1 /*
2  * Copyright © 2015 Intel Corporation
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 "brw_nir.h"
25 #include "brw_vec4.h"
26 #include "brw_vec4_builder.h"
27 #include "brw_vec4_surface_builder.h"
28 #include "brw_eu.h"
29 
30 using namespace brw;
31 using namespace brw::surface_access;
32 
33 namespace brw {
34 
35 void
emit_nir_code()36 vec4_visitor::emit_nir_code()
37 {
38    if (nir->num_uniforms > 0)
39       nir_setup_uniforms();
40 
41    nir_emit_impl(nir_shader_get_entrypoint((nir_shader *)nir));
42 }
43 
44 void
nir_setup_uniforms()45 vec4_visitor::nir_setup_uniforms()
46 {
47    uniforms = nir->num_uniforms / 16;
48 }
49 
50 void
nir_emit_impl(nir_function_impl * impl)51 vec4_visitor::nir_emit_impl(nir_function_impl *impl)
52 {
53    nir_locals = ralloc_array(mem_ctx, dst_reg, impl->reg_alloc);
54    for (unsigned i = 0; i < impl->reg_alloc; i++) {
55       nir_locals[i] = dst_reg();
56    }
57 
58    foreach_list_typed(nir_register, reg, node, &impl->registers) {
59       unsigned array_elems =
60          reg->num_array_elems == 0 ? 1 : reg->num_array_elems;
61       const unsigned num_regs = array_elems * DIV_ROUND_UP(reg->bit_size, 32);
62       nir_locals[reg->index] = dst_reg(VGRF, alloc.allocate(num_regs));
63 
64       if (reg->bit_size == 64)
65          nir_locals[reg->index].type = BRW_REGISTER_TYPE_DF;
66    }
67 
68    nir_ssa_values = ralloc_array(mem_ctx, dst_reg, impl->ssa_alloc);
69 
70    nir_emit_cf_list(&impl->body);
71 }
72 
73 void
nir_emit_cf_list(exec_list * list)74 vec4_visitor::nir_emit_cf_list(exec_list *list)
75 {
76    exec_list_validate(list);
77    foreach_list_typed(nir_cf_node, node, node, list) {
78       switch (node->type) {
79       case nir_cf_node_if:
80          nir_emit_if(nir_cf_node_as_if(node));
81          break;
82 
83       case nir_cf_node_loop:
84          nir_emit_loop(nir_cf_node_as_loop(node));
85          break;
86 
87       case nir_cf_node_block:
88          nir_emit_block(nir_cf_node_as_block(node));
89          break;
90 
91       default:
92          unreachable("Invalid CFG node block");
93       }
94    }
95 }
96 
97 void
nir_emit_if(nir_if * if_stmt)98 vec4_visitor::nir_emit_if(nir_if *if_stmt)
99 {
100    /* First, put the condition in f0 */
101    src_reg condition = get_nir_src(if_stmt->condition, BRW_REGISTER_TYPE_D, 1);
102    vec4_instruction *inst = emit(MOV(dst_null_d(), condition));
103    inst->conditional_mod = BRW_CONDITIONAL_NZ;
104 
105    /* We can just predicate based on the X channel, as the condition only
106     * goes on its own line */
107    emit(IF(BRW_PREDICATE_ALIGN16_REPLICATE_X));
108 
109    nir_emit_cf_list(&if_stmt->then_list);
110 
111    /* note: if the else is empty, dead CF elimination will remove it */
112    emit(BRW_OPCODE_ELSE);
113 
114    nir_emit_cf_list(&if_stmt->else_list);
115 
116    emit(BRW_OPCODE_ENDIF);
117 }
118 
119 void
nir_emit_loop(nir_loop * loop)120 vec4_visitor::nir_emit_loop(nir_loop *loop)
121 {
122    emit(BRW_OPCODE_DO);
123 
124    nir_emit_cf_list(&loop->body);
125 
126    emit(BRW_OPCODE_WHILE);
127 }
128 
129 void
nir_emit_block(nir_block * block)130 vec4_visitor::nir_emit_block(nir_block *block)
131 {
132    nir_foreach_instr(instr, block) {
133       nir_emit_instr(instr);
134    }
135 }
136 
137 void
nir_emit_instr(nir_instr * instr)138 vec4_visitor::nir_emit_instr(nir_instr *instr)
139 {
140    base_ir = instr;
141 
142    switch (instr->type) {
143    case nir_instr_type_load_const:
144       nir_emit_load_const(nir_instr_as_load_const(instr));
145       break;
146 
147    case nir_instr_type_intrinsic:
148       nir_emit_intrinsic(nir_instr_as_intrinsic(instr));
149       break;
150 
151    case nir_instr_type_alu:
152       nir_emit_alu(nir_instr_as_alu(instr));
153       break;
154 
155    case nir_instr_type_jump:
156       nir_emit_jump(nir_instr_as_jump(instr));
157       break;
158 
159    case nir_instr_type_tex:
160       nir_emit_texture(nir_instr_as_tex(instr));
161       break;
162 
163    case nir_instr_type_ssa_undef:
164       nir_emit_undef(nir_instr_as_ssa_undef(instr));
165       break;
166 
167    default:
168       unreachable("VS instruction not yet implemented by NIR->vec4");
169    }
170 }
171 
172 static dst_reg
dst_reg_for_nir_reg(vec4_visitor * v,nir_register * nir_reg,unsigned base_offset,nir_src * indirect)173 dst_reg_for_nir_reg(vec4_visitor *v, nir_register *nir_reg,
174                     unsigned base_offset, nir_src *indirect)
175 {
176    dst_reg reg;
177 
178    reg = v->nir_locals[nir_reg->index];
179    if (nir_reg->bit_size == 64)
180       reg.type = BRW_REGISTER_TYPE_DF;
181    reg = offset(reg, 8, base_offset);
182    if (indirect) {
183       reg.reladdr =
184          new(v->mem_ctx) src_reg(v->get_nir_src(*indirect,
185                                                 BRW_REGISTER_TYPE_D,
186                                                 1));
187    }
188    return reg;
189 }
190 
191 dst_reg
get_nir_dest(const nir_dest & dest)192 vec4_visitor::get_nir_dest(const nir_dest &dest)
193 {
194    if (dest.is_ssa) {
195       dst_reg dst =
196          dst_reg(VGRF, alloc.allocate(DIV_ROUND_UP(dest.ssa.bit_size, 32)));
197       if (dest.ssa.bit_size == 64)
198          dst.type = BRW_REGISTER_TYPE_DF;
199       nir_ssa_values[dest.ssa.index] = dst;
200       return dst;
201    } else {
202       return dst_reg_for_nir_reg(this, dest.reg.reg, dest.reg.base_offset,
203                                  dest.reg.indirect);
204    }
205 }
206 
207 dst_reg
get_nir_dest(const nir_dest & dest,enum brw_reg_type type)208 vec4_visitor::get_nir_dest(const nir_dest &dest, enum brw_reg_type type)
209 {
210    return retype(get_nir_dest(dest), type);
211 }
212 
213 dst_reg
get_nir_dest(const nir_dest & dest,nir_alu_type type)214 vec4_visitor::get_nir_dest(const nir_dest &dest, nir_alu_type type)
215 {
216    return get_nir_dest(dest, brw_type_for_nir_type(devinfo, type));
217 }
218 
219 src_reg
get_nir_src(const nir_src & src,enum brw_reg_type type,unsigned num_components)220 vec4_visitor::get_nir_src(const nir_src &src, enum brw_reg_type type,
221                           unsigned num_components)
222 {
223    dst_reg reg;
224 
225    if (src.is_ssa) {
226       assert(src.ssa != NULL);
227       reg = nir_ssa_values[src.ssa->index];
228    }
229    else {
230       reg = dst_reg_for_nir_reg(this, src.reg.reg, src.reg.base_offset,
231                                 src.reg.indirect);
232    }
233 
234    reg = retype(reg, type);
235 
236    src_reg reg_as_src = src_reg(reg);
237    reg_as_src.swizzle = brw_swizzle_for_size(num_components);
238    return reg_as_src;
239 }
240 
241 src_reg
get_nir_src(const nir_src & src,nir_alu_type type,unsigned num_components)242 vec4_visitor::get_nir_src(const nir_src &src, nir_alu_type type,
243                           unsigned num_components)
244 {
245    return get_nir_src(src, brw_type_for_nir_type(devinfo, type),
246                       num_components);
247 }
248 
249 src_reg
get_nir_src(const nir_src & src,unsigned num_components)250 vec4_visitor::get_nir_src(const nir_src &src, unsigned num_components)
251 {
252    /* if type is not specified, default to signed int */
253    return get_nir_src(src, nir_type_int32, num_components);
254 }
255 
256 src_reg
get_nir_src_imm(const nir_src & src)257 vec4_visitor::get_nir_src_imm(const nir_src &src)
258 {
259    assert(nir_src_num_components(src) == 1);
260    assert(nir_src_bit_size(src) == 32);
261    return nir_src_is_const(src) ? src_reg(brw_imm_d(nir_src_as_int(src))) :
262                                   get_nir_src(src, 1);
263 }
264 
265 src_reg
get_indirect_offset(nir_intrinsic_instr * instr)266 vec4_visitor::get_indirect_offset(nir_intrinsic_instr *instr)
267 {
268    nir_src *offset_src = nir_get_io_offset_src(instr);
269 
270    if (nir_src_is_const(*offset_src)) {
271       /* The only constant offset we should find is 0.  brw_nir.c's
272        * add_const_offset_to_base() will fold other constant offsets
273        * into instr->const_index[0].
274        */
275       assert(nir_src_as_uint(*offset_src) == 0);
276       return src_reg();
277    }
278 
279    return get_nir_src(*offset_src, BRW_REGISTER_TYPE_UD, 1);
280 }
281 
282 static src_reg
setup_imm_df(const vec4_builder & bld,double v)283 setup_imm_df(const vec4_builder &bld, double v)
284 {
285    const intel_device_info *devinfo = bld.shader->devinfo;
286    assert(devinfo->ver == 7);
287 
288    /* gfx7.5 does not support DF immediates straighforward but the DIM
289     * instruction allows to set the 64-bit immediate value.
290     */
291    if (devinfo->verx10 == 75) {
292       const vec4_builder ubld = bld.exec_all();
293       const dst_reg dst = bld.vgrf(BRW_REGISTER_TYPE_DF);
294       ubld.DIM(dst, brw_imm_df(v));
295       return swizzle(src_reg(dst), BRW_SWIZZLE_XXXX);
296    }
297 
298    /* gfx7 does not support DF immediates */
299    union {
300       double d;
301       struct {
302          uint32_t i1;
303          uint32_t i2;
304       };
305    } di;
306 
307    di.d = v;
308 
309    /* Write the low 32-bit of the constant to the X:UD channel and the
310     * high 32-bit to the Y:UD channel to build the constant in a VGRF.
311     * We have to do this twice (offset 0 and offset 1), since a DF VGRF takes
312     * two SIMD8 registers in SIMD4x2 execution. Finally, return a swizzle
313     * XXXX so any access to the VGRF only reads the constant data in these
314     * channels.
315     */
316    const dst_reg tmp = bld.vgrf(BRW_REGISTER_TYPE_UD, 2);
317    for (unsigned n = 0; n < 2; n++) {
318       const vec4_builder ubld = bld.exec_all().group(4, n);
319       ubld.MOV(writemask(offset(tmp, 8, n), WRITEMASK_X), brw_imm_ud(di.i1));
320       ubld.MOV(writemask(offset(tmp, 8, n), WRITEMASK_Y), brw_imm_ud(di.i2));
321    }
322 
323    return swizzle(src_reg(retype(tmp, BRW_REGISTER_TYPE_DF)), BRW_SWIZZLE_XXXX);
324 }
325 
326 void
nir_emit_load_const(nir_load_const_instr * instr)327 vec4_visitor::nir_emit_load_const(nir_load_const_instr *instr)
328 {
329    dst_reg reg;
330 
331    if (instr->def.bit_size == 64) {
332       reg = dst_reg(VGRF, alloc.allocate(2));
333       reg.type = BRW_REGISTER_TYPE_DF;
334    } else {
335       reg = dst_reg(VGRF, alloc.allocate(1));
336       reg.type = BRW_REGISTER_TYPE_D;
337    }
338 
339    const vec4_builder ibld = vec4_builder(this).at_end();
340    unsigned remaining = brw_writemask_for_size(instr->def.num_components);
341 
342    /* @FIXME: consider emitting vector operations to save some MOVs in
343     * cases where the components are representable in 8 bits.
344     * For now, we emit a MOV for each distinct value.
345     */
346    for (unsigned i = 0; i < instr->def.num_components; i++) {
347       unsigned writemask = 1 << i;
348 
349       if ((remaining & writemask) == 0)
350          continue;
351 
352       for (unsigned j = i; j < instr->def.num_components; j++) {
353          if ((instr->def.bit_size == 32 &&
354               instr->value[i].u32 == instr->value[j].u32) ||
355              (instr->def.bit_size == 64 &&
356               instr->value[i].f64 == instr->value[j].f64)) {
357             writemask |= 1 << j;
358          }
359       }
360 
361       reg.writemask = writemask;
362       if (instr->def.bit_size == 64) {
363          emit(MOV(reg, setup_imm_df(ibld, instr->value[i].f64)));
364       } else {
365          emit(MOV(reg, brw_imm_d(instr->value[i].i32)));
366       }
367 
368       remaining &= ~writemask;
369    }
370 
371    /* Set final writemask */
372    reg.writemask = brw_writemask_for_size(instr->def.num_components);
373 
374    nir_ssa_values[instr->def.index] = reg;
375 }
376 
377 src_reg
get_nir_ssbo_intrinsic_index(nir_intrinsic_instr * instr)378 vec4_visitor::get_nir_ssbo_intrinsic_index(nir_intrinsic_instr *instr)
379 {
380    /* SSBO stores are weird in that their index is in src[1] */
381    const unsigned src = instr->intrinsic == nir_intrinsic_store_ssbo ? 1 : 0;
382 
383    if (nir_src_is_const(instr->src[src])) {
384       return brw_imm_ud(nir_src_as_uint(instr->src[src]));
385    } else {
386       return emit_uniformize(get_nir_src(instr->src[src]));
387    }
388 }
389 
390 void
nir_emit_intrinsic(nir_intrinsic_instr * instr)391 vec4_visitor::nir_emit_intrinsic(nir_intrinsic_instr *instr)
392 {
393    dst_reg dest;
394    src_reg src;
395 
396    switch (instr->intrinsic) {
397 
398    case nir_intrinsic_load_input: {
399       assert(nir_dest_bit_size(instr->dest) == 32);
400       /* We set EmitNoIndirectInput for VS */
401       unsigned load_offset = nir_src_as_uint(instr->src[0]);
402 
403       dest = get_nir_dest(instr->dest);
404       dest.writemask = brw_writemask_for_size(instr->num_components);
405 
406       src = src_reg(ATTR, instr->const_index[0] + load_offset,
407                     glsl_type::uvec4_type);
408       src = retype(src, dest.type);
409 
410       /* Swizzle source based on component layout qualifier */
411       src.swizzle = BRW_SWZ_COMP_INPUT(nir_intrinsic_component(instr));
412       emit(MOV(dest, src));
413       break;
414    }
415 
416    case nir_intrinsic_store_output: {
417       assert(nir_src_bit_size(instr->src[0]) == 32);
418       unsigned store_offset = nir_src_as_uint(instr->src[1]);
419       int varying = instr->const_index[0] + store_offset;
420       src = get_nir_src(instr->src[0], BRW_REGISTER_TYPE_F,
421                         instr->num_components);
422 
423       unsigned c = nir_intrinsic_component(instr);
424       output_reg[varying][c] = dst_reg(src);
425       output_num_components[varying][c] = instr->num_components;
426       break;
427    }
428 
429    case nir_intrinsic_get_ssbo_size: {
430       assert(nir_src_num_components(instr->src[0]) == 1);
431       unsigned ssbo_index = nir_src_is_const(instr->src[0]) ?
432                             nir_src_as_uint(instr->src[0]) : 0;
433 
434       dst_reg result_dst = get_nir_dest(instr->dest);
435       vec4_instruction *inst = new(mem_ctx)
436          vec4_instruction(SHADER_OPCODE_GET_BUFFER_SIZE, result_dst);
437 
438       inst->base_mrf = 2;
439       inst->mlen = 1; /* always at least one */
440       inst->src[1] = brw_imm_ud(ssbo_index);
441 
442       /* MRF for the first parameter */
443       src_reg lod = brw_imm_d(0);
444       int param_base = inst->base_mrf;
445       int writemask = WRITEMASK_X;
446       emit(MOV(dst_reg(MRF, param_base, glsl_type::int_type, writemask), lod));
447 
448       emit(inst);
449       break;
450    }
451 
452    case nir_intrinsic_store_ssbo: {
453       assert(devinfo->ver == 7);
454 
455       /* brw_nir_lower_mem_access_bit_sizes takes care of this */
456       assert(nir_src_bit_size(instr->src[0]) == 32);
457       assert(nir_intrinsic_write_mask(instr) ==
458              (1u << instr->num_components) - 1);
459 
460       src_reg surf_index = get_nir_ssbo_intrinsic_index(instr);
461       src_reg offset_reg = retype(get_nir_src_imm(instr->src[2]),
462                                   BRW_REGISTER_TYPE_UD);
463 
464       /* Value */
465       src_reg val_reg = get_nir_src(instr->src[0], BRW_REGISTER_TYPE_F, 4);
466 
467       /* IvyBridge does not have a native SIMD4x2 untyped write message so untyped
468        * writes will use SIMD8 mode. In order to hide this and keep symmetry across
469        * typed and untyped messages and across hardware platforms, the
470        * current implementation of the untyped messages will transparently convert
471        * the SIMD4x2 payload into an equivalent SIMD8 payload by transposing it
472        * and enabling only channel X on the SEND instruction.
473        *
474        * The above, works well for full vector writes, but not for partial writes
475        * where we want to write some channels and not others, like when we have
476        * code such as v.xyw = vec3(1,2,4). Because the untyped write messages are
477        * quite restrictive with regards to the channel enables we can configure in
478        * the message descriptor (not all combinations are allowed) we cannot simply
479        * implement these scenarios with a single message while keeping the
480        * aforementioned symmetry in the implementation. For now we de decided that
481        * it is better to keep the symmetry to reduce complexity, so in situations
482        * such as the one described we end up emitting two untyped write messages
483        * (one for xy and another for w).
484        *
485        * The code below packs consecutive channels into a single write message,
486        * detects gaps in the vector write and if needed, sends a second message
487        * with the remaining channels. If in the future we decide that we want to
488        * emit a single message at the expense of losing the symmetry in the
489        * implementation we can:
490        *
491        * 1) For IvyBridge: Only use the red channel of the untyped write SIMD8
492        *    message payload. In this mode we can write up to 8 offsets and dwords
493        *    to the red channel only (for the two vec4s in the SIMD4x2 execution)
494        *    and select which of the 8 channels carry data to write by setting the
495        *    appropriate writemask in the dst register of the SEND instruction.
496        *    It would require to write a new generator opcode specifically for
497        *    IvyBridge since we would need to prepare a SIMD8 payload that could
498        *    use any channel, not just X.
499        *
500        * 2) For Haswell+: Simply send a single write message but set the writemask
501        *    on the dst of the SEND instruction to select the channels we want to
502        *    write. It would require to modify the current messages to receive
503        *    and honor the writemask provided.
504        */
505       const vec4_builder bld = vec4_builder(this).at_end()
506                                .annotate(current_annotation, base_ir);
507 
508       emit_untyped_write(bld, surf_index, offset_reg, val_reg,
509                          1 /* dims */, instr->num_components /* size */,
510                          BRW_PREDICATE_NONE);
511       break;
512    }
513 
514    case nir_intrinsic_load_ssbo: {
515       assert(devinfo->ver == 7);
516 
517       /* brw_nir_lower_mem_access_bit_sizes takes care of this */
518       assert(nir_dest_bit_size(instr->dest) == 32);
519 
520       src_reg surf_index = get_nir_ssbo_intrinsic_index(instr);
521       src_reg offset_reg = retype(get_nir_src_imm(instr->src[1]),
522                                   BRW_REGISTER_TYPE_UD);
523 
524       /* Read the vector */
525       const vec4_builder bld = vec4_builder(this).at_end()
526          .annotate(current_annotation, base_ir);
527 
528       src_reg read_result = emit_untyped_read(bld, surf_index, offset_reg,
529                                               1 /* dims */, 4 /* size*/,
530                                               BRW_PREDICATE_NONE);
531       dst_reg dest = get_nir_dest(instr->dest);
532       read_result.type = dest.type;
533       read_result.swizzle = brw_swizzle_for_size(instr->num_components);
534       emit(MOV(dest, read_result));
535       break;
536    }
537 
538    case nir_intrinsic_ssbo_atomic_add:
539    case nir_intrinsic_ssbo_atomic_imin:
540    case nir_intrinsic_ssbo_atomic_umin:
541    case nir_intrinsic_ssbo_atomic_imax:
542    case nir_intrinsic_ssbo_atomic_umax:
543    case nir_intrinsic_ssbo_atomic_and:
544    case nir_intrinsic_ssbo_atomic_or:
545    case nir_intrinsic_ssbo_atomic_xor:
546    case nir_intrinsic_ssbo_atomic_exchange:
547    case nir_intrinsic_ssbo_atomic_comp_swap:
548       nir_emit_ssbo_atomic(brw_aop_for_nir_intrinsic(instr), instr);
549       break;
550 
551    case nir_intrinsic_load_vertex_id:
552       unreachable("should be lowered by lower_vertex_id()");
553 
554    case nir_intrinsic_load_vertex_id_zero_base:
555    case nir_intrinsic_load_base_vertex:
556    case nir_intrinsic_load_instance_id:
557    case nir_intrinsic_load_base_instance:
558    case nir_intrinsic_load_draw_id:
559    case nir_intrinsic_load_invocation_id:
560       unreachable("should be lowered by brw_nir_lower_vs_inputs()");
561 
562    case nir_intrinsic_load_uniform: {
563       /* Offsets are in bytes but they should always be multiples of 4 */
564       assert(nir_intrinsic_base(instr) % 4 == 0);
565 
566       dest = get_nir_dest(instr->dest);
567 
568       src = src_reg(dst_reg(UNIFORM, nir_intrinsic_base(instr) / 16));
569       src.type = dest.type;
570 
571       /* Uniforms don't actually have to be vec4 aligned.  In the case that
572        * it isn't, we have to use a swizzle to shift things around.  They
573        * do still have the std140 alignment requirement that vec2's have to
574        * be vec2-aligned and vec3's and vec4's have to be vec4-aligned.
575        *
576        * The swizzle also works in the indirect case as the generator adds
577        * the swizzle to the offset for us.
578        */
579       const int type_size = type_sz(src.type);
580       unsigned shift = (nir_intrinsic_base(instr) % 16) / type_size;
581       assert(shift + instr->num_components <= 4);
582 
583       if (nir_src_is_const(instr->src[0])) {
584          const unsigned load_offset = nir_src_as_uint(instr->src[0]);
585          /* Offsets are in bytes but they should always be multiples of 4 */
586          assert(load_offset % 4 == 0);
587 
588          src.swizzle = brw_swizzle_for_size(instr->num_components);
589          dest.writemask = brw_writemask_for_size(instr->num_components);
590          unsigned offset = load_offset + shift * type_size;
591          src.offset = ROUND_DOWN_TO(offset, 16);
592          shift = (offset % 16) / type_size;
593          assert(shift + instr->num_components <= 4);
594          src.swizzle += BRW_SWIZZLE4(shift, shift, shift, shift);
595 
596          emit(MOV(dest, src));
597       } else {
598          /* Uniform arrays are vec4 aligned, because of std140 alignment
599           * rules.
600           */
601          assert(shift == 0);
602 
603          src_reg indirect = get_nir_src(instr->src[0], BRW_REGISTER_TYPE_UD, 1);
604 
605          /* MOV_INDIRECT is going to stomp the whole thing anyway */
606          dest.writemask = WRITEMASK_XYZW;
607 
608          emit(SHADER_OPCODE_MOV_INDIRECT, dest, src,
609               indirect, brw_imm_ud(instr->const_index[1]));
610       }
611       break;
612    }
613 
614    case nir_intrinsic_load_ubo: {
615       src_reg surf_index;
616 
617       dest = get_nir_dest(instr->dest);
618 
619       if (nir_src_is_const(instr->src[0])) {
620          /* The block index is a constant, so just emit the binding table entry
621           * as an immediate.
622           */
623          const unsigned index = nir_src_as_uint(instr->src[0]);
624          surf_index = brw_imm_ud(index);
625       } else {
626          /* The block index is not a constant. Evaluate the index expression
627           * per-channel and add the base UBO index; we have to select a value
628           * from any live channel.
629           */
630          surf_index = src_reg(this, glsl_type::uint_type);
631          emit(MOV(dst_reg(surf_index), get_nir_src(instr->src[0], nir_type_int32,
632                                                    instr->num_components)));
633          surf_index = emit_uniformize(surf_index);
634       }
635 
636       src_reg push_reg;
637       src_reg offset_reg;
638       if (nir_src_is_const(instr->src[1])) {
639          unsigned load_offset = nir_src_as_uint(instr->src[1]);
640          unsigned aligned_offset = load_offset & ~15;
641          offset_reg = brw_imm_ud(aligned_offset);
642 
643          /* See if we've selected this as a push constant candidate */
644          if (nir_src_is_const(instr->src[0])) {
645             const unsigned ubo_block = nir_src_as_uint(instr->src[0]);
646             const unsigned offset_256b = aligned_offset / 32;
647 
648             for (int i = 0; i < 4; i++) {
649                const struct brw_ubo_range *range = &prog_data->base.ubo_ranges[i];
650                if (range->block == ubo_block &&
651                    offset_256b >= range->start &&
652                    offset_256b < range->start + range->length) {
653 
654                   push_reg = src_reg(dst_reg(UNIFORM, UBO_START + i));
655                   push_reg.type = dest.type;
656                   push_reg.offset = aligned_offset - 32 * range->start;
657                   break;
658                }
659             }
660          }
661       } else {
662          offset_reg = src_reg(this, glsl_type::uint_type);
663          emit(MOV(dst_reg(offset_reg),
664                   get_nir_src(instr->src[1], nir_type_uint32, 1)));
665       }
666 
667       src_reg packed_consts;
668       if (push_reg.file != BAD_FILE) {
669          packed_consts = push_reg;
670       } else if (nir_dest_bit_size(instr->dest) == 32) {
671          packed_consts = src_reg(this, glsl_type::vec4_type);
672          emit_pull_constant_load_reg(dst_reg(packed_consts),
673                                      surf_index,
674                                      offset_reg,
675                                      NULL, NULL /* before_block/inst */);
676          prog_data->base.has_ubo_pull = true;
677       } else {
678          src_reg temp = src_reg(this, glsl_type::dvec4_type);
679          src_reg temp_float = retype(temp, BRW_REGISTER_TYPE_F);
680 
681          emit_pull_constant_load_reg(dst_reg(temp_float),
682                                      surf_index, offset_reg, NULL, NULL);
683          if (offset_reg.file == IMM)
684             offset_reg.ud += 16;
685          else
686             emit(ADD(dst_reg(offset_reg), offset_reg, brw_imm_ud(16u)));
687          emit_pull_constant_load_reg(dst_reg(byte_offset(temp_float, REG_SIZE)),
688                                      surf_index, offset_reg, NULL, NULL);
689          prog_data->base.has_ubo_pull = true;
690 
691          packed_consts = src_reg(this, glsl_type::dvec4_type);
692          shuffle_64bit_data(dst_reg(packed_consts), temp, false);
693       }
694 
695       packed_consts.swizzle = brw_swizzle_for_size(instr->num_components);
696       if (nir_src_is_const(instr->src[1])) {
697          unsigned load_offset = nir_src_as_uint(instr->src[1]);
698          unsigned type_size = type_sz(dest.type);
699          packed_consts.swizzle +=
700             BRW_SWIZZLE4(load_offset % 16 / type_size,
701                          load_offset % 16 / type_size,
702                          load_offset % 16 / type_size,
703                          load_offset % 16 / type_size);
704       }
705 
706       emit(MOV(dest, retype(packed_consts, dest.type)));
707 
708       break;
709    }
710 
711    case nir_intrinsic_scoped_barrier:
712       assert(nir_intrinsic_execution_scope(instr) == NIR_SCOPE_NONE);
713       FALLTHROUGH;
714    case nir_intrinsic_memory_barrier: {
715       const vec4_builder bld =
716          vec4_builder(this).at_end().annotate(current_annotation, base_ir);
717       const dst_reg tmp = bld.vgrf(BRW_REGISTER_TYPE_UD);
718       vec4_instruction *fence =
719          bld.emit(SHADER_OPCODE_MEMORY_FENCE, tmp, brw_vec8_grf(0, 0));
720       fence->sfid = GFX7_SFID_DATAPORT_DATA_CACHE;
721       break;
722    }
723 
724    case nir_intrinsic_shader_clock: {
725       /* We cannot do anything if there is an event, so ignore it for now */
726       const src_reg shader_clock = get_timestamp();
727       const enum brw_reg_type type = brw_type_for_base_type(glsl_type::uvec2_type);
728 
729       dest = get_nir_dest(instr->dest, type);
730       emit(MOV(dest, shader_clock));
731       break;
732    }
733 
734    default:
735       unreachable("Unknown intrinsic");
736    }
737 }
738 
739 void
nir_emit_ssbo_atomic(int op,nir_intrinsic_instr * instr)740 vec4_visitor::nir_emit_ssbo_atomic(int op, nir_intrinsic_instr *instr)
741 {
742    dst_reg dest;
743    if (nir_intrinsic_infos[instr->intrinsic].has_dest)
744       dest = get_nir_dest(instr->dest);
745 
746    src_reg surface = get_nir_ssbo_intrinsic_index(instr);
747    src_reg offset = get_nir_src(instr->src[1], 1);
748    src_reg data1;
749    if (op != BRW_AOP_INC && op != BRW_AOP_DEC && op != BRW_AOP_PREDEC)
750       data1 = get_nir_src(instr->src[2], 1);
751    src_reg data2;
752    if (op == BRW_AOP_CMPWR)
753       data2 = get_nir_src(instr->src[3], 1);
754 
755    /* Emit the actual atomic operation operation */
756    const vec4_builder bld =
757       vec4_builder(this).at_end().annotate(current_annotation, base_ir);
758 
759    src_reg atomic_result = emit_untyped_atomic(bld, surface, offset,
760                                                data1, data2,
761                                                1 /* dims */, 1 /* rsize */,
762                                                op,
763                                                BRW_PREDICATE_NONE);
764    dest.type = atomic_result.type;
765    bld.MOV(dest, atomic_result);
766 }
767 
768 static unsigned
brw_swizzle_for_nir_swizzle(uint8_t swizzle[4])769 brw_swizzle_for_nir_swizzle(uint8_t swizzle[4])
770 {
771    return BRW_SWIZZLE4(swizzle[0], swizzle[1], swizzle[2], swizzle[3]);
772 }
773 
774 bool
optimize_predicate(nir_alu_instr * instr,enum brw_predicate * predicate)775 vec4_visitor::optimize_predicate(nir_alu_instr *instr,
776                                  enum brw_predicate *predicate)
777 {
778    if (!instr->src[0].src.is_ssa ||
779        instr->src[0].src.ssa->parent_instr->type != nir_instr_type_alu)
780       return false;
781 
782    nir_alu_instr *cmp_instr =
783       nir_instr_as_alu(instr->src[0].src.ssa->parent_instr);
784 
785    switch (cmp_instr->op) {
786    case nir_op_b32any_fnequal2:
787    case nir_op_b32any_inequal2:
788    case nir_op_b32any_fnequal3:
789    case nir_op_b32any_inequal3:
790    case nir_op_b32any_fnequal4:
791    case nir_op_b32any_inequal4:
792       *predicate = BRW_PREDICATE_ALIGN16_ANY4H;
793       break;
794    case nir_op_b32all_fequal2:
795    case nir_op_b32all_iequal2:
796    case nir_op_b32all_fequal3:
797    case nir_op_b32all_iequal3:
798    case nir_op_b32all_fequal4:
799    case nir_op_b32all_iequal4:
800       *predicate = BRW_PREDICATE_ALIGN16_ALL4H;
801       break;
802    default:
803       return false;
804    }
805 
806    unsigned size_swizzle =
807       brw_swizzle_for_size(nir_op_infos[cmp_instr->op].input_sizes[0]);
808 
809    src_reg op[2];
810    assert(nir_op_infos[cmp_instr->op].num_inputs == 2);
811    for (unsigned i = 0; i < 2; i++) {
812       nir_alu_type type = nir_op_infos[cmp_instr->op].input_types[i];
813       unsigned bit_size = nir_src_bit_size(cmp_instr->src[i].src);
814       type = (nir_alu_type) (((unsigned) type) | bit_size);
815       op[i] = get_nir_src(cmp_instr->src[i].src, type, 4);
816       unsigned base_swizzle =
817          brw_swizzle_for_nir_swizzle(cmp_instr->src[i].swizzle);
818       op[i].swizzle = brw_compose_swizzle(size_swizzle, base_swizzle);
819    }
820 
821    emit(CMP(dst_null_d(), op[0], op[1],
822             brw_cmod_for_nir_comparison(cmp_instr->op)));
823 
824    return true;
825 }
826 
827 static void
emit_find_msb_using_lzd(const vec4_builder & bld,const dst_reg & dst,const src_reg & src,bool is_signed)828 emit_find_msb_using_lzd(const vec4_builder &bld,
829                         const dst_reg &dst,
830                         const src_reg &src,
831                         bool is_signed)
832 {
833    vec4_instruction *inst;
834    src_reg temp = src;
835 
836    if (is_signed) {
837       /* LZD of an absolute value source almost always does the right
838        * thing.  There are two problem values:
839        *
840        * * 0x80000000.  Since abs(0x80000000) == 0x80000000, LZD returns
841        *   0.  However, findMSB(int(0x80000000)) == 30.
842        *
843        * * 0xffffffff.  Since abs(0xffffffff) == 1, LZD returns
844        *   31.  Section 8.8 (Integer Functions) of the GLSL 4.50 spec says:
845        *
846        *    For a value of zero or negative one, -1 will be returned.
847        *
848        * * Negative powers of two.  LZD(abs(-(1<<x))) returns x, but
849        *   findMSB(-(1<<x)) should return x-1.
850        *
851        * For all negative number cases, including 0x80000000 and
852        * 0xffffffff, the correct value is obtained from LZD if instead of
853        * negating the (already negative) value the logical-not is used.  A
854        * conditonal logical-not can be achieved in two instructions.
855        */
856       temp = src_reg(bld.vgrf(BRW_REGISTER_TYPE_D));
857 
858       bld.ASR(dst_reg(temp), src, brw_imm_d(31));
859       bld.XOR(dst_reg(temp), temp, src);
860    }
861 
862    bld.LZD(retype(dst, BRW_REGISTER_TYPE_UD),
863            retype(temp, BRW_REGISTER_TYPE_UD));
864 
865    /* LZD counts from the MSB side, while GLSL's findMSB() wants the count
866     * from the LSB side. Subtract the result from 31 to convert the MSB count
867     * into an LSB count.  If no bits are set, LZD will return 32.  31-32 = -1,
868     * which is exactly what findMSB() is supposed to return.
869     */
870    inst = bld.ADD(dst, retype(src_reg(dst), BRW_REGISTER_TYPE_D),
871                   brw_imm_d(31));
872    inst->src[0].negate = true;
873 }
874 
875 void
emit_conversion_from_double(dst_reg dst,src_reg src)876 vec4_visitor::emit_conversion_from_double(dst_reg dst, src_reg src)
877 {
878    enum opcode op;
879    switch (dst.type) {
880    case BRW_REGISTER_TYPE_D:
881       op = VEC4_OPCODE_DOUBLE_TO_D32;
882       break;
883    case BRW_REGISTER_TYPE_UD:
884       op = VEC4_OPCODE_DOUBLE_TO_U32;
885       break;
886    case BRW_REGISTER_TYPE_F:
887       op = VEC4_OPCODE_DOUBLE_TO_F32;
888       break;
889    default:
890       unreachable("Unknown conversion");
891    }
892 
893    dst_reg temp = dst_reg(this, glsl_type::dvec4_type);
894    emit(MOV(temp, src));
895    dst_reg temp2 = dst_reg(this, glsl_type::dvec4_type);
896    emit(op, temp2, src_reg(temp));
897 
898    emit(VEC4_OPCODE_PICK_LOW_32BIT, retype(temp2, dst.type), src_reg(temp2));
899    emit(MOV(dst, src_reg(retype(temp2, dst.type))));
900 }
901 
902 void
emit_conversion_to_double(dst_reg dst,src_reg src)903 vec4_visitor::emit_conversion_to_double(dst_reg dst, src_reg src)
904 {
905    dst_reg tmp_dst = dst_reg(src_reg(this, glsl_type::dvec4_type));
906    src_reg tmp_src = retype(src_reg(this, glsl_type::vec4_type), src.type);
907    emit(MOV(dst_reg(tmp_src), src));
908    emit(VEC4_OPCODE_TO_DOUBLE, tmp_dst, tmp_src);
909    emit(MOV(dst, src_reg(tmp_dst)));
910 }
911 
912 /**
913  * Try to use an immediate value for a source
914  *
915  * In cases of flow control, constant propagation is sometimes unable to
916  * determine that a register contains a constant value.  To work around this,
917  * try to emit a literal as one of the sources.  If \c try_src0_also is set,
918  * \c op[0] will also be tried for an immediate value.
919  *
920  * If \c op[0] is modified, the operands will be exchanged so that \c op[1]
921  * will always be the immediate value.
922  *
923  * \return The index of the source that was modified, 0 or 1, if successful.
924  * Otherwise, -1.
925  *
926  * \param op - Operands to the instruction
927  * \param try_src0_also - True if \c op[0] should also be a candidate for
928  *                        getting an immediate value.  This should only be set
929  *                        for commutative operations.
930  */
931 static int
try_immediate_source(const nir_alu_instr * instr,src_reg * op,bool try_src0_also)932 try_immediate_source(const nir_alu_instr *instr, src_reg *op,
933                      bool try_src0_also)
934 {
935    unsigned idx;
936 
937    /* MOV should be the only single-source instruction passed to this
938     * function.  Any other unary instruction with a constant source should
939     * have been constant-folded away!
940     */
941    assert(nir_op_infos[instr->op].num_inputs > 1 ||
942           instr->op == nir_op_mov);
943 
944    if (instr->op != nir_op_mov &&
945        nir_src_bit_size(instr->src[1].src) == 32 &&
946        nir_src_is_const(instr->src[1].src)) {
947       idx = 1;
948    } else if (try_src0_also &&
949          nir_src_bit_size(instr->src[0].src) == 32 &&
950          nir_src_is_const(instr->src[0].src)) {
951       idx = 0;
952    } else {
953       return -1;
954    }
955 
956    const enum brw_reg_type old_type = op[idx].type;
957 
958    switch (old_type) {
959    case BRW_REGISTER_TYPE_D:
960    case BRW_REGISTER_TYPE_UD: {
961       int first_comp = -1;
962       int d = 0;
963 
964       for (unsigned i = 0; i < NIR_MAX_VEC_COMPONENTS; i++) {
965          if (nir_alu_instr_channel_used(instr, idx, i)) {
966             if (first_comp < 0) {
967                first_comp = i;
968                d = nir_src_comp_as_int(instr->src[idx].src,
969                                        instr->src[idx].swizzle[i]);
970             } else if (d != nir_src_comp_as_int(instr->src[idx].src,
971                                                 instr->src[idx].swizzle[i])) {
972                return -1;
973             }
974          }
975       }
976 
977       assert(first_comp >= 0);
978 
979       if (op[idx].abs)
980          d = MAX2(-d, d);
981 
982       if (op[idx].negate)
983          d = -d;
984 
985       op[idx] = retype(src_reg(brw_imm_d(d)), old_type);
986       break;
987    }
988 
989    case BRW_REGISTER_TYPE_F: {
990       int first_comp = -1;
991       float f[NIR_MAX_VEC_COMPONENTS] = { 0.0f };
992       bool is_scalar = true;
993 
994       for (unsigned i = 0; i < NIR_MAX_VEC_COMPONENTS; i++) {
995          if (nir_alu_instr_channel_used(instr, idx, i)) {
996             f[i] = nir_src_comp_as_float(instr->src[idx].src,
997                                          instr->src[idx].swizzle[i]);
998             if (first_comp < 0) {
999                first_comp = i;
1000             } else if (f[first_comp] != f[i]) {
1001                is_scalar = false;
1002             }
1003          }
1004       }
1005 
1006       if (is_scalar) {
1007          if (op[idx].abs)
1008             f[first_comp] = fabs(f[first_comp]);
1009 
1010          if (op[idx].negate)
1011             f[first_comp] = -f[first_comp];
1012 
1013          op[idx] = src_reg(brw_imm_f(f[first_comp]));
1014          assert(op[idx].type == old_type);
1015       } else {
1016          uint8_t vf_values[4] = { 0, 0, 0, 0 };
1017 
1018          for (unsigned i = 0; i < ARRAY_SIZE(vf_values); i++) {
1019 
1020             if (op[idx].abs)
1021                f[i] = fabs(f[i]);
1022 
1023             if (op[idx].negate)
1024                f[i] = -f[i];
1025 
1026             const int vf = brw_float_to_vf(f[i]);
1027             if (vf == -1)
1028                return -1;
1029 
1030             vf_values[i] = vf;
1031          }
1032 
1033          op[idx] = src_reg(brw_imm_vf4(vf_values[0], vf_values[1],
1034                                        vf_values[2], vf_values[3]));
1035       }
1036       break;
1037    }
1038 
1039    default:
1040       unreachable("Non-32bit type.");
1041    }
1042 
1043    /* If the instruction has more than one source, the instruction format only
1044     * allows source 1 to be an immediate value.  If the immediate value was
1045     * source 0, then the sources must be exchanged.
1046     */
1047    if (idx == 0 && instr->op != nir_op_mov) {
1048       src_reg tmp = op[0];
1049       op[0] = op[1];
1050       op[1] = tmp;
1051    }
1052 
1053    return idx;
1054 }
1055 
1056 void
fix_float_operands(src_reg op[3],nir_alu_instr * instr)1057 vec4_visitor::fix_float_operands(src_reg op[3], nir_alu_instr *instr)
1058 {
1059    bool fixed[3] = { false, false, false };
1060 
1061    for (unsigned i = 0; i < 2; i++) {
1062       if (!nir_src_is_const(instr->src[i].src))
1063          continue;
1064 
1065       for (unsigned j = i + 1; j < 3; j++) {
1066          if (fixed[j])
1067             continue;
1068 
1069          if (!nir_src_is_const(instr->src[j].src))
1070             continue;
1071 
1072          if (nir_alu_srcs_equal(instr, instr, i, j)) {
1073             if (!fixed[i])
1074                op[i] = fix_3src_operand(op[i]);
1075 
1076             op[j] = op[i];
1077 
1078             fixed[i] = true;
1079             fixed[j] = true;
1080          } else if (nir_alu_srcs_negative_equal(instr, instr, i, j)) {
1081             if (!fixed[i])
1082                op[i] = fix_3src_operand(op[i]);
1083 
1084             op[j] = op[i];
1085             op[j].negate = !op[j].negate;
1086 
1087             fixed[i] = true;
1088             fixed[j] = true;
1089          }
1090       }
1091    }
1092 
1093    for (unsigned i = 0; i < 3; i++) {
1094       if (!fixed[i])
1095          op[i] = fix_3src_operand(op[i]);
1096    }
1097 }
1098 
1099 static bool
const_src_fits_in_16_bits(const nir_src & src,brw_reg_type type)1100 const_src_fits_in_16_bits(const nir_src &src, brw_reg_type type)
1101 {
1102    assert(nir_src_is_const(src));
1103    if (brw_reg_type_is_unsigned_integer(type)) {
1104       return nir_src_comp_as_uint(src, 0) <= UINT16_MAX;
1105    } else {
1106       const int64_t c = nir_src_comp_as_int(src, 0);
1107       return c <= INT16_MAX && c >= INT16_MIN;
1108    }
1109 }
1110 
1111 void
nir_emit_alu(nir_alu_instr * instr)1112 vec4_visitor::nir_emit_alu(nir_alu_instr *instr)
1113 {
1114    vec4_instruction *inst;
1115 
1116    nir_alu_type dst_type = (nir_alu_type) (nir_op_infos[instr->op].output_type |
1117                                            nir_dest_bit_size(instr->dest.dest));
1118    dst_reg dst = get_nir_dest(instr->dest.dest, dst_type);
1119    dst.writemask = instr->dest.write_mask;
1120 
1121    assert(!instr->dest.saturate);
1122 
1123    src_reg op[4];
1124    for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
1125       /* We don't lower to source modifiers, so they shouldn't exist. */
1126       assert(!instr->src[i].abs);
1127       assert(!instr->src[i].negate);
1128 
1129       nir_alu_type src_type = (nir_alu_type)
1130          (nir_op_infos[instr->op].input_types[i] |
1131           nir_src_bit_size(instr->src[i].src));
1132       op[i] = get_nir_src(instr->src[i].src, src_type, 4);
1133       op[i].swizzle = brw_swizzle_for_nir_swizzle(instr->src[i].swizzle);
1134    }
1135 
1136 #ifndef NDEBUG
1137    /* On Gen7 and earlier, no functionality is exposed that should allow 8-bit
1138     * integer types to ever exist.
1139     */
1140    for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
1141       assert(type_sz(op[i].type) > 1);
1142 #endif
1143 
1144    switch (instr->op) {
1145    case nir_op_mov:
1146       try_immediate_source(instr, &op[0], true);
1147       inst = emit(MOV(dst, op[0]));
1148       break;
1149 
1150    case nir_op_vec2:
1151    case nir_op_vec3:
1152    case nir_op_vec4:
1153       unreachable("not reached: should be handled by lower_vec_to_movs()");
1154 
1155    case nir_op_i2f32:
1156    case nir_op_u2f32:
1157       inst = emit(MOV(dst, op[0]));
1158       break;
1159 
1160    case nir_op_f2f32:
1161    case nir_op_f2i32:
1162    case nir_op_f2u32:
1163       if (nir_src_bit_size(instr->src[0].src) == 64)
1164          emit_conversion_from_double(dst, op[0]);
1165       else
1166          inst = emit(MOV(dst, op[0]));
1167       break;
1168 
1169    case nir_op_f2f64:
1170    case nir_op_i2f64:
1171    case nir_op_u2f64:
1172       emit_conversion_to_double(dst, op[0]);
1173       break;
1174 
1175    case nir_op_fsat:
1176       inst = emit(MOV(dst, op[0]));
1177       inst->saturate = true;
1178       break;
1179 
1180    case nir_op_fneg:
1181    case nir_op_ineg:
1182       op[0].negate = true;
1183       inst = emit(MOV(dst, op[0]));
1184       break;
1185 
1186    case nir_op_fabs:
1187    case nir_op_iabs:
1188       op[0].negate = false;
1189       op[0].abs = true;
1190       inst = emit(MOV(dst, op[0]));
1191       break;
1192 
1193    case nir_op_iadd:
1194       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1195       FALLTHROUGH;
1196    case nir_op_fadd:
1197       try_immediate_source(instr, op, true);
1198       inst = emit(ADD(dst, op[0], op[1]));
1199       break;
1200 
1201    case nir_op_uadd_sat:
1202       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1203       inst = emit(ADD(dst, op[0], op[1]));
1204       inst->saturate = true;
1205       break;
1206 
1207    case nir_op_fmul:
1208       try_immediate_source(instr, op, true);
1209       inst = emit(MUL(dst, op[0], op[1]));
1210       break;
1211 
1212    case nir_op_imul: {
1213       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1214 
1215       /* For integer multiplication, the MUL uses the low 16 bits of one of
1216        * the operands (src0 through SNB, src1 on IVB and later). The MACH
1217        * accumulates in the contribution of the upper 16 bits of that
1218        * operand. If we can determine that one of the args is in the low
1219        * 16 bits, though, we can just emit a single MUL.
1220        */
1221       if (nir_src_is_const(instr->src[0].src) &&
1222           nir_alu_instr_src_read_mask(instr, 0) == 1 &&
1223           const_src_fits_in_16_bits(instr->src[0].src, op[0].type)) {
1224          if (devinfo->ver < 7)
1225             emit(MUL(dst, op[0], op[1]));
1226          else
1227             emit(MUL(dst, op[1], op[0]));
1228       } else if (nir_src_is_const(instr->src[1].src) &&
1229                  nir_alu_instr_src_read_mask(instr, 1) == 1 &&
1230                  const_src_fits_in_16_bits(instr->src[1].src, op[1].type)) {
1231          if (devinfo->ver < 7)
1232             emit(MUL(dst, op[1], op[0]));
1233          else
1234             emit(MUL(dst, op[0], op[1]));
1235       } else {
1236          struct brw_reg acc = retype(brw_acc_reg(8), dst.type);
1237 
1238          emit(MUL(acc, op[0], op[1]));
1239          emit(MACH(dst_null_d(), op[0], op[1]));
1240          emit(MOV(dst, src_reg(acc)));
1241       }
1242       break;
1243    }
1244 
1245    case nir_op_imul_high:
1246    case nir_op_umul_high: {
1247       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1248       struct brw_reg acc = retype(brw_acc_reg(8), dst.type);
1249 
1250       emit(MUL(acc, op[0], op[1]));
1251       emit(MACH(dst, op[0], op[1]));
1252       break;
1253    }
1254 
1255    case nir_op_frcp:
1256       inst = emit_math(SHADER_OPCODE_RCP, dst, op[0]);
1257       break;
1258 
1259    case nir_op_fexp2:
1260       inst = emit_math(SHADER_OPCODE_EXP2, dst, op[0]);
1261       break;
1262 
1263    case nir_op_flog2:
1264       inst = emit_math(SHADER_OPCODE_LOG2, dst, op[0]);
1265       break;
1266 
1267    case nir_op_fsin:
1268       inst = emit_math(SHADER_OPCODE_SIN, dst, op[0]);
1269       break;
1270 
1271    case nir_op_fcos:
1272       inst = emit_math(SHADER_OPCODE_COS, dst, op[0]);
1273       break;
1274 
1275    case nir_op_idiv:
1276    case nir_op_udiv:
1277       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1278       emit_math(SHADER_OPCODE_INT_QUOTIENT, dst, op[0], op[1]);
1279       break;
1280 
1281    case nir_op_umod:
1282    case nir_op_irem:
1283       /* According to the sign table for INT DIV in the Ivy Bridge PRM, it
1284        * appears that our hardware just does the right thing for signed
1285        * remainder.
1286        */
1287       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1288       emit_math(SHADER_OPCODE_INT_REMAINDER, dst, op[0], op[1]);
1289       break;
1290 
1291    case nir_op_imod: {
1292       /* Get a regular C-style remainder.  If a % b == 0, set the predicate. */
1293       inst = emit_math(SHADER_OPCODE_INT_REMAINDER, dst, op[0], op[1]);
1294 
1295       /* Math instructions don't support conditional mod */
1296       inst = emit(MOV(dst_null_d(), src_reg(dst)));
1297       inst->conditional_mod = BRW_CONDITIONAL_NZ;
1298 
1299       /* Now, we need to determine if signs of the sources are different.
1300        * When we XOR the sources, the top bit is 0 if they are the same and 1
1301        * if they are different.  We can then use a conditional modifier to
1302        * turn that into a predicate.  This leads us to an XOR.l instruction.
1303        *
1304        * Technically, according to the PRM, you're not allowed to use .l on a
1305        * XOR instruction.  However, emperical experiments and Curro's reading
1306        * of the simulator source both indicate that it's safe.
1307        */
1308       src_reg tmp = src_reg(this, glsl_type::ivec4_type);
1309       inst = emit(XOR(dst_reg(tmp), op[0], op[1]));
1310       inst->predicate = BRW_PREDICATE_NORMAL;
1311       inst->conditional_mod = BRW_CONDITIONAL_L;
1312 
1313       /* If the result of the initial remainder operation is non-zero and the
1314        * two sources have different signs, add in a copy of op[1] to get the
1315        * final integer modulus value.
1316        */
1317       inst = emit(ADD(dst, src_reg(dst), op[1]));
1318       inst->predicate = BRW_PREDICATE_NORMAL;
1319       break;
1320    }
1321 
1322    case nir_op_ldexp:
1323       unreachable("not reached: should be handled by ldexp_to_arith()");
1324 
1325    case nir_op_fsqrt:
1326       inst = emit_math(SHADER_OPCODE_SQRT, dst, op[0]);
1327       break;
1328 
1329    case nir_op_frsq:
1330       inst = emit_math(SHADER_OPCODE_RSQ, dst, op[0]);
1331       break;
1332 
1333    case nir_op_fpow:
1334       inst = emit_math(SHADER_OPCODE_POW, dst, op[0], op[1]);
1335       break;
1336 
1337    case nir_op_uadd_carry: {
1338       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1339       struct brw_reg acc = retype(brw_acc_reg(8), BRW_REGISTER_TYPE_UD);
1340 
1341       emit(ADDC(dst_null_ud(), op[0], op[1]));
1342       emit(MOV(dst, src_reg(acc)));
1343       break;
1344    }
1345 
1346    case nir_op_usub_borrow: {
1347       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1348       struct brw_reg acc = retype(brw_acc_reg(8), BRW_REGISTER_TYPE_UD);
1349 
1350       emit(SUBB(dst_null_ud(), op[0], op[1]));
1351       emit(MOV(dst, src_reg(acc)));
1352       break;
1353    }
1354 
1355    case nir_op_ftrunc:
1356       inst = emit(RNDZ(dst, op[0]));
1357       if (devinfo->ver < 6) {
1358          inst->conditional_mod = BRW_CONDITIONAL_R;
1359          inst = emit(ADD(dst, src_reg(dst), brw_imm_f(1.0f)));
1360          inst->predicate = BRW_PREDICATE_NORMAL;
1361          inst = emit(MOV(dst, src_reg(dst))); /* for potential saturation */
1362       }
1363       break;
1364 
1365    case nir_op_fceil: {
1366       src_reg tmp = src_reg(this, glsl_type::float_type);
1367       tmp.swizzle =
1368          brw_swizzle_for_size(instr->src[0].src.is_ssa ?
1369                               instr->src[0].src.ssa->num_components :
1370                               instr->src[0].src.reg.reg->num_components);
1371 
1372       op[0].negate = !op[0].negate;
1373       emit(RNDD(dst_reg(tmp), op[0]));
1374       tmp.negate = true;
1375       inst = emit(MOV(dst, tmp));
1376       break;
1377    }
1378 
1379    case nir_op_ffloor:
1380       inst = emit(RNDD(dst, op[0]));
1381       break;
1382 
1383    case nir_op_ffract:
1384       inst = emit(FRC(dst, op[0]));
1385       break;
1386 
1387    case nir_op_fround_even:
1388       inst = emit(RNDE(dst, op[0]));
1389       if (devinfo->ver < 6) {
1390          inst->conditional_mod = BRW_CONDITIONAL_R;
1391          inst = emit(ADD(dst, src_reg(dst), brw_imm_f(1.0f)));
1392          inst->predicate = BRW_PREDICATE_NORMAL;
1393          inst = emit(MOV(dst, src_reg(dst))); /* for potential saturation */
1394       }
1395       break;
1396 
1397    case nir_op_fquantize2f16: {
1398       /* See also vec4_visitor::emit_pack_half_2x16() */
1399       src_reg tmp16 = src_reg(this, glsl_type::uvec4_type);
1400       src_reg tmp32 = src_reg(this, glsl_type::vec4_type);
1401       src_reg zero = src_reg(this, glsl_type::vec4_type);
1402 
1403       /* Check for denormal */
1404       src_reg abs_src0 = op[0];
1405       abs_src0.abs = true;
1406       emit(CMP(dst_null_f(), abs_src0, brw_imm_f(ldexpf(1.0, -14)),
1407                BRW_CONDITIONAL_L));
1408       /* Get the appropriately signed zero */
1409       emit(AND(retype(dst_reg(zero), BRW_REGISTER_TYPE_UD),
1410                retype(op[0], BRW_REGISTER_TYPE_UD),
1411                brw_imm_ud(0x80000000)));
1412       /* Do the actual F32 -> F16 -> F32 conversion */
1413       emit(F32TO16(dst_reg(tmp16), op[0]));
1414       emit(F16TO32(dst_reg(tmp32), tmp16));
1415       /* Select that or zero based on normal status */
1416       inst = emit(BRW_OPCODE_SEL, dst, zero, tmp32);
1417       inst->predicate = BRW_PREDICATE_NORMAL;
1418       break;
1419    }
1420 
1421    case nir_op_imin:
1422    case nir_op_umin:
1423       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1424       FALLTHROUGH;
1425    case nir_op_fmin:
1426       try_immediate_source(instr, op, true);
1427       inst = emit_minmax(BRW_CONDITIONAL_L, dst, op[0], op[1]);
1428       break;
1429 
1430    case nir_op_imax:
1431    case nir_op_umax:
1432       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1433       FALLTHROUGH;
1434    case nir_op_fmax:
1435       try_immediate_source(instr, op, true);
1436       inst = emit_minmax(BRW_CONDITIONAL_GE, dst, op[0], op[1]);
1437       break;
1438 
1439    case nir_op_fddx:
1440    case nir_op_fddx_coarse:
1441    case nir_op_fddx_fine:
1442    case nir_op_fddy:
1443    case nir_op_fddy_coarse:
1444    case nir_op_fddy_fine:
1445       unreachable("derivatives are not valid in vertex shaders");
1446 
1447    case nir_op_ilt32:
1448    case nir_op_ult32:
1449    case nir_op_ige32:
1450    case nir_op_uge32:
1451    case nir_op_ieq32:
1452    case nir_op_ine32:
1453       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1454       FALLTHROUGH;
1455    case nir_op_flt32:
1456    case nir_op_fge32:
1457    case nir_op_feq32:
1458    case nir_op_fneu32: {
1459       enum brw_conditional_mod conditional_mod =
1460          brw_cmod_for_nir_comparison(instr->op);
1461 
1462       if (nir_src_bit_size(instr->src[0].src) < 64) {
1463          /* If the order of the sources is changed due to an immediate value,
1464           * then the condition must also be changed.
1465           */
1466          if (try_immediate_source(instr, op, true) == 0)
1467             conditional_mod = brw_swap_cmod(conditional_mod);
1468 
1469          emit(CMP(dst, op[0], op[1], conditional_mod));
1470       } else {
1471          /* Produce a 32-bit boolean result from the DF comparison by selecting
1472           * only the low 32-bit in each DF produced. Do this in a temporary
1473           * so we can then move from there to the result using align16 again
1474           * to honor the original writemask.
1475           */
1476          dst_reg temp = dst_reg(this, glsl_type::dvec4_type);
1477          emit(CMP(temp, op[0], op[1], conditional_mod));
1478          dst_reg result = dst_reg(this, glsl_type::bvec4_type);
1479          emit(VEC4_OPCODE_PICK_LOW_32BIT, result, src_reg(temp));
1480          emit(MOV(dst, src_reg(result)));
1481       }
1482       break;
1483    }
1484 
1485    case nir_op_b32all_iequal2:
1486    case nir_op_b32all_iequal3:
1487    case nir_op_b32all_iequal4:
1488       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1489       FALLTHROUGH;
1490    case nir_op_b32all_fequal2:
1491    case nir_op_b32all_fequal3:
1492    case nir_op_b32all_fequal4: {
1493       unsigned swiz =
1494          brw_swizzle_for_size(nir_op_infos[instr->op].input_sizes[0]);
1495 
1496       emit(CMP(dst_null_d(), swizzle(op[0], swiz), swizzle(op[1], swiz),
1497                brw_cmod_for_nir_comparison(instr->op)));
1498       emit(MOV(dst, brw_imm_d(0)));
1499       inst = emit(MOV(dst, brw_imm_d(~0)));
1500       inst->predicate = BRW_PREDICATE_ALIGN16_ALL4H;
1501       break;
1502    }
1503 
1504    case nir_op_b32any_inequal2:
1505    case nir_op_b32any_inequal3:
1506    case nir_op_b32any_inequal4:
1507       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1508       FALLTHROUGH;
1509    case nir_op_b32any_fnequal2:
1510    case nir_op_b32any_fnequal3:
1511    case nir_op_b32any_fnequal4: {
1512       unsigned swiz =
1513          brw_swizzle_for_size(nir_op_infos[instr->op].input_sizes[0]);
1514 
1515       emit(CMP(dst_null_d(), swizzle(op[0], swiz), swizzle(op[1], swiz),
1516                brw_cmod_for_nir_comparison(instr->op)));
1517 
1518       emit(MOV(dst, brw_imm_d(0)));
1519       inst = emit(MOV(dst, brw_imm_d(~0)));
1520       inst->predicate = BRW_PREDICATE_ALIGN16_ANY4H;
1521       break;
1522    }
1523 
1524    case nir_op_inot:
1525       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1526       emit(NOT(dst, op[0]));
1527       break;
1528 
1529    case nir_op_ixor:
1530       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1531       try_immediate_source(instr, op, true);
1532       emit(XOR(dst, op[0], op[1]));
1533       break;
1534 
1535    case nir_op_ior:
1536       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1537       try_immediate_source(instr, op, true);
1538       emit(OR(dst, op[0], op[1]));
1539       break;
1540 
1541    case nir_op_iand:
1542       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1543       try_immediate_source(instr, op, true);
1544       emit(AND(dst, op[0], op[1]));
1545       break;
1546 
1547    case nir_op_b2i32:
1548    case nir_op_b2f32:
1549    case nir_op_b2f64:
1550       if (nir_dest_bit_size(instr->dest.dest) > 32) {
1551          assert(dst.type == BRW_REGISTER_TYPE_DF);
1552          emit_conversion_to_double(dst, negate(op[0]));
1553       } else {
1554          emit(MOV(dst, negate(op[0])));
1555       }
1556       break;
1557 
1558    case nir_op_f2b32:
1559       if (nir_src_bit_size(instr->src[0].src) == 64) {
1560          /* We use a MOV with conditional_mod to check if the provided value is
1561           * 0.0. We want this to flush denormalized numbers to zero, so we set a
1562           * source modifier on the source operand to trigger this, as source
1563           * modifiers don't affect the result of the testing against 0.0.
1564           */
1565          src_reg value = op[0];
1566          value.abs = true;
1567          vec4_instruction *inst = emit(MOV(dst_null_df(), value));
1568          inst->conditional_mod = BRW_CONDITIONAL_NZ;
1569 
1570          src_reg one = src_reg(this, glsl_type::ivec4_type);
1571          emit(MOV(dst_reg(one), brw_imm_d(~0)));
1572          inst = emit(BRW_OPCODE_SEL, dst, one, brw_imm_d(0));
1573          inst->predicate = BRW_PREDICATE_NORMAL;
1574       } else {
1575          emit(CMP(dst, op[0], brw_imm_f(0.0f), BRW_CONDITIONAL_NZ));
1576       }
1577       break;
1578 
1579    case nir_op_i2b32:
1580       emit(CMP(dst, op[0], brw_imm_d(0), BRW_CONDITIONAL_NZ));
1581       break;
1582 
1583    case nir_op_unpack_half_2x16_split_x:
1584    case nir_op_unpack_half_2x16_split_y:
1585    case nir_op_pack_half_2x16_split:
1586       unreachable("not reached: should not occur in vertex shader");
1587 
1588    case nir_op_unpack_snorm_2x16:
1589    case nir_op_unpack_unorm_2x16:
1590    case nir_op_pack_snorm_2x16:
1591    case nir_op_pack_unorm_2x16:
1592       unreachable("not reached: should be handled by lower_packing_builtins");
1593 
1594    case nir_op_pack_uvec4_to_uint:
1595       unreachable("not reached");
1596 
1597    case nir_op_pack_uvec2_to_uint: {
1598       dst_reg tmp1 = dst_reg(this, glsl_type::uint_type);
1599       tmp1.writemask = WRITEMASK_X;
1600       op[0].swizzle = BRW_SWIZZLE_YYYY;
1601       emit(SHL(tmp1, op[0], src_reg(brw_imm_ud(16u))));
1602 
1603       dst_reg tmp2 = dst_reg(this, glsl_type::uint_type);
1604       tmp2.writemask = WRITEMASK_X;
1605       op[0].swizzle = BRW_SWIZZLE_XXXX;
1606       emit(AND(tmp2, op[0], src_reg(brw_imm_ud(0xffffu))));
1607 
1608       emit(OR(dst, src_reg(tmp1), src_reg(tmp2)));
1609       break;
1610    }
1611 
1612    case nir_op_pack_64_2x32_split: {
1613       dst_reg result = dst_reg(this, glsl_type::dvec4_type);
1614       dst_reg tmp = dst_reg(this, glsl_type::uvec4_type);
1615       emit(MOV(tmp, retype(op[0], BRW_REGISTER_TYPE_UD)));
1616       emit(VEC4_OPCODE_SET_LOW_32BIT, result, src_reg(tmp));
1617       emit(MOV(tmp, retype(op[1], BRW_REGISTER_TYPE_UD)));
1618       emit(VEC4_OPCODE_SET_HIGH_32BIT, result, src_reg(tmp));
1619       emit(MOV(dst, src_reg(result)));
1620       break;
1621    }
1622 
1623    case nir_op_unpack_64_2x32_split_x:
1624    case nir_op_unpack_64_2x32_split_y: {
1625       enum opcode oper = (instr->op == nir_op_unpack_64_2x32_split_x) ?
1626          VEC4_OPCODE_PICK_LOW_32BIT : VEC4_OPCODE_PICK_HIGH_32BIT;
1627       dst_reg tmp = dst_reg(this, glsl_type::dvec4_type);
1628       emit(MOV(tmp, op[0]));
1629       dst_reg tmp2 = dst_reg(this, glsl_type::uvec4_type);
1630       emit(oper, tmp2, src_reg(tmp));
1631       emit(MOV(dst, src_reg(tmp2)));
1632       break;
1633    }
1634 
1635    case nir_op_unpack_half_2x16:
1636       /* As NIR does not guarantee that we have a correct swizzle outside the
1637        * boundaries of a vector, and the implementation of emit_unpack_half_2x16
1638        * uses the source operand in an operation with WRITEMASK_Y while our
1639        * source operand has only size 1, it accessed incorrect data producing
1640        * regressions in Piglit. We repeat the swizzle of the first component on the
1641        * rest of components to avoid regressions. In the vec4_visitor IR code path
1642        * this is not needed because the operand has already the correct swizzle.
1643        */
1644       op[0].swizzle = brw_compose_swizzle(BRW_SWIZZLE_XXXX, op[0].swizzle);
1645       emit_unpack_half_2x16(dst, op[0]);
1646       break;
1647 
1648    case nir_op_pack_half_2x16:
1649       emit_pack_half_2x16(dst, op[0]);
1650       break;
1651 
1652    case nir_op_unpack_unorm_4x8:
1653       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1654       emit_unpack_unorm_4x8(dst, op[0]);
1655       break;
1656 
1657    case nir_op_pack_unorm_4x8:
1658       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1659       emit_pack_unorm_4x8(dst, op[0]);
1660       break;
1661 
1662    case nir_op_unpack_snorm_4x8:
1663       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1664       emit_unpack_snorm_4x8(dst, op[0]);
1665       break;
1666 
1667    case nir_op_pack_snorm_4x8:
1668       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1669       emit_pack_snorm_4x8(dst, op[0]);
1670       break;
1671 
1672    case nir_op_bitfield_reverse:
1673       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1674       emit(BFREV(dst, op[0]));
1675       break;
1676 
1677    case nir_op_bit_count:
1678       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1679       emit(CBIT(dst, op[0]));
1680       break;
1681 
1682    case nir_op_ufind_msb:
1683       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1684       emit_find_msb_using_lzd(vec4_builder(this).at_end(), dst, op[0], false);
1685       break;
1686 
1687    case nir_op_ifind_msb: {
1688       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1689       vec4_builder bld = vec4_builder(this).at_end();
1690       src_reg src(dst);
1691 
1692       if (devinfo->ver < 7) {
1693          emit_find_msb_using_lzd(bld, dst, op[0], true);
1694       } else {
1695          emit(FBH(retype(dst, BRW_REGISTER_TYPE_UD), op[0]));
1696 
1697          /* FBH counts from the MSB side, while GLSL's findMSB() wants the
1698           * count from the LSB side. If FBH didn't return an error
1699           * (0xFFFFFFFF), then subtract the result from 31 to convert the MSB
1700           * count into an LSB count.
1701           */
1702          bld.CMP(dst_null_d(), src, brw_imm_d(-1), BRW_CONDITIONAL_NZ);
1703 
1704          inst = bld.ADD(dst, src, brw_imm_d(31));
1705          inst->predicate = BRW_PREDICATE_NORMAL;
1706          inst->src[0].negate = true;
1707       }
1708       break;
1709    }
1710 
1711    case nir_op_find_lsb: {
1712       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1713       vec4_builder bld = vec4_builder(this).at_end();
1714 
1715       if (devinfo->ver < 7) {
1716          dst_reg temp = bld.vgrf(BRW_REGISTER_TYPE_D);
1717 
1718          /* (x & -x) generates a value that consists of only the LSB of x.
1719           * For all powers of 2, findMSB(y) == findLSB(y).
1720           */
1721          src_reg src = src_reg(retype(op[0], BRW_REGISTER_TYPE_D));
1722          src_reg negated_src = src;
1723 
1724          /* One must be negated, and the other must be non-negated.  It
1725           * doesn't matter which is which.
1726           */
1727          negated_src.negate = true;
1728          src.negate = false;
1729 
1730          bld.AND(temp, src, negated_src);
1731          emit_find_msb_using_lzd(bld, dst, src_reg(temp), false);
1732       } else {
1733          bld.FBL(dst, op[0]);
1734       }
1735       break;
1736    }
1737 
1738    case nir_op_ubitfield_extract:
1739    case nir_op_ibitfield_extract:
1740       unreachable("should have been lowered");
1741    case nir_op_ubfe:
1742    case nir_op_ibfe:
1743       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1744       op[0] = fix_3src_operand(op[0]);
1745       op[1] = fix_3src_operand(op[1]);
1746       op[2] = fix_3src_operand(op[2]);
1747 
1748       emit(BFE(dst, op[2], op[1], op[0]));
1749       break;
1750 
1751    case nir_op_bfm:
1752       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1753       emit(BFI1(dst, op[0], op[1]));
1754       break;
1755 
1756    case nir_op_bfi:
1757       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1758       op[0] = fix_3src_operand(op[0]);
1759       op[1] = fix_3src_operand(op[1]);
1760       op[2] = fix_3src_operand(op[2]);
1761 
1762       emit(BFI2(dst, op[0], op[1], op[2]));
1763       break;
1764 
1765    case nir_op_bitfield_insert:
1766       unreachable("not reached: should have been lowered");
1767 
1768    case nir_op_fsign:
1769        if (type_sz(op[0].type) < 8) {
1770          /* AND(val, 0x80000000) gives the sign bit.
1771           *
1772           * Predicated OR ORs 1.0 (0x3f800000) with the sign bit if val is not
1773           * zero.
1774           */
1775          emit(CMP(dst_null_f(), op[0], brw_imm_f(0.0f), BRW_CONDITIONAL_NZ));
1776 
1777          op[0].type = BRW_REGISTER_TYPE_UD;
1778          dst.type = BRW_REGISTER_TYPE_UD;
1779          emit(AND(dst, op[0], brw_imm_ud(0x80000000u)));
1780 
1781          inst = emit(OR(dst, src_reg(dst), brw_imm_ud(0x3f800000u)));
1782          inst->predicate = BRW_PREDICATE_NORMAL;
1783          dst.type = BRW_REGISTER_TYPE_F;
1784       } else {
1785          /* For doubles we do the same but we need to consider:
1786           *
1787           * - We use a MOV with conditional_mod instead of a CMP so that we can
1788           *   skip loading a 0.0 immediate. We use a source modifier on the
1789           *   source of the MOV so that we flush denormalized values to 0.
1790           *   Since we want to compare against 0, this won't alter the result.
1791           * - We need to extract the high 32-bit of each DF where the sign
1792           *   is stored.
1793           * - We need to produce a DF result.
1794           */
1795 
1796          /* Check for zero */
1797          src_reg value = op[0];
1798          value.abs = true;
1799          inst = emit(MOV(dst_null_df(), value));
1800          inst->conditional_mod = BRW_CONDITIONAL_NZ;
1801 
1802          /* AND each high 32-bit channel with 0x80000000u */
1803          dst_reg tmp = dst_reg(this, glsl_type::uvec4_type);
1804          emit(VEC4_OPCODE_PICK_HIGH_32BIT, tmp, op[0]);
1805          emit(AND(tmp, src_reg(tmp), brw_imm_ud(0x80000000u)));
1806 
1807          /* Add 1.0 to each channel, predicated to skip the cases where the
1808           * channel's value was 0
1809           */
1810          inst = emit(OR(tmp, src_reg(tmp), brw_imm_ud(0x3f800000u)));
1811          inst->predicate = BRW_PREDICATE_NORMAL;
1812 
1813          /* Now convert the result from float to double */
1814          emit_conversion_to_double(dst, retype(src_reg(tmp),
1815                                                BRW_REGISTER_TYPE_F));
1816       }
1817       break;
1818 
1819    case nir_op_ishl:
1820       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1821       try_immediate_source(instr, op, false);
1822       emit(SHL(dst, op[0], op[1]));
1823       break;
1824 
1825    case nir_op_ishr:
1826       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1827       try_immediate_source(instr, op, false);
1828       emit(ASR(dst, op[0], op[1]));
1829       break;
1830 
1831    case nir_op_ushr:
1832       assert(nir_dest_bit_size(instr->dest.dest) < 64);
1833       try_immediate_source(instr, op, false);
1834       emit(SHR(dst, op[0], op[1]));
1835       break;
1836 
1837    case nir_op_ffma:
1838       if (type_sz(dst.type) == 8) {
1839          dst_reg mul_dst = dst_reg(this, glsl_type::dvec4_type);
1840          emit(MUL(mul_dst, op[1], op[0]));
1841          inst = emit(ADD(dst, src_reg(mul_dst), op[2]));
1842       } else {
1843          fix_float_operands(op, instr);
1844          inst = emit(MAD(dst, op[2], op[1], op[0]));
1845       }
1846       break;
1847 
1848    case nir_op_flrp:
1849       fix_float_operands(op, instr);
1850       inst = emit(LRP(dst, op[2], op[1], op[0]));
1851       break;
1852 
1853    case nir_op_b32csel:
1854       enum brw_predicate predicate;
1855       if (!optimize_predicate(instr, &predicate)) {
1856          emit(CMP(dst_null_d(), op[0], brw_imm_d(0), BRW_CONDITIONAL_NZ));
1857          switch (dst.writemask) {
1858          case WRITEMASK_X:
1859             predicate = BRW_PREDICATE_ALIGN16_REPLICATE_X;
1860             break;
1861          case WRITEMASK_Y:
1862             predicate = BRW_PREDICATE_ALIGN16_REPLICATE_Y;
1863             break;
1864          case WRITEMASK_Z:
1865             predicate = BRW_PREDICATE_ALIGN16_REPLICATE_Z;
1866             break;
1867          case WRITEMASK_W:
1868             predicate = BRW_PREDICATE_ALIGN16_REPLICATE_W;
1869             break;
1870          default:
1871             predicate = BRW_PREDICATE_NORMAL;
1872             break;
1873          }
1874       }
1875       inst = emit(BRW_OPCODE_SEL, dst, op[1], op[2]);
1876       inst->predicate = predicate;
1877       break;
1878 
1879    case nir_op_fdot2_replicated:
1880       try_immediate_source(instr, op, true);
1881       inst = emit(BRW_OPCODE_DP2, dst, op[0], op[1]);
1882       break;
1883 
1884    case nir_op_fdot3_replicated:
1885       try_immediate_source(instr, op, true);
1886       inst = emit(BRW_OPCODE_DP3, dst, op[0], op[1]);
1887       break;
1888 
1889    case nir_op_fdot4_replicated:
1890       try_immediate_source(instr, op, true);
1891       inst = emit(BRW_OPCODE_DP4, dst, op[0], op[1]);
1892       break;
1893 
1894    case nir_op_fdph_replicated:
1895       try_immediate_source(instr, op, false);
1896       inst = emit(BRW_OPCODE_DPH, dst, op[0], op[1]);
1897       break;
1898 
1899    case nir_op_fdiv:
1900       unreachable("not reached: should be lowered by DIV_TO_MUL_RCP in the compiler");
1901 
1902    case nir_op_fmod:
1903       unreachable("not reached: should be lowered by MOD_TO_FLOOR in the compiler");
1904 
1905    case nir_op_fsub:
1906    case nir_op_isub:
1907       unreachable("not reached: should be handled by ir_sub_to_add_neg");
1908 
1909    default:
1910       unreachable("Unimplemented ALU operation");
1911    }
1912 
1913    /* If we need to do a boolean resolve, replace the result with -(x & 1)
1914     * to sign extend the low bit to 0/~0
1915     */
1916    if (devinfo->ver <= 5 &&
1917        (instr->instr.pass_flags & BRW_NIR_BOOLEAN_MASK) ==
1918        BRW_NIR_BOOLEAN_NEEDS_RESOLVE) {
1919       dst_reg masked = dst_reg(this, glsl_type::int_type);
1920       masked.writemask = dst.writemask;
1921       emit(AND(masked, src_reg(dst), brw_imm_d(1)));
1922       src_reg masked_neg = src_reg(masked);
1923       masked_neg.negate = true;
1924       emit(MOV(retype(dst, BRW_REGISTER_TYPE_D), masked_neg));
1925    }
1926 }
1927 
1928 void
nir_emit_jump(nir_jump_instr * instr)1929 vec4_visitor::nir_emit_jump(nir_jump_instr *instr)
1930 {
1931    switch (instr->type) {
1932    case nir_jump_break:
1933       emit(BRW_OPCODE_BREAK);
1934       break;
1935 
1936    case nir_jump_continue:
1937       emit(BRW_OPCODE_CONTINUE);
1938       break;
1939 
1940    case nir_jump_return:
1941       FALLTHROUGH;
1942    default:
1943       unreachable("unknown jump");
1944    }
1945 }
1946 
1947 static bool
is_high_sampler(const struct intel_device_info * devinfo,src_reg sampler)1948 is_high_sampler(const struct intel_device_info *devinfo, src_reg sampler)
1949 {
1950    if (devinfo->verx10 != 75)
1951       return false;
1952 
1953    return sampler.file != IMM || sampler.ud >= 16;
1954 }
1955 
1956 void
nir_emit_texture(nir_tex_instr * instr)1957 vec4_visitor::nir_emit_texture(nir_tex_instr *instr)
1958 {
1959    unsigned texture = instr->texture_index;
1960    unsigned sampler = instr->sampler_index;
1961    src_reg texture_reg = brw_imm_ud(texture);
1962    src_reg sampler_reg = brw_imm_ud(sampler);
1963    src_reg coordinate;
1964    const glsl_type *coord_type = NULL;
1965    src_reg shadow_comparator;
1966    src_reg offset_value;
1967    src_reg lod, lod2;
1968    src_reg sample_index;
1969    src_reg mcs;
1970 
1971    dst_reg dest = get_nir_dest(instr->dest, instr->dest_type);
1972 
1973    /* The hardware requires a LOD for buffer textures */
1974    if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF)
1975       lod = brw_imm_d(0);
1976 
1977    /* Load the texture operation sources */
1978    uint32_t constant_offset = 0;
1979    for (unsigned i = 0; i < instr->num_srcs; i++) {
1980       switch (instr->src[i].src_type) {
1981       case nir_tex_src_comparator:
1982          shadow_comparator = get_nir_src(instr->src[i].src,
1983                                          BRW_REGISTER_TYPE_F, 1);
1984          break;
1985 
1986       case nir_tex_src_coord: {
1987          unsigned src_size = nir_tex_instr_src_size(instr, i);
1988 
1989          switch (instr->op) {
1990          case nir_texop_txf:
1991          case nir_texop_txf_ms:
1992          case nir_texop_samples_identical:
1993             coordinate = get_nir_src(instr->src[i].src, BRW_REGISTER_TYPE_D,
1994                                      src_size);
1995             coord_type = glsl_type::ivec(src_size);
1996             break;
1997 
1998          default:
1999             coordinate = get_nir_src(instr->src[i].src, BRW_REGISTER_TYPE_F,
2000                                      src_size);
2001             coord_type = glsl_type::vec(src_size);
2002             break;
2003          }
2004          break;
2005       }
2006 
2007       case nir_tex_src_ddx:
2008          lod = get_nir_src(instr->src[i].src, BRW_REGISTER_TYPE_F,
2009                            nir_tex_instr_src_size(instr, i));
2010          break;
2011 
2012       case nir_tex_src_ddy:
2013          lod2 = get_nir_src(instr->src[i].src, BRW_REGISTER_TYPE_F,
2014                            nir_tex_instr_src_size(instr, i));
2015          break;
2016 
2017       case nir_tex_src_lod:
2018          switch (instr->op) {
2019          case nir_texop_txs:
2020          case nir_texop_txf:
2021             lod = get_nir_src(instr->src[i].src, BRW_REGISTER_TYPE_D, 1);
2022             break;
2023 
2024          default:
2025             lod = get_nir_src(instr->src[i].src, BRW_REGISTER_TYPE_F, 1);
2026             break;
2027          }
2028          break;
2029 
2030       case nir_tex_src_ms_index: {
2031          sample_index = get_nir_src(instr->src[i].src, BRW_REGISTER_TYPE_D, 1);
2032          break;
2033       }
2034 
2035       case nir_tex_src_offset:
2036          if (!brw_texture_offset(instr, i, &constant_offset)) {
2037             offset_value =
2038                get_nir_src(instr->src[i].src, BRW_REGISTER_TYPE_D, 2);
2039          }
2040          break;
2041 
2042       case nir_tex_src_texture_offset: {
2043          /* Emit code to evaluate the actual indexing expression */
2044          src_reg src = get_nir_src(instr->src[i].src, 1);
2045          src_reg temp(this, glsl_type::uint_type);
2046          emit(ADD(dst_reg(temp), src, brw_imm_ud(texture)));
2047          texture_reg = emit_uniformize(temp);
2048          break;
2049       }
2050 
2051       case nir_tex_src_sampler_offset: {
2052          /* Emit code to evaluate the actual indexing expression */
2053          src_reg src = get_nir_src(instr->src[i].src, 1);
2054          src_reg temp(this, glsl_type::uint_type);
2055          emit(ADD(dst_reg(temp), src, brw_imm_ud(sampler)));
2056          sampler_reg = emit_uniformize(temp);
2057          break;
2058       }
2059 
2060       case nir_tex_src_projector:
2061          unreachable("Should be lowered by nir_lower_tex");
2062 
2063       case nir_tex_src_bias:
2064          unreachable("LOD bias is not valid for vertex shaders.\n");
2065 
2066       default:
2067          unreachable("unknown texture source");
2068       }
2069    }
2070 
2071    if (instr->op == nir_texop_txf_ms ||
2072        instr->op == nir_texop_samples_identical) {
2073       assert(coord_type != NULL);
2074       if (devinfo->ver >= 7 &&
2075           key_tex->compressed_multisample_layout_mask & (1 << texture)) {
2076          mcs = emit_mcs_fetch(coord_type, coordinate, texture_reg);
2077       } else {
2078          mcs = brw_imm_ud(0u);
2079       }
2080    }
2081 
2082    /* Stuff the channel select bits in the top of the texture offset */
2083    if (instr->op == nir_texop_tg4) {
2084       if (instr->component == 1 &&
2085           (key_tex->gather_channel_quirk_mask & (1 << texture))) {
2086          /* gather4 sampler is broken for green channel on RG32F --
2087           * we must ask for blue instead.
2088           */
2089          constant_offset |= 2 << 16;
2090       } else {
2091          constant_offset |= instr->component << 16;
2092       }
2093    }
2094 
2095    enum opcode opcode;
2096    switch (instr->op) {
2097    case nir_texop_tex:             opcode = SHADER_OPCODE_TXL;        break;
2098    case nir_texop_txl:             opcode = SHADER_OPCODE_TXL;        break;
2099    case nir_texop_txd:             opcode = SHADER_OPCODE_TXD;        break;
2100    case nir_texop_txf:             opcode = SHADER_OPCODE_TXF;        break;
2101    case nir_texop_txf_ms:          opcode = SHADER_OPCODE_TXF_CMS;    break;
2102    case nir_texop_txs:             opcode = SHADER_OPCODE_TXS;        break;
2103    case nir_texop_query_levels:    opcode = SHADER_OPCODE_TXS;        break;
2104    case nir_texop_texture_samples: opcode = SHADER_OPCODE_SAMPLEINFO; break;
2105    case nir_texop_tg4:
2106       opcode = offset_value.file != BAD_FILE ? SHADER_OPCODE_TG4_OFFSET
2107                                              : SHADER_OPCODE_TG4;
2108       break;
2109    case nir_texop_samples_identical: {
2110       /* There are some challenges implementing this for vec4, and it seems
2111        * unlikely to be used anyway.  For now, just return false ways.
2112        */
2113       emit(MOV(dest, brw_imm_ud(0u)));
2114       return;
2115    }
2116    case nir_texop_txb:
2117    case nir_texop_lod:
2118       unreachable("Implicit LOD is only valid inside fragment shaders.");
2119    default:
2120       unreachable("Unrecognized tex op");
2121    }
2122 
2123    vec4_instruction *inst = new(mem_ctx) vec4_instruction(opcode, dest);
2124 
2125    inst->offset = constant_offset;
2126 
2127    /* The message header is necessary for:
2128     * - Gfx4 (always)
2129     * - Texel offsets
2130     * - Gather channel selection
2131     * - Sampler indices too large to fit in a 4-bit value.
2132     * - Sampleinfo message - takes no parameters, but mlen = 0 is illegal
2133     */
2134    inst->header_size =
2135       (devinfo->ver < 5 ||
2136        inst->offset != 0 ||
2137        opcode == SHADER_OPCODE_TG4 ||
2138        opcode == SHADER_OPCODE_TG4_OFFSET ||
2139        opcode == SHADER_OPCODE_SAMPLEINFO ||
2140        is_high_sampler(devinfo, sampler_reg)) ? 1 : 0;
2141    inst->base_mrf = 2;
2142    inst->mlen = inst->header_size;
2143    inst->dst.writemask = WRITEMASK_XYZW;
2144    inst->shadow_compare = shadow_comparator.file != BAD_FILE;
2145 
2146    inst->src[1] = texture_reg;
2147    inst->src[2] = sampler_reg;
2148 
2149    /* MRF for the first parameter */
2150    int param_base = inst->base_mrf + inst->header_size;
2151 
2152    if (opcode == SHADER_OPCODE_TXS) {
2153       int writemask = devinfo->ver == 4 ? WRITEMASK_W : WRITEMASK_X;
2154       emit(MOV(dst_reg(MRF, param_base, lod.type, writemask), lod));
2155       inst->mlen++;
2156    } else if (opcode == SHADER_OPCODE_SAMPLEINFO) {
2157       inst->dst.writemask = WRITEMASK_X;
2158    } else {
2159       /* Load the coordinate */
2160       /* FINISHME: gl_clamp_mask and saturate */
2161       int coord_mask = (1 << instr->coord_components) - 1;
2162       int zero_mask = 0xf & ~coord_mask;
2163 
2164       emit(MOV(dst_reg(MRF, param_base, coordinate.type, coord_mask),
2165                coordinate));
2166       inst->mlen++;
2167 
2168       if (zero_mask != 0) {
2169          emit(MOV(dst_reg(MRF, param_base, coordinate.type, zero_mask),
2170                   brw_imm_d(0)));
2171       }
2172       /* Load the shadow comparator */
2173       if (shadow_comparator.file != BAD_FILE &&
2174           opcode != SHADER_OPCODE_TXD &&
2175           opcode != SHADER_OPCODE_TG4_OFFSET) {
2176 	 emit(MOV(dst_reg(MRF, param_base + 1, shadow_comparator.type,
2177 			  WRITEMASK_X),
2178 		  shadow_comparator));
2179 	 inst->mlen++;
2180       }
2181 
2182       /* Load the LOD info */
2183       switch (opcode) {
2184       case SHADER_OPCODE_TXL: {
2185 	 int mrf, writemask;
2186 	 if (devinfo->ver >= 5) {
2187 	    mrf = param_base + 1;
2188 	    if (shadow_comparator.file != BAD_FILE) {
2189 	       writemask = WRITEMASK_Y;
2190 	       /* mlen already incremented */
2191 	    } else {
2192 	       writemask = WRITEMASK_X;
2193 	       inst->mlen++;
2194 	    }
2195 	 } else /* devinfo->ver == 4 */ {
2196 	    mrf = param_base;
2197 	    writemask = WRITEMASK_W;
2198 	 }
2199 	 emit(MOV(dst_reg(MRF, mrf, lod.type, writemask), lod));
2200          break;
2201       }
2202 
2203       case SHADER_OPCODE_TXF:
2204          emit(MOV(dst_reg(MRF, param_base, lod.type, WRITEMASK_W), lod));
2205          break;
2206 
2207       case SHADER_OPCODE_TXF_CMS:
2208          emit(MOV(dst_reg(MRF, param_base + 1, sample_index.type, WRITEMASK_X),
2209                   sample_index));
2210          if (devinfo->ver >= 7) {
2211             /* MCS data is in the first channel of `mcs`, but we need to get it into
2212              * the .y channel of the second vec4 of params, so replicate .x across
2213              * the whole vec4 and then mask off everything except .y
2214              */
2215             mcs.swizzle = BRW_SWIZZLE_XXXX;
2216             emit(MOV(dst_reg(MRF, param_base + 1, glsl_type::uint_type, WRITEMASK_Y),
2217                      mcs));
2218          }
2219          inst->mlen++;
2220          break;
2221 
2222       case SHADER_OPCODE_TXD: {
2223          const brw_reg_type type = lod.type;
2224 
2225 	 if (devinfo->ver >= 5) {
2226 	    lod.swizzle = BRW_SWIZZLE4(SWIZZLE_X,SWIZZLE_X,SWIZZLE_Y,SWIZZLE_Y);
2227 	    lod2.swizzle = BRW_SWIZZLE4(SWIZZLE_X,SWIZZLE_X,SWIZZLE_Y,SWIZZLE_Y);
2228 	    emit(MOV(dst_reg(MRF, param_base + 1, type, WRITEMASK_XZ), lod));
2229 	    emit(MOV(dst_reg(MRF, param_base + 1, type, WRITEMASK_YW), lod2));
2230 	    inst->mlen++;
2231 
2232 	    if (nir_tex_instr_dest_size(instr) == 3 ||
2233                 shadow_comparator.file != BAD_FILE) {
2234 	       lod.swizzle = BRW_SWIZZLE_ZZZZ;
2235 	       lod2.swizzle = BRW_SWIZZLE_ZZZZ;
2236 	       emit(MOV(dst_reg(MRF, param_base + 2, type, WRITEMASK_X), lod));
2237 	       emit(MOV(dst_reg(MRF, param_base + 2, type, WRITEMASK_Y), lod2));
2238 	       inst->mlen++;
2239 
2240                if (shadow_comparator.file != BAD_FILE) {
2241                   emit(MOV(dst_reg(MRF, param_base + 2,
2242                                    shadow_comparator.type, WRITEMASK_Z),
2243                            shadow_comparator));
2244                }
2245 	    }
2246 	 } else /* devinfo->ver == 4 */ {
2247 	    emit(MOV(dst_reg(MRF, param_base + 1, type, WRITEMASK_XYZ), lod));
2248 	    emit(MOV(dst_reg(MRF, param_base + 2, type, WRITEMASK_XYZ), lod2));
2249 	    inst->mlen += 2;
2250 	 }
2251          break;
2252       }
2253 
2254       case SHADER_OPCODE_TG4_OFFSET:
2255          if (shadow_comparator.file != BAD_FILE) {
2256             emit(MOV(dst_reg(MRF, param_base, shadow_comparator.type, WRITEMASK_W),
2257                      shadow_comparator));
2258          }
2259 
2260          emit(MOV(dst_reg(MRF, param_base + 1, glsl_type::ivec2_type, WRITEMASK_XY),
2261                   offset_value));
2262          inst->mlen++;
2263          break;
2264 
2265       default:
2266          break;
2267       }
2268    }
2269 
2270    emit(inst);
2271 
2272    /* fixup num layers (z) for cube arrays: hardware returns faces * layers;
2273     * spec requires layers.
2274     */
2275    if (instr->op == nir_texop_txs && devinfo->ver < 7) {
2276       /* Gfx4-6 return 0 instead of 1 for single layer surfaces. */
2277       emit_minmax(BRW_CONDITIONAL_GE, writemask(inst->dst, WRITEMASK_Z),
2278                   src_reg(inst->dst), brw_imm_d(1));
2279    }
2280 
2281    if (instr->op == nir_texop_query_levels) {
2282       /* # levels is in .w */
2283       src_reg swizzled(dest);
2284       swizzled.swizzle = BRW_SWIZZLE4(SWIZZLE_W, SWIZZLE_W,
2285                                       SWIZZLE_W, SWIZZLE_W);
2286       emit(MOV(dest, swizzled));
2287    }
2288 }
2289 
2290 src_reg
emit_mcs_fetch(const glsl_type * coordinate_type,src_reg coordinate,src_reg surface)2291 vec4_visitor::emit_mcs_fetch(const glsl_type *coordinate_type,
2292                              src_reg coordinate, src_reg surface)
2293 {
2294    vec4_instruction *inst =
2295       new(mem_ctx) vec4_instruction(SHADER_OPCODE_TXF_MCS,
2296                                     dst_reg(this, glsl_type::uvec4_type));
2297    inst->base_mrf = 2;
2298    inst->src[1] = surface;
2299    inst->src[2] = brw_imm_ud(0); /* sampler */
2300    inst->mlen = 1;
2301 
2302    const int param_base = inst->base_mrf;
2303 
2304    /* parameters are: u, v, r, lod; lod will always be zero due to api restrictions */
2305    int coord_mask = (1 << coordinate_type->vector_elements) - 1;
2306    int zero_mask = 0xf & ~coord_mask;
2307 
2308    emit(MOV(dst_reg(MRF, param_base, coordinate_type, coord_mask),
2309             coordinate));
2310 
2311    emit(MOV(dst_reg(MRF, param_base, coordinate_type, zero_mask),
2312             brw_imm_d(0)));
2313 
2314    emit(inst);
2315    return src_reg(inst->dst);
2316 }
2317 
2318 void
nir_emit_undef(nir_ssa_undef_instr * instr)2319 vec4_visitor::nir_emit_undef(nir_ssa_undef_instr *instr)
2320 {
2321    nir_ssa_values[instr->def.index] =
2322       dst_reg(VGRF, alloc.allocate(DIV_ROUND_UP(instr->def.bit_size, 32)));
2323 }
2324 
2325 /* SIMD4x2 64bit data is stored in register space like this:
2326  *
2327  * r0.0:DF  x0 y0 z0 w0
2328  * r1.0:DF  x1 y1 z1 w1
2329  *
2330  * When we need to write data such as this to memory using 32-bit write
2331  * messages we need to shuffle it in this fashion:
2332  *
2333  * r0.0:DF  x0 y0 x1 y1 (to be written at base offset)
2334  * r0.0:DF  z0 w0 z1 w1 (to be written at base offset + 16)
2335  *
2336  * We need to do the inverse operation when we read using 32-bit messages,
2337  * which we can do by applying the same exact shuffling on the 64-bit data
2338  * read, only that because the data for each vertex is positioned differently
2339  * we need to apply different channel enables.
2340  *
2341  * This function takes 64bit data and shuffles it as explained above.
2342  *
2343  * The @for_write parameter is used to specify if the shuffling is being done
2344  * for proper SIMD4x2 64-bit data that needs to be shuffled prior to a 32-bit
2345  * write message (for_write = true), or instead we are doing the inverse
2346  * operation and we have just read 64-bit data using a 32-bit messages that we
2347  * need to shuffle to create valid SIMD4x2 64-bit data (for_write = false).
2348  *
2349  * If @block and @ref are non-NULL, then the shuffling is done after @ref,
2350  * otherwise the instructions are emitted normally at the end. The function
2351  * returns the last instruction inserted.
2352  *
2353  * Notice that @src and @dst cannot be the same register.
2354  */
2355 vec4_instruction *
shuffle_64bit_data(dst_reg dst,src_reg src,bool for_write,bool for_scratch,bblock_t * block,vec4_instruction * ref)2356 vec4_visitor::shuffle_64bit_data(dst_reg dst, src_reg src, bool for_write,
2357                                  bool for_scratch,
2358                                  bblock_t *block, vec4_instruction *ref)
2359 {
2360    assert(type_sz(src.type) == 8);
2361    assert(type_sz(dst.type) == 8);
2362    assert(!regions_overlap(dst, 2 * REG_SIZE, src, 2 * REG_SIZE));
2363    assert(!ref == !block);
2364 
2365    opcode mov_op = for_scratch ? VEC4_OPCODE_MOV_FOR_SCRATCH : BRW_OPCODE_MOV;
2366 
2367    const vec4_builder bld = !ref ? vec4_builder(this).at_end() :
2368                                    vec4_builder(this).at(block, ref->next);
2369 
2370    /* Resolve swizzle in src */
2371    if (src.swizzle != BRW_SWIZZLE_XYZW) {
2372       dst_reg data = dst_reg(this, glsl_type::dvec4_type);
2373       bld.emit(mov_op, data, src);
2374       src = src_reg(data);
2375    }
2376 
2377    /* dst+0.XY = src+0.XY */
2378    bld.group(4, 0).emit(mov_op, writemask(dst, WRITEMASK_XY), src);
2379 
2380    /* dst+0.ZW = src+1.XY */
2381    bld.group(4, for_write ? 1 : 0)
2382             .emit(mov_op, writemask(dst, WRITEMASK_ZW),
2383                   swizzle(byte_offset(src, REG_SIZE), BRW_SWIZZLE_XYXY));
2384 
2385    /* dst+1.XY = src+0.ZW */
2386    bld.group(4, for_write ? 0 : 1)
2387             .emit(mov_op, writemask(byte_offset(dst, REG_SIZE), WRITEMASK_XY),
2388                   swizzle(src, BRW_SWIZZLE_ZWZW));
2389 
2390    /* dst+1.ZW = src+1.ZW */
2391    return bld.group(4, 1)
2392             .emit(mov_op, writemask(byte_offset(dst, REG_SIZE), WRITEMASK_ZW),
2393                   byte_offset(src, REG_SIZE));
2394 }
2395 
2396 }
2397