src[0], src[1]);
856                break;
857
858        case nir_op_f2i32: {
859                nir_alu_instr *src0_alu = ntq_get_alu_parent(instr->src[0].src);
860                if (src0_alu && src0_alu->op == nir_op_fround_even) {
861                        result = vir_FTOIN(c, ntq_get_alu_src(c, src0_alu, 0));
862                } else {
863                        result = vir_FTOIZ(c, src[0]);
864                }
865                break;
866        }
867
868        case nir_op_f2u32:
869                result = vir_FTOUZ(c, src[0]);
870                break;
871        case nir_op_i2f32:
872                result = vir_ITOF(c, src[0]);
873                break;
874        case nir_op_u2f32:
875                result = vir_UTOF(c, src[0]);
876                break;
877        case nir_op_b2f32:
878                result = vir_AND(c, src[0], vir_uniform_f(c, 1.0));
879                break;
880        case nir_op_b2i32:
881                result = vir_AND(c, src[0], vir_uniform_ui(c, 1));
882                break;
883
884        case nir_op_iadd:
885                result = vir_ADD(c, src[0], src[1]);
886                break;
887        case nir_op_ushr:
888                result = vir_SHR(c, src[0], src[1]);
889                break;
890        case nir_op_isub:
891                result = vir_SUB(c, src[0], src[1]);
892                break;
893        case nir_op_ishr:
894                result = vir_ASR(c, src[0], src[1]);
895                break;
896        case nir_op_ishl:
897                result = vir_SHL(c, src[0], src[1]);
898                break;
899        case nir_op_imin:
900                result = vir_MIN(c, src[0], src[1]);
901                break;
902        case nir_op_umin:
903                result = vir_UMIN(c, src[0], src[1]);
904                break;
905        case nir_op_imax:
906                result = vir_MAX(c, src[0], src[1]);
907                break;
908        case nir_op_umax:
909                result = vir_UMAX(c, src[0], src[1]);
910                break;
911        case nir_op_iand:
912                result = vir_AND(c, src[0], src[1]);
913                break;
914        case nir_op_ior:
915                result = vir_OR(c, src[0], src[1]);
916                break;
917        case nir_op_ixor:
918                result = vir_XOR(c, src[0], src[1]);
919                break;
920        case nir_op_inot:
921                result = vir_NOT(c, src[0]);
922                break;
923
924        case nir_op_ufind_msb:
925                result = vir_SUB(c, vir_uniform_ui(c, 31), vir_CLZ(c, src[0]));
926                break;
927
928        case nir_op_imul:
929                result = vir_UMUL(c, src[0], src[1]);
930                break;
931
932        case nir_op_seq:
933        case nir_op_sne:
934        case nir_op_sge:
935        case nir_op_slt: {
936                enum v3d_qpu_cond cond;
937                MAYBE_UNUSED bool ok = ntq_emit_comparison(c, instr, &cond);
938                assert(ok);
939                result = vir_MOV(c, vir_SEL(c, cond,
940                                            vir_uniform_f(c, 1.0),
941                                            vir_uniform_f(c, 0.0)));
942                break;
943        }
944
945        case nir_op_i2b32:
946        case nir_op_f2b32:
947        case nir_op_feq32:
948        case nir_op_fne32:
949        case nir_op_fge32:
950        case nir_op_flt32:
951        case nir_op_ieq32:
952        case nir_op_ine32:
953        case nir_op_ige32:
954        case nir_op_uge32:
955        case nir_op_ilt32:
956        case nir_op_ult32: {
957                enum v3d_qpu_cond cond;
958                MAYBE_UNUSED bool ok = ntq_emit_comparison(c, instr, &cond);
959                assert(ok);
960                result = vir_MOV(c, vir_SEL(c, cond,
961                                            vir_uniform_ui(c, ~0),
962                                            vir_uniform_ui(c, 0)));
963                break;
964        }
965
966        case nir_op_b32csel:
967                result = vir_MOV(c,
968                                 vir_SEL(c,
969                                         ntq_emit_bool_to_cond(c, instr->src[0].src),
970                                         src[1], src[2]));
971                break;
972
973        case nir_op_fcsel:
974                vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), src[0]),
975                           V3D_QPU_PF_PUSHZ);
976                result = vir_MOV(c, vir_SEL(c, V3D_QPU_COND_IFNA,
977                                            src[1], src[2]));
978                break;
979
980        case nir_op_frcp:
981                result = vir_RECIP(c, src[0]);
982                break;
983        case nir_op_frsq:
984                result = vir_RSQRT(c, src[0]);
985                break;
986        case nir_op_fexp2:
987                result = vir_EXP(c, src[0]);
988                break;
989        case nir_op_flog2:
990                result = vir_LOG(c, src[0]);
991                break;
992
993        case nir_op_fceil:
994                result = vir_FCEIL(c, src[0]);
995                break;
996        case nir_op_ffloor:
997                result = vir_FFLOOR(c, src[0]);
998                break;
999        case nir_op_fround_even:
1000                result = vir_FROUND(c, src[0]);
1001                break;
1002        case nir_op_ftrunc:
1003                result = vir_FTRUNC(c, src[0]);
1004                break;
1005
1006        case nir_op_fsin:
1007                result = ntq_fsincos(c, src[0], false);
1008                break;
1009        case nir_op_fcos:
1010                result = ntq_fsincos(c, src[0], true);
1011                break;
1012
1013        case nir_op_fsign:
1014                result = ntq_fsign(c, src[0]);
1015                break;
1016
1017        case nir_op_fabs: {
1018                result = vir_FMOV(c, src[0]);
1019                vir_set_unpack(c->defs[result.index], 0, V3D_QPU_UNPACK_ABS);
1020                break;
1021        }
1022
1023        case nir_op_iabs:
1024                result = vir_MAX(c, src[0], vir_NEG(c, src[0]));
1025                break;
1026
1027        case nir_op_fddx:
1028        case nir_op_fddx_coarse:
1029        case nir_op_fddx_fine:
1030                result = vir_FDX(c, src[0]);
1031                break;
1032
1033        case nir_op_fddy:
1034        case nir_op_fddy_coarse:
1035        case nir_op_fddy_fine:
1036                result = vir_FDY(c, src[0]);
1037                break;
1038
1039        case nir_op_uadd_carry:
1040                vir_set_pf(vir_ADD_dest(c, vir_nop_reg(), src[0], src[1]),
1041                           V3D_QPU_PF_PUSHC);
1042                result = vir_MOV(c, vir_SEL(c, V3D_QPU_COND_IFA,
1043                                            vir_uniform_ui(c, ~0),
1044                                            vir_uniform_ui(c, 0)));
1045                break;
1046
1047        case nir_op_pack_half_2x16_split:
1048                result = vir_VFPACK(c, src[0], src[1]);
1049                break;
1050
1051        case nir_op_unpack_half_2x16_split_x:
1052                result = vir_FMOV(c, src[0]);
1053                vir_set_unpack(c->defs[result.index], 0, V3D_QPU_UNPACK_L);
1054                break;
1055
1056        case nir_op_unpack_half_2x16_split_y:
1057                result = vir_FMOV(c, src[0]);
1058                vir_set_unpack(c->defs[result.index], 0, V3D_QPU_UNPACK_H);
1059                break;
1060
1061        default:
1062                fprintf(stderr, "unknown NIR ALU inst: ");
1063                nir_print_instr(&instr->instr, stderr);
1064                fprintf(stderr, "\n");
1065                abort();
1066        }
1067
1068        /* We have a scalar result, so the instruction should only have a
1069         * single channel written to.
1070         */
1071        assert(util_is_power_of_two_or_zero(instr->dest.write_mask));
1072        ntq_store_dest(c, &instr->dest.dest,
1073                       ffs(instr->dest.write_mask) - 1, result);
1074}
1075
1076/* Each TLB read/write setup (a render target or depth buffer) takes an 8-bit
1077 * specifier.  They come from a register that's preloaded with 0xffffffff
1078 * (0xff gets you normal vec4 f16 RT0 writes), and when one is neaded the low
1079 * 8 bits are shifted off the bottom and 0xff shifted in from the top.
1080 */
1081#define TLB_TYPE_F16_COLOR         (3 << 6)
1082#define TLB_TYPE_I32_COLOR         (1 << 6)
1083#define TLB_TYPE_F32_COLOR         (0 << 6)
1084#define TLB_RENDER_TARGET_SHIFT    3 /* Reversed!  7 = RT 0, 0 = RT 7. */
1085#define TLB_SAMPLE_MODE_PER_SAMPLE (0 << 2)
1086#define TLB_SAMPLE_MODE_PER_PIXEL  (1 << 2)
1087#define TLB_F16_SWAP_HI_LO         (1 << 1)
1088#define TLB_VEC_SIZE_4_F16         (1 << 0)
1089#define TLB_VEC_SIZE_2_F16         (0 << 0)
1090#define TLB_VEC_SIZE_MINUS_1_SHIFT 0
1091
1092/* Triggers Z/Stencil testing, used when the shader state's "FS modifies Z"
1093 * flag is set.
1094 */
1095#define TLB_TYPE_DEPTH             ((2 << 6) | (0 << 4))
1096#define TLB_DEPTH_TYPE_INVARIANT   (0 << 2) /* Unmodified sideband input used */
1097#define TLB_DEPTH_TYPE_PER_PIXEL   (1 << 2) /* QPU result used */
1098#define TLB_V42_DEPTH_TYPE_INVARIANT   (0 << 3) /* Unmodified sideband input used */
1099#define TLB_V42_DEPTH_TYPE_PER_PIXEL   (1 << 3) /* QPU result used */
1100
1101/* Stencil is a single 32-bit write. */
1102#define TLB_TYPE_STENCIL_ALPHA     ((2 << 6) | (1 << 4))
1103
1104static void
1105emit_frag_end(struct v3d_compile *c)
1106{
1107        /* XXX
1108        if (c->output_sample_mask_index != -1) {
1109                vir_MS_MASK(c, c->outputs[c->output_sample_mask_index]);
1110        }
1111        */
1112
1113        bool has_any_tlb_color_write = false;
1114        for (int rt = 0; rt < V3D_MAX_DRAW_BUFFERS; rt++) {
1115                if (c->fs_key->cbufs & (1 << rt) && c->output_color_var[rt])
1116                        has_any_tlb_color_write = true;
1117        }
1118
1119        if (c->fs_key->sample_alpha_to_coverage && c->output_color_var[0]) {
1120                struct nir_variable *var = c->output_color_var[0];
1121                struct qreg *color = &c->outputs[var->data.driver_location * 4];
1122
1123                vir_SETMSF_dest(c, vir_nop_reg(),
1124                                vir_AND(c,
1125                                        vir_MSF(c),
1126                                        vir_FTOC(c, color[3])));
1127        }
1128
1129        struct qreg tlb_reg = vir_magic_reg(V3D_QPU_WADDR_TLB);
1130        struct qreg tlbu_reg = vir_magic_reg(V3D_QPU_WADDR_TLBU);
1131        if (c->output_position_index != -1) {
1132                struct qinst *inst = vir_MOV_dest(c, tlbu_reg,
1133                                                  c->outputs[c->output_position_index]);
1134                uint8_t tlb_specifier = TLB_TYPE_DEPTH;
1135
1136                if (c->devinfo->ver >= 42) {
1137                        tlb_specifier |= (TLB_V42_DEPTH_TYPE_PER_PIXEL |
1138                                          TLB_SAMPLE_MODE_PER_PIXEL);
1139                } else
1140                        tlb_specifier |= TLB_DEPTH_TYPE_PER_PIXEL;
1141
1142                inst->uniform = vir_get_uniform_index(c, QUNIFORM_CONSTANT,
1143                                                      tlb_specifier |
1144                                                      0xffffff00);
1145                c->writes_z = true;
1146        } else if (c->s->info.fs.uses_discard ||
1147                   !c->s->info.fs.early_fragment_tests ||
1148                   c->fs_key->sample_alpha_to_coverage ||
1149                   !has_any_tlb_color_write) {
1150                /* Emit passthrough Z if it needed to be delayed until shader
1151                 * end due to potential discards.
1152                 *
1153                 * Since (single-threaded) fragment shaders always need a TLB
1154                 * write, emit passthrouh Z if we didn't have any color
1155                 * buffers and flag us as potentially discarding, so that we
1156                 * can use Z as the TLB write.
1157                 */
1158                c->s->info.fs.uses_discard = true;
1159
1160                struct qinst *inst = vir_MOV_dest(c, tlbu_reg,
1161                                                  vir_nop_reg());
1162                uint8_t tlb_specifier = TLB_TYPE_DEPTH;
1163
1164                if (c->devinfo->ver >= 42) {
1165                        /* The spec says the PER_PIXEL flag is ignored for
1166                         * invariant writes, but the simulator demands it.
1167                         */
1168                        tlb_specifier |= (TLB_V42_DEPTH_TYPE_INVARIANT |
1169                                          TLB_SAMPLE_MODE_PER_PIXEL);
1170                } else {
1171                        tlb_specifier |= TLB_DEPTH_TYPE_INVARIANT;
1172                }
1173
1174                inst->uniform = vir_get_uniform_index(c,
1175                                                      QUNIFORM_CONSTANT,
1176                                                      tlb_specifier |
1177                                                      0xffffff00);
1178                c->writes_z = true;
1179        }
1180
1181        /* XXX: Performance improvement: Merge Z write and color writes TLB
1182         * uniform setup
1183         */
1184
1185        for (int rt = 0; rt < V3D_MAX_DRAW_BUFFERS; rt++) {
1186                if (!(c->fs_key->cbufs & (1 << rt)) || !c->output_color_var[rt])
1187                        continue;
1188
1189                nir_variable *var = c->output_color_var[rt];
1190                struct qreg *color = &c->outputs[var->data.driver_location * 4];
1191                int num_components = glsl_get_vector_elements(var->type);
1192                uint32_t conf = 0xffffff00;
1193                struct qinst *inst;
1194
1195                conf |= TLB_SAMPLE_MODE_PER_PIXEL;
1196                conf |= (7 - rt) << TLB_RENDER_TARGET_SHIFT;
1197
1198                if (c->fs_key->swap_color_rb & (1 << rt))
1199                        num_components = MAX2(num_components, 3);
1200
1201                assert(num_components != 0);
1202                switch (glsl_get_base_type(var->type)) {
1203                case GLSL_TYPE_UINT:
1204                case GLSL_TYPE_INT:
1205                        /* The F32 vs I32 distinction was dropped in 4.2. */
1206                        if (c->devinfo->ver < 42)
1207                                conf |= TLB_TYPE_I32_COLOR;
1208                        else
1209                                conf |= TLB_TYPE_F32_COLOR;
1210                        conf |= ((num_components - 1) <<
1211                                 TLB_VEC_SIZE_MINUS_1_SHIFT);
1212
1213                        inst = vir_MOV_dest(c, tlbu_reg, color[0]);
1214                        inst->uniform = vir_get_uniform_index(c,
1215                                                              QUNIFORM_CONSTANT,
1216                                                              conf);
1217
1218                        for (int i = 1; i < num_components; i++) {
1219                                inst = vir_MOV_dest(c, tlb_reg, color[i]);
1220                        }
1221                        break;
1222
1223                default: {
1224                        struct qreg r = color[0];
1225                        struct qreg g = color[1];
1226                        struct qreg b = color[2];
1227                        struct qreg a = color[3];
1228
1229                        if (c->fs_key->f32_color_rb & (1 << rt)) {
1230                                conf |= TLB_TYPE_F32_COLOR;
1231                                conf |= ((num_components - 1) <<
1232                                         TLB_VEC_SIZE_MINUS_1_SHIFT);
1233                        } else {
1234                                conf |= TLB_TYPE_F16_COLOR;
1235                                conf |= TLB_F16_SWAP_HI_LO;
1236                                if (num_components >= 3)
1237                                        conf |= TLB_VEC_SIZE_4_F16;
1238                                else
1239                                        conf |= TLB_VEC_SIZE_2_F16;
1240                        }
1241
1242                        if (c->fs_key->swap_color_rb & (1 << rt))  {
1243                                r = color[2];
1244                                b = color[0];
1245                        }
1246
1247                        if (c->fs_key->sample_alpha_to_one)
1248                                a = vir_uniform_f(c, 1.0);
1249
1250                        if (c->fs_key->f32_color_rb & (1 << rt)) {
1251                                inst = vir_MOV_dest(c, tlbu_reg, r);
1252                                inst->uniform = vir_get_uniform_index(c,
1253                                                                      QUNIFORM_CONSTANT,
1254                                                                      conf);
1255
1256                                if (num_components >= 2)
1257                                        vir_MOV_dest(c, tlb_reg, g);
1258                                if (num_components >= 3)
1259                                        vir_MOV_dest(c, tlb_reg, b);
1260                                if (num_components >= 4)
1261                                        vir_MOV_dest(c, tlb_reg, a);
1262                        } else {
1263                                inst = vir_VFPACK_dest(c, tlb_reg, r, g);
1264                                if (conf != ~0) {
1265                                        inst->dst = tlbu_reg;
1266                                        inst->uniform = vir_get_uniform_index(c,
1267                                                                              QUNIFORM_CONSTANT,
1268                                                                              conf);
1269                                }
1270
1271                                if (num_components >= 3)
1272                                        inst = vir_VFPACK_dest(c, tlb_reg, b, a);
1273                        }
1274                        break;
1275                }
1276                }
1277        }
1278}
1279
1280static void
1281vir_VPM_WRITE(struct v3d_compile *c, struct qreg val, uint32_t vpm_index)
1282{
1283        if (c->devinfo->ver >= 40) {
1284                vir_STVPMV(c, vir_uniform_ui(c, vpm_index), val);
1285        } else {
1286                /* XXX: v3d33_vir_vpm_write_setup(c); */
1287                vir_MOV_dest(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_VPM), val);
1288        }
1289}
1290
1291static void
1292emit_vert_end(struct v3d_compile *c)
1293{
1294        /* GFXH-1684: VPM writes need to be complete by the end of the shader.
1295         */
1296        if (c->devinfo->ver >= 40 && c->devinfo->ver <= 42)
1297                vir_VPMWT(c);
1298}
1299
1300void
1301v3d_optimize_nir(struct nir_shader *s)
1302{
1303        bool progress;
1304
1305        do {
1306                progress = false;
1307
1308                NIR_PASS_V(s, nir_lower_vars_to_ssa);
1309                NIR_PASS(progress, s, nir_lower_alu_to_scalar);
1310                NIR_PASS(progress, s, nir_lower_phis_to_scalar);
1311                NIR_PASS(progress, s, nir_copy_prop);
1312                NIR_PASS(progress, s, nir_opt_remove_phis);
1313                NIR_PASS(progress, s, nir_opt_dce);
1314                NIR_PASS(progress, s, nir_opt_dead_cf);
1315                NIR_PASS(progress, s, nir_opt_cse);
1316                NIR_PASS(progress, s, nir_opt_peephole_select, 8, true, true);
1317                NIR_PASS(progress, s, nir_opt_algebraic);
1318                NIR_PASS(progress, s, nir_opt_constant_folding);
1319                NIR_PASS(progress, s, nir_opt_undef);
1320        } while (progress);
1321
1322        NIR_PASS(progress, s, nir_opt_move_load_ubo);
1323}
1324
1325static int
1326driver_location_compare(const void *in_a, const void *in_b)
1327{
1328        const nir_variable *const *a = in_a;
1329        const nir_variable *const *b = in_b;
1330
1331        return (*a)->data.driver_location - (*b)->data.driver_location;
1332}
1333
1334static struct qreg
1335ntq_emit_vpm_read(struct v3d_compile *c,
1336                  uint32_t *num_components_queued,
1337                  uint32_t *remaining,
1338                  uint32_t vpm_index)
1339{
1340        struct qreg vpm = vir_reg(QFILE_VPM, vpm_index);
1341
1342        if (c->devinfo->ver >= 40 ) {
1343                return vir_LDVPMV_IN(c,
1344                                     vir_uniform_ui(c,
1345                                                    (*num_components_queued)++));
1346        }
1347
1348        if (*num_components_queued != 0) {
1349                (*num_components_queued)--;
1350                return vir_MOV(c, vpm);
1351        }
1352
1353        uint32_t num_components = MIN2(*remaining, 32);
1354
1355        v3d33_vir_vpm_read_setup(c, num_components);
1356
1357        *num_components_queued = num_components - 1;
1358        *remaining -= num_components;
1359
1360        return vir_MOV(c, vpm);
1361}
1362
1363static void
1364ntq_setup_vpm_inputs(struct v3d_compile *c)
1365{
1366        /* Figure out how many components of each vertex attribute the shader
1367         * uses.  Each variable should have been split to individual
1368         * components and unused ones DCEed.  The vertex fetcher will load
1369         * from the start of the attribute to the number of components we
1370         * declare we need in c->vattr_sizes[].
1371         */
1372        nir_foreach_variable(var, &c->s->inputs) {
1373                /* No VS attribute array support. */
1374                assert(MAX2(glsl_get_length(var->type), 1) == 1);
1375
1376                unsigned loc = var->data.driver_location;
1377                int start_component = var->data.location_frac;
1378                int num_components = glsl_get_components(var->type);
1379
1380                c->vattr_sizes[loc] = MAX2(c->vattr_sizes[loc],
1381                                           start_component + num_components);
1382        }
1383
1384        unsigned num_components = 0;
1385        uint32_t vpm_components_queued = 0;
1386        bool uses_iid = c->s->info.system_values_read &
1387                (1ull << SYSTEM_VALUE_INSTANCE_ID);
1388        bool uses_vid = c->s->info.system_values_read &
1389                (1ull << SYSTEM_VALUE_VERTEX_ID);
1390        num_components += uses_iid;
1391        num_components += uses_vid;
1392
1393        for (int i = 0; i < ARRAY_SIZE(c->vattr_sizes); i++)
1394                num_components += c->vattr_sizes[i];
1395
1396        if (uses_iid) {
1397                c->iid = ntq_emit_vpm_read(c, &vpm_components_queued,
1398                                           &num_components, ~0);
1399        }
1400
1401        if (uses_vid) {
1402                c->vid = ntq_emit_vpm_read(c, &vpm_components_queued,
1403                                           &num_components, ~0);
1404        }
1405
1406        /* The actual loads will happen directly in nir_intrinsic_load_input
1407         * on newer versions.
1408         */
1409        if (c->devinfo->ver >= 40)
1410                return;
1411
1412        for (int loc = 0; loc < ARRAY_SIZE(c->vattr_sizes); loc++) {
1413                resize_qreg_array(c, &c->inputs, &c->inputs_array_size,
1414                                  (loc + 1) * 4);
1415
1416                for (int i = 0; i < c->vattr_sizes[loc]; i++) {
1417                        c->inputs[loc * 4 + i] =
1418                                ntq_emit_vpm_read(c,
1419                                                  &vpm_components_queued,
1420                                                  &num_components,
1421                                                  loc * 4 + i);
1422
1423                }
1424        }
1425
1426        if (c->devinfo->ver >= 40) {
1427                assert(vpm_components_queued == num_components);
1428        } else {
1429                assert(vpm_components_queued == 0);
1430                assert(num_components == 0);
1431        }
1432}
1433
1434static void
1435ntq_setup_fs_inputs(struct v3d_compile *c)
1436{
1437        unsigned num_entries = 0;
1438        unsigned num_components = 0;
1439        nir_foreach_variable(var, &c->s->inputs) {
1440                num_entries++;
1441                num_components += glsl_get_components(var->type);
1442        }
1443
1444        nir_variable *vars[num_entries];
1445
1446        unsigned i = 0;
1447        nir_foreach_variable(var, &c->s->inputs)
1448                vars[i++] = var;
1449
1450        /* Sort the variables so that we emit the input setup in
1451         * driver_location order.  This is required for VPM reads, whose data
1452         * is fetched into the VPM in driver_location (TGSI register index)
1453         * order.
1454         */
1455        qsort(&vars, num_entries, sizeof(*vars), driver_location_compare);
1456
1457        for (unsigned i = 0; i < num_entries; i++) {
1458                nir_variable *var = vars[i];
1459                unsigned array_len = MAX2(glsl_get_length(var->type), 1);
1460                unsigned loc = var->data.driver_location;
1461
1462                resize_qreg_array(c, &c->inputs, &c->inputs_array_size,
1463                                  (loc + array_len) * 4);
1464
1465                if (var->data.location == VARYING_SLOT_POS) {
1466                        emit_fragcoord_input(c, loc);
1467                } else if (var->data.location == VARYING_SLOT_PNTC ||
1468                           (var->data.location >= VARYING_SLOT_VAR0 &&
1469                            (c->fs_key->point_sprite_mask &
1470                             (1 << (var->data.location -
1471                                    VARYING_SLOT_VAR0))))) {
1472                        c->inputs[loc * 4 + 0] = c->point_x;
1473                        c->inputs[loc * 4 + 1] = c->point_y;
1474                } else {
1475                        for (int j = 0; j < array_len; j++)
1476                                emit_fragment_input(c, loc + j, var, j);
1477                }
1478        }
1479}
1480
1481static void
1482ntq_setup_outputs(struct v3d_compile *c)
1483{
1484        if (c->s->info.stage != MESA_SHADER_FRAGMENT)
1485                return;
1486
1487        nir_foreach_variable(var, &c->s->outputs) {
1488                unsigned array_len = MAX2(glsl_get_length(var->type), 1);
1489                unsigned loc = var->data.driver_location * 4;
1490
1491                assert(array_len == 1);
1492                (void)array_len;
1493
1494                for (int i = 0; i < 4 - var->data.location_frac; i++) {
1495                        add_output(c, loc + var->data.location_frac + i,
1496                                   var->data.location,
1497                                   var->data.location_frac + i);
1498                }
1499
1500                switch (var->data.location) {
1501                case FRAG_RESULT_COLOR:
1502                        c->output_color_var[0] = var;
1503                        c->output_color_var[1] = var;
1504                        c->output_color_var[2] = var;
1505                        c->output_color_var[3] = var;
1506                        break;
1507                case FRAG_RESULT_DATA0:
1508                case FRAG_RESULT_DATA1:
1509                case FRAG_RESULT_DATA2:
1510                case FRAG_RESULT_DATA3:
1511                        c->output_color_var[var->data.location -
1512                                            FRAG_RESULT_DATA0] = var;
1513                        break;
1514                case FRAG_RESULT_DEPTH:
1515                        c->output_position_index = loc;
1516                        break;
1517                case FRAG_RESULT_SAMPLE_MASK:
1518                        c->output_sample_mask_index = loc;
1519                        break;
1520                }
1521        }
1522}
1523
1524/**
1525 * Sets up the mapping from nir_register to struct qreg *.
1526 *
1527 * Each nir_register gets a struct qreg per 32-bit component being stored.
1528 */
1529static void
1530ntq_setup_registers(struct v3d_compile *c, struct exec_list *list)
1531{
1532        foreach_list_typed(nir_register, nir_reg, node, list) {
1533                unsigned array_len = MAX2(nir_reg->num_array_elems, 1);
1534                struct qreg *qregs = ralloc_array(c->def_ht, struct qreg,
1535                                                  array_len *
1536                                                  nir_reg->num_components);
1537
1538                _mesa_hash_table_insert(c->def_ht, nir_reg, qregs);
1539
1540                for (int i = 0; i < array_len * nir_reg->num_components; i++)
1541                        qregs[i] = vir_get_temp(c);
1542        }
1543}
1544
1545static void
1546ntq_emit_load_const(struct v3d_compile *c, nir_load_const_instr *instr)
1547{
1548        /* XXX perf: Experiment with using immediate loads to avoid having
1549         * these end up in the uniform stream.  Watch out for breaking the
1550         * small immediates optimization in the process!
1551         */
1552        struct qreg *qregs = ntq_init_ssa_def(c, &instr->def);
1553        for (int i = 0; i < instr->def.num_components; i++)
1554                qregs[i] = vir_uniform_ui(c, instr->value[i].u32);
1555
1556        _mesa_hash_table_insert(c->def_ht, &instr->def, qregs);
1557}
1558
1559static void
1560ntq_emit_ssa_undef(struct v3d_compile *c, nir_ssa_undef_instr *instr)
1561{
1562        struct qreg *qregs = ntq_init_ssa_def(c, &instr->def);
1563
1564        /* VIR needs there to be *some* value, so pick 0 (same as for
1565         * ntq_setup_registers().
1566         */
1567        for (int i = 0; i < instr->def.num_components; i++)
1568                qregs[i] = vir_uniform_ui(c, 0);
1569}
1570
1571static void
1572ntq_emit_image_size(struct v3d_compile *c, nir_intrinsic_instr *instr)
1573{
1574        assert(instr->intrinsic == nir_intrinsic_image_deref_size);
1575        nir_variable *var = nir_intrinsic_get_var(instr, 0);
1576        unsigned image_index = var->data.driver_location;
1577        const struct glsl_type *sampler_type = glsl_without_array(var->type);
1578        bool is_array = glsl_sampler_type_is_array(sampler_type);
1579
1580        ntq_store_dest(c, &instr->dest, 0,
1581                       vir_uniform(c, QUNIFORM_IMAGE_WIDTH, image_index));
1582        if (instr->num_components > 1) {
1583                ntq_store_dest(c, &instr->dest, 1,
1584                               vir_uniform(c, QUNIFORM_IMAGE_HEIGHT,
1585                                           image_index));
1586        }
1587        if (instr->num_components > 2) {
1588                ntq_store_dest(c, &instr->dest, 2,
1589                               vir_uniform(c,
1590                                           is_array ?
1591                                           QUNIFORM_IMAGE_ARRAY_SIZE :
1592                                           QUNIFORM_IMAGE_DEPTH,
1593                                           image_index));
1594        }
1595}
1596
1597static void
1598ntq_emit_intrinsic(struct v3d_compile *c, nir_intrinsic_instr *instr)
1599{
1600        unsigned offset;
1601
1602        switch (instr->intrinsic) {
1603        case nir_intrinsic_load_uniform:
1604                if (nir_src_is_const(instr->src[0])) {
1605                        int offset = (nir_intrinsic_base(instr) +
1606                                      nir_src_as_uint(instr->src[0]));
1607                        assert(offset % 4 == 0);
1608                        /* We need dwords */
1609                        offset = offset / 4;
1610                        for (int i = 0; i < instr->num_components; i++) {
1611                                ntq_store_dest(c, &instr->dest, i,
1612                                               vir_uniform(c, QUNIFORM_UNIFORM,
1613                                                           offset + i));
1614                        }
1615                } else {
1616                        ntq_emit_tmu_general(c, instr, false);
1617                }
1618                break;
1619
1620        case nir_intrinsic_load_ubo:
1621                ntq_emit_tmu_general(c, instr, false);
1622                break;
1623
1624        case nir_intrinsic_ssbo_atomic_add:
1625        case nir_intrinsic_ssbo_atomic_imin:
1626        case nir_intrinsic_ssbo_atomic_umin:
1627        case nir_intrinsic_ssbo_atomic_imax:
1628        case nir_intrinsic_ssbo_atomic_umax:
1629        case nir_intrinsic_ssbo_atomic_and:
1630        case nir_intrinsic_ssbo_atomic_or:
1631        case nir_intrinsic_ssbo_atomic_xor:
1632        case nir_intrinsic_ssbo_atomic_exchange:
1633        case nir_intrinsic_ssbo_atomic_comp_swap:
1634        case nir_intrinsic_load_ssbo:
1635        case nir_intrinsic_store_ssbo:
1636                ntq_emit_tmu_general(c, instr, false);
1637                break;
1638
1639        case nir_intrinsic_shared_atomic_add:
1640        case nir_intrinsic_shared_atomic_imin:
1641        case nir_intrinsic_shared_atomic_umin:
1642        case nir_intrinsic_shared_atomic_imax:
1643        case nir_intrinsic_shared_atomic_umax:
1644        case nir_intrinsic_shared_atomic_and:
1645        case nir_intrinsic_shared_atomic_or:
1646        case nir_intrinsic_shared_atomic_xor:
1647        case nir_intrinsic_shared_atomic_exchange:
1648        case nir_intrinsic_shared_atomic_comp_swap:
1649        case nir_intrinsic_load_shared:
1650        case nir_intrinsic_store_shared:
1651        case nir_intrinsic_load_scratch:
1652        case nir_intrinsic_store_scratch:
1653                ntq_emit_tmu_general(c, instr, true);
1654                break;
1655
1656        case nir_intrinsic_image_deref_load:
1657        case nir_intrinsic_image_deref_store:
1658        case nir_intrinsic_image_deref_atomic_add:
1659        case nir_intrinsic_image_deref_atomic_min:
1660        case nir_intrinsic_image_deref_atomic_max:
1661        case nir_intrinsic_image_deref_atomic_and:
1662        case nir_intrinsic_image_deref_atomic_or:
1663        case nir_intrinsic_image_deref_atomic_xor:
1664        case nir_intrinsic_image_deref_atomic_exchange:
1665        case nir_intrinsic_image_deref_atomic_comp_swap:
1666                v3d40_vir_emit_image_load_store(c, instr);
1667                break;
1668
1669        case nir_intrinsic_get_buffer_size:
1670                ntq_store_dest(c, &instr->dest, 0,
1671                               vir_uniform(c, QUNIFORM_GET_BUFFER_SIZE,
1672                                           nir_src_as_uint(instr->src[0])));
1673                break;
1674
1675        case nir_intrinsic_load_user_clip_plane:
1676                for (int i = 0; i < instr->num_components; i++) {
1677                        ntq_store_dest(c, &instr->dest, i,
1678                                       vir_uniform(c, QUNIFORM_USER_CLIP_PLANE,
1679                                                   nir_intrinsic_ucp_id(instr) *
1680                                                   4 + i));
1681                }
1682                break;
1683
1684        case nir_intrinsic_load_viewport_x_scale:
1685                ntq_store_dest(c, &instr->dest, 0,
1686                               vir_uniform(c, QUNIFORM_VIEWPORT_X_SCALE, 0));
1687                break;
1688
1689        case nir_intrinsic_load_viewport_y_scale:
1690                ntq_store_dest(c, &instr->dest, 0,
1691                               vir_uniform(c, QUNIFORM_VIEWPORT_Y_SCALE, 0));
1692                break;
1693
1694        case nir_intrinsic_load_viewport_z_scale:
1695                ntq_store_dest(c, &instr->dest, 0,
1696                               vir_uniform(c, QUNIFORM_VIEWPORT_Z_SCALE, 0));
1697                break;
1698
1699        case nir_intrinsic_load_viewport_z_offset:
1700                ntq_store_dest(c, &instr->dest, 0,
1701                               vir_uniform(c, QUNIFORM_VIEWPORT_Z_OFFSET, 0));
1702                break;
1703
1704        case nir_intrinsic_load_alpha_ref_float:
1705                ntq_store_dest(c, &instr->dest, 0,
1706                               vir_uniform(c, QUNIFORM_ALPHA_REF, 0));
1707                break;
1708
1709        case nir_intrinsic_load_sample_mask_in:
1710                ntq_store_dest(c, &instr->dest, 0, vir_MSF(c));
1711                break;
1712
1713        case nir_intrinsic_load_helper_invocation:
1714                vir_set_pf(vir_MSF_dest(c, vir_nop_reg()), V3D_QPU_PF_PUSHZ);
1715                ntq_store_dest(c, &instr->dest, 0,
1716                               vir_MOV(c, vir_SEL(c, V3D_QPU_COND_IFA,
1717                                                  vir_uniform_ui(c, ~0),
1718                                                  vir_uniform_ui(c, 0))));
1719                break;
1720
1721        case nir_intrinsic_load_front_face:
1722                /* The register contains 0 (front) or 1 (back), and we need to
1723                 * turn it into a NIR bool where true means front.
1724                 */
1725                ntq_store_dest(c, &instr->dest, 0,
1726                               vir_ADD(c,
1727                                       vir_uniform_ui(c, -1),
1728                                       vir_REVF(c)));
1729                break;
1730
1731        case nir_intrinsic_load_instance_id:
1732                ntq_store_dest(c, &instr->dest, 0, vir_MOV(c, c->iid));
1733                break;
1734
1735        case nir_intrinsic_load_vertex_id:
1736                ntq_store_dest(c, &instr->dest, 0, vir_MOV(c, c->vid));
1737                break;
1738
1739        case nir_intrinsic_load_input:
1740                /* Use ldvpmv (uniform offset) or ldvpmd (non-uniform offset)
1741                 * and enable PIPE_SHADER_CAP_INDIRECT_INPUT_ADDR.
1742                 */
1743                offset = (nir_intrinsic_base(instr) +
1744                          nir_src_as_uint(instr->src[0]));
1745                if (c->s->info.stage != MESA_SHADER_FRAGMENT &&
1746                    c->devinfo->ver >= 40) {
1747                        /* Emit the LDVPM directly now, rather than at the top
1748                         * of the shader like we did for V3D 3.x (which needs
1749                         * vpmsetup when not just taking the next offset).
1750                         *
1751                         * Note that delaying like this may introduce stalls,
1752                         * as LDVPMV takes a minimum of 1 instruction but may
1753                         * be slower if the VPM unit is busy with another QPU.
1754                         */
1755                        int index = 0;
1756                        if (c->s->info.system_values_read &
1757                            (1ull << SYSTEM_VALUE_INSTANCE_ID)) {
1758                                index++;
1759                        }
1760                        if (c->s->info.system_values_read &
1761                            (1ull << SYSTEM_VALUE_VERTEX_ID)) {
1762                                index++;
1763                        }
1764                        for (int i = 0; i < offset; i++)
1765                                index += c->vattr_sizes[i];
1766                        index += nir_intrinsic_component(instr);
1767                        for (int i = 0; i < instr->num_components; i++) {
1768                                struct qreg vpm_offset =
1769                                        vir_uniform_ui(c, index++);
1770                                ntq_store_dest(c, &instr->dest, i,
1771                                               vir_LDVPMV_IN(c, vpm_offset));
1772                        }
1773                } else {
1774                        for (int i = 0; i < instr->num_components; i++) {
1775                                int comp = nir_intrinsic_component(instr) + i;
1776                                ntq_store_dest(c, &instr->dest, i,
1777                                               vir_MOV(c, c->inputs[offset * 4 +
1778                                                                    comp]));
1779                        }
1780                }
1781                break;
1782
1783        case nir_intrinsic_store_output:
1784                /* XXX perf: Use stvpmv with uniform non-constant offsets and
1785                 * stvpmd with non-uniform offsets and enable
1786                 * PIPE_SHADER_CAP_INDIRECT_OUTPUT_ADDR.
1787                 */
1788                if (c->s->info.stage == MESA_SHADER_FRAGMENT) {
1789                        offset = ((nir_intrinsic_base(instr) +
1790                                   nir_src_as_uint(instr->src[1])) * 4 +
1791                                  nir_intrinsic_component(instr));
1792                        for (int i = 0; i < instr->num_components; i++) {
1793                                c->outputs[offset + i] =
1794                                        vir_MOV(c,
1795                                                ntq_get_src(c,
1796                                                            instr->src[0], i));
1797                        }
1798                } else {
1799                        assert(instr->num_components == 1);
1800
1801                        vir_VPM_WRITE(c,
1802                                      ntq_get_src(c, instr->src[0], 0),
1803                                      nir_intrinsic_base(instr));
1804                }
1805                break;
1806
1807        case nir_intrinsic_image_deref_size:
1808                ntq_emit_image_size(c, instr);
1809                break;
1810
1811        case nir_intrinsic_discard:
1812                if (vir_in_nonuniform_control_flow(c)) {
1813                        vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), c->execute),
1814                                   V3D_QPU_PF_PUSHZ);
1815                        vir_set_cond(vir_SETMSF_dest(c, vir_nop_reg(),
1816                                                     vir_uniform_ui(c, 0)),
1817                                V3D_QPU_COND_IFA);
1818                } else {
1819                        vir_SETMSF_dest(c, vir_nop_reg(),
1820                                        vir_uniform_ui(c, 0));
1821                }
1822                break;
1823
1824        case nir_intrinsic_discard_if: {
1825                enum v3d_qpu_cond cond = ntq_emit_bool_to_cond(c, instr->src[0]);
1826
1827                if (vir_in_nonuniform_control_flow(c)) {
1828                        struct qinst *exec_flag = vir_MOV_dest(c, vir_nop_reg(),
1829                                                               c->execute);
1830                        if (cond == V3D_QPU_COND_IFA) {
1831                                vir_set_uf(exec_flag, V3D_QPU_UF_ANDZ);
1832                        } else {
1833                                vir_set_uf(exec_flag, V3D_QPU_UF_NORNZ);
1834                                cond = V3D_QPU_COND_IFA;
1835                        }
1836                }
1837
1838                vir_set_cond(vir_SETMSF_dest(c, vir_nop_reg(),
1839                                             vir_uniform_ui(c, 0)), cond);
1840
1841                break;
1842        }
1843
1844        case nir_intrinsic_memory_barrier:
1845        case nir_intrinsic_memory_barrier_atomic_counter:
1846        case nir_intrinsic_memory_barrier_buffer:
1847        case nir_intrinsic_memory_barrier_image:
1848        case nir_intrinsic_memory_barrier_shared:
1849        case nir_intrinsic_group_memory_barrier:
1850                /* We don't do any instruction scheduling of these NIR
1851                 * instructions between each other, so we just need to make
1852                 * sure that the TMU operations before the barrier are flushed
1853                 * before the ones after the barrier.  That is currently
1854                 * handled by having a THRSW in each of them and a LDTMU
1855                 * series or a TMUWT after.
1856                 */
1857                break;
1858
1859        case nir_intrinsic_barrier:
1860                /* Emit a TSY op to get all invocations in the workgroup
1861                 * (actually supergroup) to block until the last invocation
1862                 * reaches the TSY op.
1863                 */
1864                if (c->devinfo->ver >= 42) {
1865                        vir_BARRIERID_dest(c, vir_reg(QFILE_MAGIC,
1866                                                      V3D_QPU_WADDR_SYNCB));
1867                } else {
1868                        struct qinst *sync =
1869                                vir_BARRIERID_dest(c,
1870                                                   vir_reg(QFILE_MAGIC,
1871                                                           V3D_QPU_WADDR_SYNCU));
1872                        sync->uniform =
1873                                vir_get_uniform_index(c, QUNIFORM_CONSTANT,
1874                                                      0xffffff00 |
1875                                                      V3D_TSY_WAIT_INC_CHECK);
1876
1877                }
1878
1879                /* The blocking of a TSY op only happens at the next thread
1880                 * switch.  No texturing may be outstanding at the time of a
1881                 * TSY blocking operation.
1882                 */
1883                vir_emit_thrsw(c);
1884                break;
1885
1886        case nir_intrinsic_load_num_work_groups:
1887                for (int i = 0; i < 3; i++) {
1888                        ntq_store_dest(c, &instr->dest, i,
1889                                       vir_uniform(c, QUNIFORM_NUM_WORK_GROUPS,
1890                                                   i));
1891                }
1892                break;
1893
1894        case nir_intrinsic_load_local_invocation_index:
1895                ntq_store_dest(c, &instr->dest, 0,
1896                               vir_SHR(c, c->cs_payload[1],
1897                                       vir_uniform_ui(c, 32 - c->local_invocation_index_bits)));
1898                break;
1899
1900        case nir_intrinsic_load_work_group_id:
1901                ntq_store_dest(c, &instr->dest, 0,
1902                               vir_AND(c, c->cs_payload[0],
1903                                       vir_uniform_ui(c, 0xffff)));
1904                ntq_store_dest(c, &instr->dest, 1,
1905                               vir_SHR(c, c->cs_payload[0],
1906                                       vir_uniform_ui(c, 16)));
1907                ntq_store_dest(c, &instr->dest, 2,
1908                               vir_AND(c, c->cs_payload[1],
1909                                       vir_uniform_ui(c, 0xffff)));
1910                break;
1911
1912        case nir_intrinsic_load_subgroup_id:
1913                ntq_store_dest(c, &instr->dest, 0, vir_EIDX(c));
1914                break;
1915
1916        default:
1917                fprintf(stderr, "Unknown intrinsic: ");
1918                nir_print_instr(&instr->instr, stderr);
1919                fprintf(stderr, "\n");
1920                break;
1921        }
1922}
1923
1924/* Clears (activates) the execute flags for any channels whose jump target
1925 * matches this block.
1926 *
1927 * XXX perf: Could we be using flpush/flpop somehow for our execution channel
1928 * enabling?
1929 *
1930 * XXX perf: For uniform control flow, we should be able to skip c->execute
1931 * handling entirely.
1932 */
1933static void
1934ntq_activate_execute_for_block(struct v3d_compile *c)
1935{
1936        vir_set_pf(vir_XOR_dest(c, vir_nop_reg(),
1937                                c->execute, vir_uniform_ui(c, c->cur_block->index)),
1938                   V3D_QPU_PF_PUSHZ);
1939
1940        vir_MOV_cond(c, V3D_QPU_COND_IFA, c->execute, vir_uniform_ui(c, 0));
1941}
1942
1943static void
1944ntq_emit_uniform_if(struct v3d_compile *c, nir_if *if_stmt)
1945{
1946        nir_block *nir_else_block = nir_if_first_else_block(if_stmt);
1947        bool empty_else_block =
1948                (nir_else_block == nir_if_last_else_block(if_stmt) &&
1949                 exec_list_is_empty(&nir_else_block->instr_list));
1950
1951        struct qblock *then_block = vir_new_block(c);
1952        struct qblock *after_block = vir_new_block(c);
1953        struct qblock *else_block;
1954        if (empty_else_block)
1955                else_block = after_block;
1956        else
1957                else_block = vir_new_block(c);
1958
1959        /* Set up the flags for the IF condition (taking the THEN branch). */
1960        enum v3d_qpu_cond cond = ntq_emit_bool_to_cond(c, if_stmt->condition);
1961
1962        /* Jump to ELSE. */
1963        vir_BRANCH(c, cond == V3D_QPU_COND_IFA ?
1964                   V3D_QPU_BRANCH_COND_ALLNA :
1965                   V3D_QPU_BRANCH_COND_ALLA);
1966        vir_link_blocks(c->cur_block, else_block);
1967        vir_link_blocks(c->cur_block, then_block);
1968
1969        /* Process the THEN block. */
1970        vir_set_emit_block(c, then_block);
1971        ntq_emit_cf_list(c, &if_stmt->then_list);
1972
1973        if (!empty_else_block) {
1974                /* At the end of the THEN block, jump to ENDIF */
1975                vir_BRANCH(c, V3D_QPU_BRANCH_COND_ALWAYS);
1976                vir_link_blocks(c->cur_block, after_block);
1977
1978                /* Emit the else block. */
1979                vir_set_emit_block(c, else_block);
1980                ntq_activate_execute_for_block(c);
1981                ntq_emit_cf_list(c, &if_stmt->else_list);
1982        }
1983
1984        vir_link_blocks(c->cur_block, after_block);
1985
1986        vir_set_emit_block(c, after_block);
1987}
1988
1989static void
1990ntq_emit_nonuniform_if(struct v3d_compile *c, nir_if *if_stmt)
1991{
1992        nir_block *nir_else_block = nir_if_first_else_block(if_stmt);
1993        bool empty_else_block =
1994                (nir_else_block == nir_if_last_else_block(if_stmt) &&
1995                 exec_list_is_empty(&nir_else_block->instr_list));
1996
1997        struct qblock *then_block = vir_new_block(c);
1998        struct qblock *after_block = vir_new_block(c);
1999        struct qblock *else_block;
2000        if (empty_else_block)
2001                else_block = after_block;
2002        else
2003                else_block = vir_new_block(c);
2004
2005        bool was_uniform_control_flow = false;
2006        if (!vir_in_nonuniform_control_flow(c)) {
2007                c->execute = vir_MOV(c, vir_uniform_ui(c, 0));
2008                was_uniform_control_flow = true;
2009        }
2010
2011        /* Set up the flags for the IF condition (taking the THEN branch). */
2012        enum v3d_qpu_cond cond = ntq_emit_bool_to_cond(c, if_stmt->condition);
2013
2014        /* Update the flags+cond to mean "Taking the ELSE branch (!cond) and
2015         * was previously active (execute Z) for updating the exec flags.
2016         */
2017        if (was_uniform_control_flow) {
2018                cond = v3d_qpu_cond_invert(cond);
2019        } else {
2020                struct qinst *inst = vir_MOV_dest(c, vir_nop_reg(), c->execute);
2021                if (cond == V3D_QPU_COND_IFA) {
2022                        vir_set_uf(inst, V3D_QPU_UF_NORNZ);
2023                } else {
2024                        vir_set_uf(inst, V3D_QPU_UF_ANDZ);
2025                        cond = V3D_QPU_COND_IFA;
2026                }
2027        }
2028
2029        vir_MOV_cond(c, cond,
2030                     c->execute,
2031                     vir_uniform_ui(c, else_block->index));
2032
2033        /* Jump to ELSE if nothing is active for THEN, otherwise fall
2034         * through.
2035         */
2036        vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), c->execute), V3D_QPU_PF_PUSHZ);
2037        vir_BRANCH(c, V3D_QPU_BRANCH_COND_ALLNA);
2038        vir_link_blocks(c->cur_block, else_block);
2039        vir_link_blocks(c->cur_block, then_block);
2040
2041        /* Process the THEN block. */
2042        vir_set_emit_block(c, then_block);
2043        ntq_emit_cf_list(c, &if_stmt->then_list);
2044
2045        if (!empty_else_block) {
2046                /* Handle the end of the THEN block.  First, all currently
2047                 * active channels update their execute flags to point to
2048                 * ENDIF
2049                 */
2050                vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), c->execute),
2051                           V3D_QPU_PF_PUSHZ);
2052                vir_MOV_cond(c, V3D_QPU_COND_IFA, c->execute,
2053                             vir_uniform_ui(c, after_block->index));
2054
2055                /* If everything points at ENDIF, then jump there immediately. */
2056                vir_set_pf(vir_XOR_dest(c, vir_nop_reg(),
2057                                        c->execute,
2058                                        vir_uniform_ui(c, after_block->index)),
2059                           V3D_QPU_PF_PUSHZ);
2060                vir_BRANCH(c, V3D_QPU_BRANCH_COND_ALLA);
2061                vir_link_blocks(c->cur_block, after_block);
2062                vir_link_blocks(c->cur_block, else_block);
2063
2064                vir_set_emit_block(c, else_block);
2065                ntq_activate_execute_for_block(c);
2066                ntq_emit_cf_list(c, &if_stmt->else_list);
2067        }
2068
2069        vir_link_blocks(c->cur_block, after_block);
2070
2071        vir_set_emit_block(c, after_block);
2072        if (was_uniform_control_flow)
2073                c->execute = c->undef;
2074        else
2075                ntq_activate_execute_for_block(c);
2076}
2077
2078static void
2079ntq_emit_if(struct v3d_compile *c, nir_if *nif)
2080{
2081        bool was_in_control_flow = c->in_control_flow;
2082        c->in_control_flow = true;
2083        if (!vir_in_nonuniform_control_flow(c) &&
2084            nir_src_is_dynamically_uniform(nif->condition)) {
2085                ntq_emit_uniform_if(c, nif);
2086        } else {
2087                ntq_emit_nonuniform_if(c, nif);
2088        }
2089        c->in_control_flow = was_in_control_flow;
2090}
2091
2092static void
2093ntq_emit_jump(struct v3d_compile *c, nir_jump_instr *jump)
2094{
2095        switch (jump->type) {
2096        case nir_jump_break:
2097                vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), c->execute),
2098                           V3D_QPU_PF_PUSHZ);
2099                vir_MOV_cond(c, V3D_QPU_COND_IFA, c->execute,
2100                             vir_uniform_ui(c, c->loop_break_block->index));
2101                break;
2102
2103        case nir_jump_continue:
2104                vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), c->execute),
2105                           V3D_QPU_PF_PUSHZ);
2106                vir_MOV_cond(c, V3D_QPU_COND_IFA, c->execute,
2107                             vir_uniform_ui(c, c->loop_cont_block->index));
2108                break;
2109
2110        case nir_jump_return:
2111                unreachable("All returns shouold be lowered\n");
2112        }
2113}
2114
2115static void
2116ntq_emit_instr(struct v3d_compile *c, nir_instr *instr)
2117{
2118        switch (instr->type) {
2119        case nir_instr_type_deref:
2120                /* ignored, will be walked by the intrinsic using it. */
2121                break;
2122
2123        case nir_instr_type_alu:
2124                ntq_emit_alu(c, nir_instr_as_alu(instr));
2125                break;
2126
2127        case nir_instr_type_intrinsic:
2128                ntq_emit_intrinsic(c, nir_instr_as_intrinsic(instr));
2129                break;
2130
2131        case nir_instr_type_load_const:
2132                ntq_emit_load_const(c, nir_instr_as_load_const(instr));
2133                break;
2134
2135        case nir_instr_type_ssa_undef:
2136                ntq_emit_ssa_undef(c, nir_instr_as_ssa_undef(instr));
2137                break;
2138
2139        case nir_instr_type_tex:
2140                ntq_emit_tex(c, nir_instr_as_tex(instr));
2141                break;
2142
2143        case nir_instr_type_jump:
2144                ntq_emit_jump(c, nir_instr_as_jump(instr));
2145                break;
2146
2147        default:
2148                fprintf(stderr, "Unknown NIR instr type: ");
2149                nir_print_instr(instr, stderr);
2150                fprintf(stderr, "\n");
2151                abort();
2152        }
2153}
2154
2155static void
2156ntq_emit_block(struct v3d_compile *c, nir_block *block)
2157{
2158        nir_foreach_instr(instr, block) {
2159                ntq_emit_instr(c, instr);
2160        }
2161}
2162
2163static void ntq_emit_cf_list(struct v3d_compile *c, struct exec_list *list);
2164
2165static void
2166ntq_emit_loop(struct v3d_compile *c, nir_loop *loop)
2167{
2168        bool was_in_control_flow = c->in_control_flow;
2169        c->in_control_flow = true;
2170
2171        bool was_uniform_control_flow = false;
2172        if (!vir_in_nonuniform_control_flow(c)) {
2173                c->execute = vir_MOV(c, vir_uniform_ui(c, 0));
2174                was_uniform_control_flow = true;
2175        }
2176
2177        struct qblock *save_loop_cont_block = c->loop_cont_block;
2178        struct qblock *save_loop_break_block = c->loop_break_block;
2179
2180        c->loop_cont_block = vir_new_block(c);
2181        c->loop_break_block = vir_new_block(c);
2182
2183        vir_link_blocks(c->cur_block, c->loop_cont_block);
2184        vir_set_emit_block(c, c->loop_cont_block);
2185        ntq_activate_execute_for_block(c);
2186
2187        ntq_emit_cf_list(c, &loop->body);
2188
2189        /* Re-enable any previous continues now, so our ANYA check below
2190         * works.
2191         *
2192         * XXX: Use the .ORZ flags update, instead.
2193         */
2194        vir_set_pf(vir_XOR_dest(c,
2195                                vir_nop_reg(),
2196                                c->execute,
2197                                vir_uniform_ui(c, c->loop_cont_block->index)),
2198                   V3D_QPU_PF_PUSHZ);
2199        vir_MOV_cond(c, V3D_QPU_COND_IFA, c->execute, vir_uniform_ui(c, 0));
2200
2201        vir_set_pf(vir_MOV_dest(c, vir_nop_reg(), c->execute), V3D_QPU_PF_PUSHZ);
2202
2203        struct qinst *branch = vir_BRANCH(c, V3D_QPU_BRANCH_COND_ANYA);
2204        /* Pixels that were not dispatched or have been discarded should not
2205         * contribute to looping again.
2206         */
2207        branch->qpu.branch.msfign = V3D_QPU_MSFIGN_P;
2208        vir_link_blocks(c->cur_block, c->loop_cont_block);
2209        vir_link_blocks(c->cur_block, c->loop_break_block);
2210
2211        vir_set_emit_block(c, c->loop_break_block);
2212        if (was_uniform_control_flow)
2213                c->execute = c->undef;
2214        else
2215                ntq_activate_execute_for_block(c);
2216
2217        c->loop_break_block = save_loop_break_block;
2218        c->loop_cont_block = save_loop_cont_block;
2219
2220        c->loops++;
2221
2222        c->in_control_flow = was_in_control_flow;
2223}
2224
2225static void
2226ntq_emit_function(struct v3d_compile *c, nir_function_impl *func)
2227{
2228        fprintf(stderr, "FUNCTIONS not handled.\n");
2229        abort();
2230}
2231
2232static void
2233ntq_emit_cf_list(struct v3d_compile *c, struct exec_list *list)
2234{
2235        foreach_list_typed(nir_cf_node, node, node, list) {
2236                switch (node->type) {
2237                case nir_cf_node_block:
2238                        ntq_emit_block(c, nir_cf_node_as_block(node));
2239                        break;
2240
2241                case nir_cf_node_if:
2242                        ntq_emit_if(c, nir_cf_node_as_if(node));
2243                        break;
2244
2245                case nir_cf_node_loop:
2246                        ntq_emit_loop(c, nir_cf_node_as_loop(node));
2247                        break;
2248
2249                case nir_cf_node_function:
2250                        ntq_emit_function(c, nir_cf_node_as_function(node));
2251                        break;
2252
2253                default:
2254                        fprintf(stderr, "Unknown NIR node type\n");
2255                        abort();
2256                }
2257        }
2258}
2259
2260static void
2261ntq_emit_impl(struct v3d_compile *c, nir_function_impl *impl)
2262{
2263        ntq_setup_registers(c, &impl->registers);
2264        ntq_emit_cf_list(c, &impl->body);
2265}
2266
2267static void
2268nir_to_vir(struct v3d_compile *c)
2269{
2270        switch (c->s->info.stage) {
2271        case MESA_SHADER_FRAGMENT:
2272                c->payload_w = vir_MOV(c, vir_reg(QFILE_REG, 0));
2273                c->payload_w_centroid = vir_MOV(c, vir_reg(QFILE_REG, 1));
2274                c->payload_z = vir_MOV(c, vir_reg(QFILE_REG, 2));
2275
2276                /* XXX perf: We could set the "disable implicit point/line
2277                 * varyings" field in the shader record and not emit these, if
2278                 * they're not going to be used.
2279                 */
2280                if (c->fs_key->is_points) {
2281                        c->point_x = emit_fragment_varying(c, NULL, 0, 0);
2282                        c->point_y = emit_fragment_varying(c, NULL, 0, 0);
2283                } else if (c->fs_key->is_lines) {
2284                        c->line_x = emit_fragment_varying(c, NULL, 0, 0);
2285                }
2286                break;
2287        case MESA_SHADER_COMPUTE:
2288                /* Set up the TSO for barriers, assuming we do some. */
2289                if (c->devinfo->ver < 42) {
2290                        vir_BARRIERID_dest(c, vir_reg(QFILE_MAGIC,
2291                                                      V3D_QPU_WADDR_SYNC));
2292                }
2293
2294                c->cs_payload[0] = vir_MOV(c, vir_reg(QFILE_REG, 0));
2295                c->cs_payload[1] = vir_MOV(c, vir_reg(QFILE_REG, 2));
2296
2297                /* Set up the division between gl_LocalInvocationIndex and
2298                 * wg_in_mem in the payload reg.
2299                 */
2300                int wg_size = (c->s->info.cs.local_size[0] *
2301                               c->s->info.cs.local_size[1] *
2302                               c->s->info.cs.local_size[2]);
2303                c->local_invocation_index_bits =
2304                        ffs(util_next_power_of_two(MAX2(wg_size, 64))) - 1;
2305                assert(c->local_invocation_index_bits <= 8);
2306
2307                if (c->s->info.cs.shared_size) {
2308                        struct qreg wg_in_mem = vir_SHR(c, c->cs_payload[1],
2309                                                        vir_uniform_ui(c, 16));
2310                        if (c->s->info.cs.local_size[0] != 1 ||
2311                            c->s->info.cs.local_size[1] != 1 ||
2312                            c->s->info.cs.local_size[2] != 1) {
2313                                int wg_bits = (16 -
2314                                               c->local_invocation_index_bits);
2315                                int wg_mask = (1 << wg_bits) - 1;
2316                                wg_in_mem = vir_AND(c, wg_in_mem,
2317                                                    vir_uniform_ui(c, wg_mask));
2318                        }
2319                        struct qreg shared_per_wg =
2320                                vir_uniform_ui(c, c->s->info.cs.shared_size);
2321
2322                        c->cs_shared_offset =
2323                                vir_ADD(c,
2324                                        vir_uniform(c, QUNIFORM_SHARED_OFFSET,0),
2325                                        vir_UMUL(c, wg_in_mem, shared_per_wg));
2326                }
2327                break;
2328        default:
2329                break;
2330        }
2331
2332        if (c->s->scratch_size) {
2333                v3d_setup_spill_base(c);
2334                c->spill_size += V3D_CHANNELS * c->s->scratch_size;
2335        }
2336
2337        if (c->s->info.stage == MESA_SHADER_FRAGMENT)
2338                ntq_setup_fs_inputs(c);
2339        else
2340                ntq_setup_vpm_inputs(c);
2341
2342        ntq_setup_outputs(c);
2343
2344        /* Find the main function and emit the body. */
2345        nir_foreach_function(function, c->s) {
2346                assert(strcmp(function->name, "main") == 0);
2347                assert(function->impl);
2348                ntq_emit_impl(c, function->impl);
2349        }
2350}
2351
2352const nir_shader_compiler_options v3d_nir_options = {
2353        .lower_all_io_to_temps = true,
2354        .lower_extract_byte = true,
2355        .lower_extract_word = true,
2356        .lower_bfm = true,
2357        .lower_bitfield_insert_to_shifts = true,
2358        .lower_bitfield_extract_to_shifts = true,
2359        .lower_bitfield_reverse = true,
2360        .lower_bit_count = true,
2361        .lower_cs_local_id_from_index = true,
2362        .lower_ffract = true,
2363        .lower_pack_unorm_2x16 = true,
2364        .lower_pack_snorm_2x16 = true,
2365        .lower_pack_unorm_4x8 = true,
2366        .lower_pack_snorm_4x8 = true,
2367        .lower_unpack_unorm_4x8 = true,
2368        .lower_unpack_snorm_4x8 = true,
2369        .lower_pack_half_2x16 = true,
2370        .lower_unpack_half_2x16 = true,
2371        .lower_fdiv = true,
2372        .lower_find_lsb = true,
2373        .lower_ffma = true,
2374        .lower_flrp32 = true,
2375        .lower_fpow = true,
2376        .lower_fsat = true,
2377        .lower_fsqrt = true,
2378        .lower_ifind_msb = true,
2379        .lower_isign = true,
2380        .lower_ldexp = true,
2381        .lower_mul_high = true,
2382        .lower_wpos_pntc = true,
2383        .native_integers = true,
2384};
2385
2386/**
2387 * When demoting a shader down to single-threaded, removes the THRSW
2388 * instructions (one will still be inserted at v3d_vir_to_qpu() for the
2389 * program end).
2390 */
2391static void
2392vir_remove_thrsw(struct v3d_compile *c)
2393{
2394        vir_for_each_block(block, c) {
2395                vir_for_each_inst_safe(inst, block) {
2396                        if (inst->qpu.sig.thrsw)
2397                                vir_remove_instruction(c, inst);
2398                }
2399        }
2400
2401        c->last_thrsw = NULL;
2402}
2403
2404void
2405vir_emit_last_thrsw(struct v3d_compile *c)
2406{
2407        /* On V3D before 4.1, we need a TMU op to be outstanding when thread
2408         * switching, so disable threads if we didn't do any TMU ops (each of
2409         * which would have emitted a THRSW).
2410         */
2411        if (!c->last_thrsw_at_top_level && c->devinfo->ver < 41) {
2412                c->threads = 1;
2413                if (c->last_thrsw)
2414                        vir_remove_thrsw(c);
2415                return;
2416        }
2417
2418        /* If we're threaded and the last THRSW was in conditional code, then
2419         * we need to emit another one so that we can flag it as the last
2420         * thrsw.
2421         */
2422        if (c->last_thrsw && !c->last_thrsw_at_top_level) {
2423                assert(c->devinfo->ver >= 41);
2424                vir_emit_thrsw(c);
2425        }
2426
2427        /* If we're threaded, then we need to mark the last THRSW instruction
2428         * so we can emit a pair of them at QPU emit time.
2429         *
2430         * For V3D 4.x, we can spawn the non-fragment shaders already in the
2431         * post-last-THRSW state, so we can skip this.
2432         */
2433        if (!c->last_thrsw && c->s->info.stage == MESA_SHADER_FRAGMENT) {
2434                assert(c->devinfo->ver >= 41);
2435                vir_emit_thrsw(c);
2436        }
2437
2438        if (c->last_thrsw)
2439                c->last_thrsw->is_last_thrsw = true;
2440}
2441
2442/* There's a flag in the shader for "center W is needed for reasons other than
2443 * non-centroid varyings", so we just walk the program after VIR optimization
2444 * to see if it's used.  It should be harmless to set even if we only use
2445 * center W for varyings.
2446 */
2447static void
2448vir_check_payload_w(struct v3d_compile *c)
2449{
2450        if (c->s->info.stage != MESA_SHADER_FRAGMENT)
2451                return;
2452
2453        vir_for_each_inst_inorder(inst, c) {
2454                for (int i = 0; i < vir_get_nsrc(inst); i++) {
2455                        if (inst->src[i].file == QFILE_REG &&
2456                            inst->src[i].index == 0) {
2457                                c->uses_center_w = true;
2458                                return;
2459                        }
2460                }
2461        }
2462
2463}
2464
2465void
2466v3d_nir_to_vir(struct v3d_compile *c)
2467{
2468        if (V3D_DEBUG & (V3D_DEBUG_NIR |
2469                         v3d_debug_flag_for_shader_stage(c->s->info.stage))) {
2470                fprintf(stderr, "%s prog %d/%d NIR:\n",
2471                        vir_get_stage_name(c),
2472                        c->program_id, c->variant_id);
2473                nir_print_shader(c->s, stderr);
2474        }
2475
2476        nir_to_vir(c);
2477
2478        /* Emit the last THRSW before STVPM and TLB writes. */
2479        vir_emit_last_thrsw(c);
2480
2481        switch (c->s->info.stage) {
2482        case MESA_SHADER_FRAGMENT:
2483                emit_frag_end(c);
2484                break;
2485        case MESA_SHADER_VERTEX:
2486                emit_vert_end(c);
2487                break;
2488        case MESA_SHADER_COMPUTE:
2489                break;
2490        default:
2491                unreachable("bad stage");
2492        }
2493
2494        if (V3D_DEBUG & (V3D_DEBUG_VIR |
2495                         v3d_debug_flag_for_shader_stage(c->s->info.stage))) {
2496                fprintf(stderr, "%s prog %d/%d pre-opt VIR:\n",
2497                        vir_get_stage_name(c),
2498                        c->program_id, c->variant_id);
2499                vir_dump(c);
2500                fprintf(stderr, "\n");
2501        }
2502
2503        vir_optimize(c);
2504
2505        vir_check_payload_w(c);
2506
2507        /* XXX perf: On VC4, we do a VIR-level instruction scheduling here.
2508         * We used that on that platform to pipeline TMU writes and reduce the
2509         * number of thread switches, as well as try (mostly successfully) to
2510         * reduce maximum register pressure to allow more threads.  We should
2511         * do something of that sort for V3D -- either instruction scheduling
2512         * here, or delay the the THRSW and LDTMUs from our texture
2513         * instructions until the results are needed.
2514         */
2515
2516        if (V3D_DEBUG & (V3D_DEBUG_VIR |
2517                         v3d_debug_flag_for_shader_stage(c->s->info.stage))) {
2518                fprintf(stderr, "%s prog %d/%d VIR:\n",
2519                        vir_get_stage_name(c),
2520                        c->program_id, c->variant_id);
2521                vir_dump(c);
2522                fprintf(stderr, "\n");
2523        }
2524
2525        /* Attempt to allocate registers for the temporaries.  If we fail,
2526         * reduce thread count and try again.
2527         */
2528        int min_threads = (c->devinfo->ver >= 41) ? 2 : 1;
2529        struct qpu_reg *temp_registers;
2530        while (true) {
2531                bool spilled;
2532                temp_registers = v3d_register_allocate(c, &spilled);
2533                if (spilled)
2534                        continue;
2535
2536                if (temp_registers)
2537                        break;
2538
2539                if (c->threads == min_threads) {
2540                        fprintf(stderr, "Failed to register allocate at %d threads:\n",
2541                                c->threads);
2542                        vir_dump(c);
2543                        c->failed = true;
2544                        return;
2545                }
2546
2547                c->threads /= 2;
2548
2549                if (c->threads == 1)
2550                        vir_remove_thrsw(c);
2551        }
2552
2553        if (c->spills &&
2554            (V3D_DEBUG & (V3D_DEBUG_VIR |
2555                          v3d_debug_flag_for_shader_stage(c->s->info.stage)))) {
2556                fprintf(stderr, "%s prog %d/%d spilled VIR:\n",
2557                        vir_get_stage_name(c),
2558                        c->program_id, c->variant_id);
2559                vir_dump(c);
2560                fprintf(stderr, "\n");
2561        }
2562
2563        v3d_vir_to_qpu(c, temp_registers);
2564}
2565