projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
ac/nir: support 16-bit data in buffer_load_format opcodes
[mesa.git] / src / amd / llvm / ac_nir_to_llvm.c
1 /*
2 * Copyright © 2016 Bas Nieuwenhuizen
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 <llvm/Config/llvm-config.h>
25
26 #include "ac_nir_to_llvm.h"
27 #include "ac_llvm_build.h"
28 #include "ac_llvm_util.h"
29 #include "ac_binary.h"
30 #include "sid.h"
31 #include "nir/nir.h"
32 #include "nir/nir_deref.h"
33 #include "util/bitscan.h"
34 #include "util/u_math.h"
35 #include "ac_shader_abi.h"
36 #include "ac_shader_util.h"
37
38 struct ac_nir_context {
39 struct ac_llvm_context ac;
40 struct ac_shader_abi *abi;
41 const struct ac_shader_args *args;
42
43 gl_shader_stage stage;
44 shader_info *info;
45
46 LLVMValueRef *ssa_defs;
47
48 LLVMValueRef scratch;
49 LLVMValueRef constant_data;
50
51 struct hash_table *defs;
52 struct hash_table *phis;
53 struct hash_table *vars;
54 struct hash_table *verified_interp;
55
56 LLVMValueRef main_function;
57 LLVMBasicBlockRef continue_block;
58 LLVMBasicBlockRef break_block;
59
60 int num_locals;
61 LLVMValueRef *locals;
62 };
63
64 static LLVMValueRef get_sampler_desc_index(struct ac_nir_context *ctx,
65 nir_deref_instr *deref_instr,
66 const nir_instr *instr,
67 bool image);
68
69 static LLVMValueRef get_sampler_desc(struct ac_nir_context *ctx,
70 nir_deref_instr *deref_instr,
71 enum ac_descriptor_type desc_type,
72 const nir_instr *instr,
73 LLVMValueRef index,
74 bool image, bool write);
75
76 static void
77 build_store_values_extended(struct ac_llvm_context *ac,
78 LLVMValueRef *values,
79 unsigned value_count,
80 unsigned value_stride,
81 LLVMValueRef vec)
82 {
83 LLVMBuilderRef builder = ac->builder;
84 unsigned i;
85
86 for (i = 0; i < value_count; i++) {
87 LLVMValueRef ptr = values[i * value_stride];
88 LLVMValueRef index = LLVMConstInt(ac->i32, i, false);
89 LLVMValueRef value = LLVMBuildExtractElement(builder, vec, index, "");
90 LLVMBuildStore(builder, value, ptr);
91 }
92 }
93
94 static LLVMTypeRef get_def_type(struct ac_nir_context *ctx,
95 const nir_ssa_def *def)
96 {
97 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, def->bit_size);
98 if (def->num_components > 1) {
99 type = LLVMVectorType(type, def->num_components);
100 }
101 return type;
102 }
103
104 static LLVMValueRef get_src(struct ac_nir_context *nir, nir_src src)
105 {
106 assert(src.is_ssa);
107 return nir->ssa_defs[src.ssa->index];
108 }
109
110 static LLVMValueRef
111 get_memory_ptr(struct ac_nir_context *ctx, nir_src src, unsigned bit_size)
112 {
113 LLVMValueRef ptr = get_src(ctx, src);
114 ptr = LLVMBuildGEP(ctx->ac.builder, ctx->ac.lds, &ptr, 1, "");
115 int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
116
117 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, bit_size);
118
119 return LLVMBuildBitCast(ctx->ac.builder, ptr,
120 LLVMPointerType(type, addr_space), "");
121 }
122
123 static LLVMBasicBlockRef get_block(struct ac_nir_context *nir,
124 const struct nir_block *b)
125 {
126 struct hash_entry *entry = _mesa_hash_table_search(nir->defs, b);
127 return (LLVMBasicBlockRef)entry->data;
128 }
129
130 static LLVMValueRef get_alu_src(struct ac_nir_context *ctx,
131 nir_alu_src src,
132 unsigned num_components)
133 {
134 LLVMValueRef value = get_src(ctx, src.src);
135 bool need_swizzle = false;
136
137 assert(value);
138 unsigned src_components = ac_get_llvm_num_components(value);
139 for (unsigned i = 0; i < num_components; ++i) {
140 assert(src.swizzle[i] < src_components);
141 if (src.swizzle[i] != i)
142 need_swizzle = true;
143 }
144
145 if (need_swizzle || num_components != src_components) {
146 LLVMValueRef masks[] = {
147 LLVMConstInt(ctx->ac.i32, src.swizzle[0], false),
148 LLVMConstInt(ctx->ac.i32, src.swizzle[1], false),
149 LLVMConstInt(ctx->ac.i32, src.swizzle[2], false),
150 LLVMConstInt(ctx->ac.i32, src.swizzle[3], false)};
151
152 if (src_components > 1 && num_components == 1) {
153 value = LLVMBuildExtractElement(ctx->ac.builder, value,
154 masks[0], "");
155 } else if (src_components == 1 && num_components > 1) {
156 LLVMValueRef values[] = {value, value, value, value};
157 value = ac_build_gather_values(&ctx->ac, values, num_components);
158 } else {
159 LLVMValueRef swizzle = LLVMConstVector(masks, num_components);
160 value = LLVMBuildShuffleVector(ctx->ac.builder, value, value,
161 swizzle, "");
162 }
163 }
164 assert(!src.negate);
165 assert(!src.abs);
166 return value;
167 }
168
169 static LLVMValueRef emit_int_cmp(struct ac_llvm_context *ctx,
170 LLVMIntPredicate pred, LLVMValueRef src0,
171 LLVMValueRef src1)
172 {
173 LLVMValueRef result = LLVMBuildICmp(ctx->builder, pred, src0, src1, "");
174 return LLVMBuildSelect(ctx->builder, result,
175 LLVMConstInt(ctx->i32, 0xFFFFFFFF, false),
176 ctx->i32_0, "");
177 }
178
179 static LLVMValueRef emit_float_cmp(struct ac_llvm_context *ctx,
180 LLVMRealPredicate pred, LLVMValueRef src0,
181 LLVMValueRef src1)
182 {
183 LLVMValueRef result;
184 src0 = ac_to_float(ctx, src0);
185 src1 = ac_to_float(ctx, src1);
186 result = LLVMBuildFCmp(ctx->builder, pred, src0, src1, "");
187 return LLVMBuildSelect(ctx->builder, result,
188 LLVMConstInt(ctx->i32, 0xFFFFFFFF, false),
189 ctx->i32_0, "");
190 }
191
192 static LLVMValueRef emit_intrin_1f_param(struct ac_llvm_context *ctx,
193 const char *intrin,
194 LLVMTypeRef result_type,
195 LLVMValueRef src0)
196 {
197 char name[64], type[64];
198 LLVMValueRef params[] = {
199 ac_to_float(ctx, src0),
200 };
201
202 ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
203 ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
204 assert(length < sizeof(name));
205 return ac_build_intrinsic(ctx, name, result_type, params, 1, AC_FUNC_ATTR_READNONE);
206 }
207
208 static LLVMValueRef emit_intrin_2f_param(struct ac_llvm_context *ctx,
209 const char *intrin,
210 LLVMTypeRef result_type,
211 LLVMValueRef src0, LLVMValueRef src1)
212 {
213 char name[64], type[64];
214 LLVMValueRef params[] = {
215 ac_to_float(ctx, src0),
216 ac_to_float(ctx, src1),
217 };
218
219 ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
220 ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
221 assert(length < sizeof(name));
222 return ac_build_intrinsic(ctx, name, result_type, params, 2, AC_FUNC_ATTR_READNONE);
223 }
224
225 static LLVMValueRef emit_intrin_3f_param(struct ac_llvm_context *ctx,
226 const char *intrin,
227 LLVMTypeRef result_type,
228 LLVMValueRef src0, LLVMValueRef src1, LLVMValueRef src2)
229 {
230 char name[64], type[64];
231 LLVMValueRef params[] = {
232 ac_to_float(ctx, src0),
233 ac_to_float(ctx, src1),
234 ac_to_float(ctx, src2),
235 };
236
237 ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
238 ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
239 assert(length < sizeof(name));
240 return ac_build_intrinsic(ctx, name, result_type, params, 3, AC_FUNC_ATTR_READNONE);
241 }
242
243 static LLVMValueRef emit_bcsel(struct ac_llvm_context *ctx,
244 LLVMValueRef src0, LLVMValueRef src1, LLVMValueRef src2)
245 {
246 LLVMTypeRef src1_type = LLVMTypeOf(src1);
247 LLVMTypeRef src2_type = LLVMTypeOf(src2);
248
249 assert(LLVMGetTypeKind(LLVMTypeOf(src0)) != LLVMVectorTypeKind);
250
251 if (LLVMGetTypeKind(src1_type) == LLVMPointerTypeKind &&
252 LLVMGetTypeKind(src2_type) != LLVMPointerTypeKind) {
253 src2 = LLVMBuildIntToPtr(ctx->builder, src2, src1_type, "");
254 } else if (LLVMGetTypeKind(src2_type) == LLVMPointerTypeKind &&
255 LLVMGetTypeKind(src1_type) != LLVMPointerTypeKind) {
256 src1 = LLVMBuildIntToPtr(ctx->builder, src1, src2_type, "");
257 }
258
259 LLVMValueRef v = LLVMBuildICmp(ctx->builder, LLVMIntNE, src0,
260 ctx->i32_0, "");
261 return LLVMBuildSelect(ctx->builder, v,
262 ac_to_integer_or_pointer(ctx, src1),
263 ac_to_integer_or_pointer(ctx, src2), "");
264 }
265
266 static LLVMValueRef emit_iabs(struct ac_llvm_context *ctx,
267 LLVMValueRef src0)
268 {
269 return ac_build_imax(ctx, src0, LLVMBuildNeg(ctx->builder, src0, ""));
270 }
271
272 static LLVMValueRef emit_uint_carry(struct ac_llvm_context *ctx,
273 const char *intrin,
274 LLVMValueRef src0, LLVMValueRef src1)
275 {
276 LLVMTypeRef ret_type;
277 LLVMTypeRef types[] = { ctx->i32, ctx->i1 };
278 LLVMValueRef res;
279 LLVMValueRef params[] = { src0, src1 };
280 ret_type = LLVMStructTypeInContext(ctx->context, types,
281 2, true);
282
283 res = ac_build_intrinsic(ctx, intrin, ret_type,
284 params, 2, AC_FUNC_ATTR_READNONE);
285
286 res = LLVMBuildExtractValue(ctx->builder, res, 1, "");
287 res = LLVMBuildZExt(ctx->builder, res, ctx->i32, "");
288 return res;
289 }
290
291 static LLVMValueRef emit_b2f(struct ac_llvm_context *ctx,
292 LLVMValueRef src0,
293 unsigned bitsize)
294 {
295 LLVMValueRef result = LLVMBuildAnd(ctx->builder, src0,
296 LLVMBuildBitCast(ctx->builder, LLVMConstReal(ctx->f32, 1.0), ctx->i32, ""),
297 "");
298 result = LLVMBuildBitCast(ctx->builder, result, ctx->f32, "");
299
300 switch (bitsize) {
301 case 16:
302 return LLVMBuildFPTrunc(ctx->builder, result, ctx->f16, "");
303 case 32:
304 return result;
305 case 64:
306 return LLVMBuildFPExt(ctx->builder, result, ctx->f64, "");
307 default:
308 unreachable("Unsupported bit size.");
309 }
310 }
311
312 static LLVMValueRef emit_f2b(struct ac_llvm_context *ctx,
313 LLVMValueRef src0)
314 {
315 src0 = ac_to_float(ctx, src0);
316 LLVMValueRef zero = LLVMConstNull(LLVMTypeOf(src0));
317 return LLVMBuildSExt(ctx->builder,
318 LLVMBuildFCmp(ctx->builder, LLVMRealUNE, src0, zero, ""),
319 ctx->i32, "");
320 }
321
322 static LLVMValueRef emit_b2i(struct ac_llvm_context *ctx,
323 LLVMValueRef src0,
324 unsigned bitsize)
325 {
326 LLVMValueRef result = LLVMBuildAnd(ctx->builder, src0, ctx->i32_1, "");
327
328 switch (bitsize) {
329 case 8:
330 return LLVMBuildTrunc(ctx->builder, result, ctx->i8, "");
331 case 16:
332 return LLVMBuildTrunc(ctx->builder, result, ctx->i16, "");
333 case 32:
334 return result;
335 case 64:
336 return LLVMBuildZExt(ctx->builder, result, ctx->i64, "");
337 default:
338 unreachable("Unsupported bit size.");
339 }
340 }
341
342 static LLVMValueRef emit_i2b(struct ac_llvm_context *ctx,
343 LLVMValueRef src0)
344 {
345 LLVMValueRef zero = LLVMConstNull(LLVMTypeOf(src0));
346 return LLVMBuildSExt(ctx->builder,
347 LLVMBuildICmp(ctx->builder, LLVMIntNE, src0, zero, ""),
348 ctx->i32, "");
349 }
350
351 static LLVMValueRef emit_f2f16(struct ac_llvm_context *ctx,
352 LLVMValueRef src0)
353 {
354 LLVMValueRef result;
355 LLVMValueRef cond = NULL;
356
357 src0 = ac_to_float(ctx, src0);
358 result = LLVMBuildFPTrunc(ctx->builder, src0, ctx->f16, "");
359
360 if (ctx->chip_class >= GFX8) {
361 LLVMValueRef args[2];
362 /* Check if the result is a denormal - and flush to 0 if so. */
363 args[0] = result;
364 args[1] = LLVMConstInt(ctx->i32, N_SUBNORMAL | P_SUBNORMAL, false);
365 cond = ac_build_intrinsic(ctx, "llvm.amdgcn.class.f16", ctx->i1, args, 2, AC_FUNC_ATTR_READNONE);
366 }
367
368 /* need to convert back up to f32 */
369 result = LLVMBuildFPExt(ctx->builder, result, ctx->f32, "");
370
371 if (ctx->chip_class >= GFX8)
372 result = LLVMBuildSelect(ctx->builder, cond, ctx->f32_0, result, "");
373 else {
374 /* for GFX6-GFX7 */
375 /* 0x38800000 is smallest half float value (2^-14) in 32-bit float,
376 * so compare the result and flush to 0 if it's smaller.
377 */
378 LLVMValueRef temp, cond2;
379 temp = emit_intrin_1f_param(ctx, "llvm.fabs", ctx->f32, result);
380 cond = LLVMBuildFCmp(ctx->builder, LLVMRealOGT,
381 LLVMBuildBitCast(ctx->builder, LLVMConstInt(ctx->i32, 0x38800000, false), ctx->f32, ""),
382 temp, "");
383 cond2 = LLVMBuildFCmp(ctx->builder, LLVMRealONE,
384 temp, ctx->f32_0, "");
385 cond = LLVMBuildAnd(ctx->builder, cond, cond2, "");
386 result = LLVMBuildSelect(ctx->builder, cond, ctx->f32_0, result, "");
387 }
388 return result;
389 }
390
391 static LLVMValueRef emit_umul_high(struct ac_llvm_context *ctx,
392 LLVMValueRef src0, LLVMValueRef src1)
393 {
394 LLVMValueRef dst64, result;
395 src0 = LLVMBuildZExt(ctx->builder, src0, ctx->i64, "");
396 src1 = LLVMBuildZExt(ctx->builder, src1, ctx->i64, "");
397
398 dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
399 dst64 = LLVMBuildLShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
400 result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
401 return result;
402 }
403
404 static LLVMValueRef emit_imul_high(struct ac_llvm_context *ctx,
405 LLVMValueRef src0, LLVMValueRef src1)
406 {
407 LLVMValueRef dst64, result;
408 src0 = LLVMBuildSExt(ctx->builder, src0, ctx->i64, "");
409 src1 = LLVMBuildSExt(ctx->builder, src1, ctx->i64, "");
410
411 dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
412 dst64 = LLVMBuildAShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
413 result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
414 return result;
415 }
416
417 static LLVMValueRef emit_bfm(struct ac_llvm_context *ctx,
418 LLVMValueRef bits, LLVMValueRef offset)
419 {
420 /* mask = ((1 << bits) - 1) << offset */
421 return LLVMBuildShl(ctx->builder,
422 LLVMBuildSub(ctx->builder,
423 LLVMBuildShl(ctx->builder,
424 ctx->i32_1,
425 bits, ""),
426 ctx->i32_1, ""),
427 offset, "");
428 }
429
430 static LLVMValueRef emit_bitfield_select(struct ac_llvm_context *ctx,
431 LLVMValueRef mask, LLVMValueRef insert,
432 LLVMValueRef base)
433 {
434 /* Calculate:
435 * (mask & insert) | (~mask & base) = base ^ (mask & (insert ^ base))
436 * Use the right-hand side, which the LLVM backend can convert to V_BFI.
437 */
438 return LLVMBuildXor(ctx->builder, base,
439 LLVMBuildAnd(ctx->builder, mask,
440 LLVMBuildXor(ctx->builder, insert, base, ""), ""), "");
441 }
442
443 static LLVMValueRef emit_pack_2x16(struct ac_llvm_context *ctx,
444 LLVMValueRef src0,
445 LLVMValueRef (*pack)(struct ac_llvm_context *ctx,
446 LLVMValueRef args[2]))
447 {
448 LLVMValueRef comp[2];
449
450 src0 = ac_to_float(ctx, src0);
451 comp[0] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_0, "");
452 comp[1] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_1, "");
453
454 return LLVMBuildBitCast(ctx->builder, pack(ctx, comp), ctx->i32, "");
455 }
456
457 static LLVMValueRef emit_unpack_half_2x16(struct ac_llvm_context *ctx,
458 LLVMValueRef src0)
459 {
460 LLVMValueRef const16 = LLVMConstInt(ctx->i32, 16, false);
461 LLVMValueRef temps[2], val;
462 int i;
463
464 for (i = 0; i < 2; i++) {
465 val = i == 1 ? LLVMBuildLShr(ctx->builder, src0, const16, "") : src0;
466 val = LLVMBuildTrunc(ctx->builder, val, ctx->i16, "");
467 val = LLVMBuildBitCast(ctx->builder, val, ctx->f16, "");
468 temps[i] = LLVMBuildFPExt(ctx->builder, val, ctx->f32, "");
469 }
470 return ac_build_gather_values(ctx, temps, 2);
471 }
472
473 static LLVMValueRef emit_ddxy(struct ac_nir_context *ctx,
474 nir_op op,
475 LLVMValueRef src0)
476 {
477 unsigned mask;
478 int idx;
479 LLVMValueRef result;
480
481 if (op == nir_op_fddx_fine)
482 mask = AC_TID_MASK_LEFT;
483 else if (op == nir_op_fddy_fine)
484 mask = AC_TID_MASK_TOP;
485 else
486 mask = AC_TID_MASK_TOP_LEFT;
487
488 /* for DDX we want to next X pixel, DDY next Y pixel. */
489 if (op == nir_op_fddx_fine ||
490 op == nir_op_fddx_coarse ||
491 op == nir_op_fddx)
492 idx = 1;
493 else
494 idx = 2;
495
496 result = ac_build_ddxy(&ctx->ac, mask, idx, src0);
497 return result;
498 }
499
500 struct waterfall_context {
501 LLVMBasicBlockRef phi_bb[2];
502 bool use_waterfall;
503 };
504
505 /* To deal with divergent descriptors we can create a loop that handles all
506 * lanes with the same descriptor on a given iteration (henceforth a
507 * waterfall loop).
508 *
509 * These helper create the begin and end of the loop leaving the caller
510 * to implement the body.
511 *
512 * params:
513 * - ctx is the usal nir context
514 * - wctx is a temporary struct containing some loop info. Can be left uninitialized.
515 * - value is the possibly divergent value for which we built the loop
516 * - divergent is whether value is actually divergent. If false we just pass
517 * things through.
518 */
519 static LLVMValueRef enter_waterfall(struct ac_nir_context *ctx,
520 struct waterfall_context *wctx,
521 LLVMValueRef value, bool divergent)
522 {
523 /* If the app claims the value is divergent but it is constant we can
524 * end up with a dynamic index of NULL. */
525 if (!value)
526 divergent = false;
527
528 wctx->use_waterfall = divergent;
529 if (!divergent)
530 return value;
531
532 ac_build_bgnloop(&ctx->ac, 6000);
533
534 LLVMValueRef scalar_value = ac_build_readlane(&ctx->ac, value, NULL);
535
536 LLVMValueRef active = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, value,
537 scalar_value, "uniform_active");
538
539 wctx->phi_bb[0] = LLVMGetInsertBlock(ctx->ac.builder);
540 ac_build_ifcc(&ctx->ac, active, 6001);
541
542 return scalar_value;
543 }
544
545 static LLVMValueRef exit_waterfall(struct ac_nir_context *ctx,
546 struct waterfall_context *wctx,
547 LLVMValueRef value)
548 {
549 LLVMValueRef ret = NULL;
550 LLVMValueRef phi_src[2];
551 LLVMValueRef cc_phi_src[2] = {
552 LLVMConstInt(ctx->ac.i32, 0, false),
553 LLVMConstInt(ctx->ac.i32, 0xffffffff, false),
554 };
555
556 if (!wctx->use_waterfall)
557 return value;
558
559 wctx->phi_bb[1] = LLVMGetInsertBlock(ctx->ac.builder);
560
561 ac_build_endif(&ctx->ac, 6001);
562
563 if (value) {
564 phi_src[0] = LLVMGetUndef(LLVMTypeOf(value));
565 phi_src[1] = value;
566
567 ret = ac_build_phi(&ctx->ac, LLVMTypeOf(value), 2, phi_src, wctx->phi_bb);
568 }
569
570 /*
571 * By using the optimization barrier on the exit decision, we decouple
572 * the operations from the break, and hence avoid LLVM hoisting the
573 * opteration into the break block.
574 */
575 LLVMValueRef cc = ac_build_phi(&ctx->ac, ctx->ac.i32, 2, cc_phi_src, wctx->phi_bb);
576 ac_build_optimization_barrier(&ctx->ac, &cc);
577
578 LLVMValueRef active = LLVMBuildICmp(ctx->ac.builder, LLVMIntNE, cc, ctx->ac.i32_0, "uniform_active2");
579 ac_build_ifcc(&ctx->ac, active, 6002);
580 ac_build_break(&ctx->ac);
581 ac_build_endif(&ctx->ac, 6002);
582
583 ac_build_endloop(&ctx->ac, 6000);
584 return ret;
585 }
586
587 static void visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
588 {
589 LLVMValueRef src[4], result = NULL;
590 unsigned num_components = instr->dest.dest.ssa.num_components;
591 unsigned src_components;
592 LLVMTypeRef def_type = get_def_type(ctx, &instr->dest.dest.ssa);
593 bool saved_inexact = false;
594
595 if (instr->exact)
596 saved_inexact = ac_disable_inexact_math(ctx->ac.builder);
597
598 assert(nir_op_infos[instr->op].num_inputs <= ARRAY_SIZE(src));
599 switch (instr->op) {
600 case nir_op_vec2:
601 case nir_op_vec3:
602 case nir_op_vec4:
603 src_components = 1;
604 break;
605 case nir_op_pack_half_2x16:
606 case nir_op_pack_snorm_2x16:
607 case nir_op_pack_unorm_2x16:
608 src_components = 2;
609 break;
610 case nir_op_unpack_half_2x16:
611 src_components = 1;
612 break;
613 case nir_op_cube_face_coord:
614 case nir_op_cube_face_index:
615 src_components = 3;
616 break;
617 default:
618 src_components = num_components;
619 break;
620 }
621 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
622 src[i] = get_alu_src(ctx, instr->src[i], src_components);
623
624 switch (instr->op) {
625 case nir_op_mov:
626 result = src[0];
627 break;
628 case nir_op_fneg:
629 src[0] = ac_to_float(&ctx->ac, src[0]);
630 result = LLVMBuildFNeg(ctx->ac.builder, src[0], "");
631 if (ctx->ac.float_mode == AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO) {
632 /* fneg will be optimized by backend compiler with sign
633 * bit removed via XOR. This is probably a LLVM bug.
634 */
635 result = ac_build_canonicalize(&ctx->ac, result,
636 instr->dest.dest.ssa.bit_size);
637 }
638 break;
639 case nir_op_ineg:
640 result = LLVMBuildNeg(ctx->ac.builder, src[0], "");
641 break;
642 case nir_op_inot:
643 result = LLVMBuildNot(ctx->ac.builder, src[0], "");
644 break;
645 case nir_op_iadd:
646 result = LLVMBuildAdd(ctx->ac.builder, src[0], src[1], "");
647 break;
648 case nir_op_fadd:
649 src[0] = ac_to_float(&ctx->ac, src[0]);
650 src[1] = ac_to_float(&ctx->ac, src[1]);
651 result = LLVMBuildFAdd(ctx->ac.builder, src[0], src[1], "");
652 break;
653 case nir_op_fsub:
654 src[0] = ac_to_float(&ctx->ac, src[0]);
655 src[1] = ac_to_float(&ctx->ac, src[1]);
656 result = LLVMBuildFSub(ctx->ac.builder, src[0], src[1], "");
657 break;
658 case nir_op_isub:
659 result = LLVMBuildSub(ctx->ac.builder, src[0], src[1], "");
660 break;
661 case nir_op_imul:
662 result = LLVMBuildMul(ctx->ac.builder, src[0], src[1], "");
663 break;
664 case nir_op_imod:
665 result = LLVMBuildSRem(ctx->ac.builder, src[0], src[1], "");
666 break;
667 case nir_op_umod:
668 result = LLVMBuildURem(ctx->ac.builder, src[0], src[1], "");
669 break;
670 case nir_op_fmod:
671 /* lower_fmod only lower 16-bit and 32-bit fmod */
672 assert(instr->dest.dest.ssa.bit_size == 64);
673 src[0] = ac_to_float(&ctx->ac, src[0]);
674 src[1] = ac_to_float(&ctx->ac, src[1]);
675 result = ac_build_fdiv(&ctx->ac, src[0], src[1]);
676 result = emit_intrin_1f_param(&ctx->ac, "llvm.floor",
677 ac_to_float_type(&ctx->ac, def_type), result);
678 result = LLVMBuildFMul(ctx->ac.builder, src[1] , result, "");
679 result = LLVMBuildFSub(ctx->ac.builder, src[0], result, "");
680 break;
681 case nir_op_irem:
682 result = LLVMBuildSRem(ctx->ac.builder, src[0], src[1], "");
683 break;
684 case nir_op_idiv:
685 result = LLVMBuildSDiv(ctx->ac.builder, src[0], src[1], "");
686 break;
687 case nir_op_udiv:
688 result = LLVMBuildUDiv(ctx->ac.builder, src[0], src[1], "");
689 break;
690 case nir_op_fmul:
691 src[0] = ac_to_float(&ctx->ac, src[0]);
692 src[1] = ac_to_float(&ctx->ac, src[1]);
693 result = LLVMBuildFMul(ctx->ac.builder, src[0], src[1], "");
694 break;
695 case nir_op_frcp:
696 result = emit_intrin_1f_param(&ctx->ac, "llvm.amdgcn.rcp",
697 ac_to_float_type(&ctx->ac, def_type), src[0]);
698 break;
699 case nir_op_iand:
700 result = LLVMBuildAnd(ctx->ac.builder, src[0], src[1], "");
701 break;
702 case nir_op_ior:
703 result = LLVMBuildOr(ctx->ac.builder, src[0], src[1], "");
704 break;
705 case nir_op_ixor:
706 result = LLVMBuildXor(ctx->ac.builder, src[0], src[1], "");
707 break;
708 case nir_op_ishl:
709 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) < ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
710 src[1] = LLVMBuildZExt(ctx->ac.builder, src[1],
711 LLVMTypeOf(src[0]), "");
712 else if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) > ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
713 src[1] = LLVMBuildTrunc(ctx->ac.builder, src[1],
714 LLVMTypeOf(src[0]), "");
715 result = LLVMBuildShl(ctx->ac.builder, src[0], src[1], "");
716 break;
717 case nir_op_ishr:
718 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) < ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
719 src[1] = LLVMBuildZExt(ctx->ac.builder, src[1],
720 LLVMTypeOf(src[0]), "");
721 else if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) > ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
722 src[1] = LLVMBuildTrunc(ctx->ac.builder, src[1],
723 LLVMTypeOf(src[0]), "");
724 result = LLVMBuildAShr(ctx->ac.builder, src[0], src[1], "");
725 break;
726 case nir_op_ushr:
727 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) < ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
728 src[1] = LLVMBuildZExt(ctx->ac.builder, src[1],
729 LLVMTypeOf(src[0]), "");
730 else if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) > ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
731 src[1] = LLVMBuildTrunc(ctx->ac.builder, src[1],
732 LLVMTypeOf(src[0]), "");
733 result = LLVMBuildLShr(ctx->ac.builder, src[0], src[1], "");
734 break;
735 case nir_op_ilt32:
736 result = emit_int_cmp(&ctx->ac, LLVMIntSLT, src[0], src[1]);
737 break;
738 case nir_op_ine32:
739 result = emit_int_cmp(&ctx->ac, LLVMIntNE, src[0], src[1]);
740 break;
741 case nir_op_ieq32:
742 result = emit_int_cmp(&ctx->ac, LLVMIntEQ, src[0], src[1]);
743 break;
744 case nir_op_ige32:
745 result = emit_int_cmp(&ctx->ac, LLVMIntSGE, src[0], src[1]);
746 break;
747 case nir_op_ult32:
748 result = emit_int_cmp(&ctx->ac, LLVMIntULT, src[0], src[1]);
749 break;
750 case nir_op_uge32:
751 result = emit_int_cmp(&ctx->ac, LLVMIntUGE, src[0], src[1]);
752 break;
753 case nir_op_feq32:
754 result = emit_float_cmp(&ctx->ac, LLVMRealOEQ, src[0], src[1]);
755 break;
756 case nir_op_fne32:
757 result = emit_float_cmp(&ctx->ac, LLVMRealUNE, src[0], src[1]);
758 break;
759 case nir_op_flt32:
760 result = emit_float_cmp(&ctx->ac, LLVMRealOLT, src[0], src[1]);
761 break;
762 case nir_op_fge32:
763 result = emit_float_cmp(&ctx->ac, LLVMRealOGE, src[0], src[1]);
764 break;
765 case nir_op_fabs:
766 result = emit_intrin_1f_param(&ctx->ac, "llvm.fabs",
767 ac_to_float_type(&ctx->ac, def_type), src[0]);
768 if (ctx->ac.float_mode == AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO) {
769 /* fabs will be optimized by backend compiler with sign
770 * bit removed via AND.
771 */
772 result = ac_build_canonicalize(&ctx->ac, result,
773 instr->dest.dest.ssa.bit_size);
774 }
775 break;
776 case nir_op_iabs:
777 result = emit_iabs(&ctx->ac, src[0]);
778 break;
779 case nir_op_imax:
780 result = ac_build_imax(&ctx->ac, src[0], src[1]);
781 break;
782 case nir_op_imin:
783 result = ac_build_imin(&ctx->ac, src[0], src[1]);
784 break;
785 case nir_op_umax:
786 result = ac_build_umax(&ctx->ac, src[0], src[1]);
787 break;
788 case nir_op_umin:
789 result = ac_build_umin(&ctx->ac, src[0], src[1]);
790 break;
791 case nir_op_isign:
792 result = ac_build_isign(&ctx->ac, src[0],
793 instr->dest.dest.ssa.bit_size);
794 break;
795 case nir_op_fsign:
796 src[0] = ac_to_float(&ctx->ac, src[0]);
797 result = ac_build_fsign(&ctx->ac, src[0],
798 instr->dest.dest.ssa.bit_size);
799 break;
800 case nir_op_ffloor:
801 result = emit_intrin_1f_param(&ctx->ac, "llvm.floor",
802 ac_to_float_type(&ctx->ac, def_type), src[0]);
803 break;
804 case nir_op_ftrunc:
805 result = emit_intrin_1f_param(&ctx->ac, "llvm.trunc",
806 ac_to_float_type(&ctx->ac, def_type), src[0]);
807 break;
808 case nir_op_fceil:
809 result = emit_intrin_1f_param(&ctx->ac, "llvm.ceil",
810 ac_to_float_type(&ctx->ac, def_type), src[0]);
811 break;
812 case nir_op_fround_even:
813 result = emit_intrin_1f_param(&ctx->ac, "llvm.rint",
814 ac_to_float_type(&ctx->ac, def_type),src[0]);
815 break;
816 case nir_op_ffract:
817 src[0] = ac_to_float(&ctx->ac, src[0]);
818 result = ac_build_fract(&ctx->ac, src[0],
819 instr->dest.dest.ssa.bit_size);
820 break;
821 case nir_op_fsin:
822 result = emit_intrin_1f_param(&ctx->ac, "llvm.sin",
823 ac_to_float_type(&ctx->ac, def_type), src[0]);
824 break;
825 case nir_op_fcos:
826 result = emit_intrin_1f_param(&ctx->ac, "llvm.cos",
827 ac_to_float_type(&ctx->ac, def_type), src[0]);
828 break;
829 case nir_op_fsqrt:
830 result = emit_intrin_1f_param(&ctx->ac, "llvm.sqrt",
831 ac_to_float_type(&ctx->ac, def_type), src[0]);
832 break;
833 case nir_op_fexp2:
834 result = emit_intrin_1f_param(&ctx->ac, "llvm.exp2",
835 ac_to_float_type(&ctx->ac, def_type), src[0]);
836 break;
837 case nir_op_flog2:
838 result = emit_intrin_1f_param(&ctx->ac, "llvm.log2",
839 ac_to_float_type(&ctx->ac, def_type), src[0]);
840 break;
841 case nir_op_frsq:
842 result = emit_intrin_1f_param(&ctx->ac, "llvm.amdgcn.rsq",
843 ac_to_float_type(&ctx->ac, def_type), src[0]);
844 break;
845 case nir_op_frexp_exp:
846 src[0] = ac_to_float(&ctx->ac, src[0]);
847 result = ac_build_frexp_exp(&ctx->ac, src[0],
848 ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])));
849 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) == 16)
850 result = LLVMBuildSExt(ctx->ac.builder, result,
851 ctx->ac.i32, "");
852 break;
853 case nir_op_frexp_sig:
854 src[0] = ac_to_float(&ctx->ac, src[0]);
855 result = ac_build_frexp_mant(&ctx->ac, src[0],
856 instr->dest.dest.ssa.bit_size);
857 break;
858 case nir_op_fpow:
859 result = emit_intrin_2f_param(&ctx->ac, "llvm.pow",
860 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
861 break;
862 case nir_op_fmax:
863 result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
864 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
865 if (ctx->ac.chip_class < GFX9 &&
866 instr->dest.dest.ssa.bit_size == 32) {
867 /* Only pre-GFX9 chips do not flush denorms. */
868 result = ac_build_canonicalize(&ctx->ac, result,
869 instr->dest.dest.ssa.bit_size);
870 }
871 break;
872 case nir_op_fmin:
873 result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
874 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
875 if (ctx->ac.chip_class < GFX9 &&
876 instr->dest.dest.ssa.bit_size == 32) {
877 /* Only pre-GFX9 chips do not flush denorms. */
878 result = ac_build_canonicalize(&ctx->ac, result,
879 instr->dest.dest.ssa.bit_size);
880 }
881 break;
882 case nir_op_ffma:
883 /* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
884 result = emit_intrin_3f_param(&ctx->ac, ctx->ac.chip_class >= GFX10 ? "llvm.fma" : "llvm.fmuladd",
885 ac_to_float_type(&ctx->ac, def_type), src[0], src[1], src[2]);
886 break;
887 case nir_op_ldexp:
888 src[0] = ac_to_float(&ctx->ac, src[0]);
889 if (ac_get_elem_bits(&ctx->ac, def_type) == 32)
890 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f32", ctx->ac.f32, src, 2, AC_FUNC_ATTR_READNONE);
891 else if (ac_get_elem_bits(&ctx->ac, def_type) == 16)
892 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f16", ctx->ac.f16, src, 2, AC_FUNC_ATTR_READNONE);
893 else
894 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f64", ctx->ac.f64, src, 2, AC_FUNC_ATTR_READNONE);
895 break;
896 case nir_op_bfm:
897 result = emit_bfm(&ctx->ac, src[0], src[1]);
898 break;
899 case nir_op_bitfield_select:
900 result = emit_bitfield_select(&ctx->ac, src[0], src[1], src[2]);
901 break;
902 case nir_op_ubfe:
903 result = ac_build_bfe(&ctx->ac, src[0], src[1], src[2], false);
904 break;
905 case nir_op_ibfe:
906 result = ac_build_bfe(&ctx->ac, src[0], src[1], src[2], true);
907 break;
908 case nir_op_bitfield_reverse:
909 result = ac_build_bitfield_reverse(&ctx->ac, src[0]);
910 break;
911 case nir_op_bit_count:
912 result = ac_build_bit_count(&ctx->ac, src[0]);
913 break;
914 case nir_op_vec2:
915 case nir_op_vec3:
916 case nir_op_vec4:
917 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
918 src[i] = ac_to_integer(&ctx->ac, src[i]);
919 result = ac_build_gather_values(&ctx->ac, src, num_components);
920 break;
921 case nir_op_f2i8:
922 case nir_op_f2i16:
923 case nir_op_f2i32:
924 case nir_op_f2i64:
925 src[0] = ac_to_float(&ctx->ac, src[0]);
926 result = LLVMBuildFPToSI(ctx->ac.builder, src[0], def_type, "");
927 break;
928 case nir_op_f2u8:
929 case nir_op_f2u16:
930 case nir_op_f2u32:
931 case nir_op_f2u64:
932 src[0] = ac_to_float(&ctx->ac, src[0]);
933 result = LLVMBuildFPToUI(ctx->ac.builder, src[0], def_type, "");
934 break;
935 case nir_op_i2f16:
936 case nir_op_i2f32:
937 case nir_op_i2f64:
938 result = LLVMBuildSIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
939 break;
940 case nir_op_u2f16:
941 case nir_op_u2f32:
942 case nir_op_u2f64:
943 result = LLVMBuildUIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
944 break;
945 case nir_op_f2f16_rtz:
946 src[0] = ac_to_float(&ctx->ac, src[0]);
947 if (LLVMTypeOf(src[0]) == ctx->ac.f64)
948 src[0] = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ctx->ac.f32, "");
949 LLVMValueRef param[2] = { src[0], ctx->ac.f32_0 };
950 result = ac_build_cvt_pkrtz_f16(&ctx->ac, param);
951 result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
952 break;
953 case nir_op_f2f16_rtne:
954 case nir_op_f2f16:
955 case nir_op_f2f32:
956 case nir_op_f2f64:
957 src[0] = ac_to_float(&ctx->ac, src[0]);
958 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
959 result = LLVMBuildFPExt(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
960 else
961 result = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
962 break;
963 case nir_op_u2u8:
964 case nir_op_u2u16:
965 case nir_op_u2u32:
966 case nir_op_u2u64:
967 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
968 result = LLVMBuildZExt(ctx->ac.builder, src[0], def_type, "");
969 else
970 result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
971 break;
972 case nir_op_i2i8:
973 case nir_op_i2i16:
974 case nir_op_i2i32:
975 case nir_op_i2i64:
976 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
977 result = LLVMBuildSExt(ctx->ac.builder, src[0], def_type, "");
978 else
979 result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
980 break;
981 case nir_op_b32csel:
982 result = emit_bcsel(&ctx->ac, src[0], src[1], src[2]);
983 break;
984 case nir_op_find_lsb:
985 result = ac_find_lsb(&ctx->ac, ctx->ac.i32, src[0]);
986 break;
987 case nir_op_ufind_msb:
988 result = ac_build_umsb(&ctx->ac, src[0], ctx->ac.i32);
989 break;
990 case nir_op_ifind_msb:
991 result = ac_build_imsb(&ctx->ac, src[0], ctx->ac.i32);
992 break;
993 case nir_op_uadd_carry:
994 result = emit_uint_carry(&ctx->ac, "llvm.uadd.with.overflow.i32", src[0], src[1]);
995 break;
996 case nir_op_usub_borrow:
997 result = emit_uint_carry(&ctx->ac, "llvm.usub.with.overflow.i32", src[0], src[1]);
998 break;
999 case nir_op_b2f16:
1000 case nir_op_b2f32:
1001 case nir_op_b2f64:
1002 result = emit_b2f(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
1003 break;
1004 case nir_op_f2b32:
1005 result = emit_f2b(&ctx->ac, src[0]);
1006 break;
1007 case nir_op_b2i8:
1008 case nir_op_b2i16:
1009 case nir_op_b2i32:
1010 case nir_op_b2i64:
1011 result = emit_b2i(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
1012 break;
1013 case nir_op_i2b32:
1014 result = emit_i2b(&ctx->ac, src[0]);
1015 break;
1016 case nir_op_fquantize2f16:
1017 result = emit_f2f16(&ctx->ac, src[0]);
1018 break;
1019 case nir_op_umul_high:
1020 result = emit_umul_high(&ctx->ac, src[0], src[1]);
1021 break;
1022 case nir_op_imul_high:
1023 result = emit_imul_high(&ctx->ac, src[0], src[1]);
1024 break;
1025 case nir_op_pack_half_2x16:
1026 result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pkrtz_f16);
1027 break;
1028 case nir_op_pack_snorm_2x16:
1029 result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pknorm_i16);
1030 break;
1031 case nir_op_pack_unorm_2x16:
1032 result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pknorm_u16);
1033 break;
1034 case nir_op_unpack_half_2x16:
1035 result = emit_unpack_half_2x16(&ctx->ac, src[0]);
1036 break;
1037 case nir_op_fddx:
1038 case nir_op_fddy:
1039 case nir_op_fddx_fine:
1040 case nir_op_fddy_fine:
1041 case nir_op_fddx_coarse:
1042 case nir_op_fddy_coarse:
1043 result = emit_ddxy(ctx, instr->op, src[0]);
1044 break;
1045
1046 case nir_op_unpack_64_2x32_split_x: {
1047 assert(ac_get_llvm_num_components(src[0]) == 1);
1048 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1049 ctx->ac.v2i32,
1050 "");
1051 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1052 ctx->ac.i32_0, "");
1053 break;
1054 }
1055
1056 case nir_op_unpack_64_2x32_split_y: {
1057 assert(ac_get_llvm_num_components(src[0]) == 1);
1058 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1059 ctx->ac.v2i32,
1060 "");
1061 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1062 ctx->ac.i32_1, "");
1063 break;
1064 }
1065
1066 case nir_op_pack_64_2x32_split: {
1067 LLVMValueRef tmp = ac_build_gather_values(&ctx->ac, src, 2);
1068 result = LLVMBuildBitCast(ctx->ac.builder, tmp, ctx->ac.i64, "");
1069 break;
1070 }
1071
1072 case nir_op_pack_32_2x16_split: {
1073 LLVMValueRef tmp = ac_build_gather_values(&ctx->ac, src, 2);
1074 result = LLVMBuildBitCast(ctx->ac.builder, tmp, ctx->ac.i32, "");
1075 break;
1076 }
1077
1078 case nir_op_unpack_32_2x16_split_x: {
1079 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1080 ctx->ac.v2i16,
1081 "");
1082 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1083 ctx->ac.i32_0, "");
1084 break;
1085 }
1086
1087 case nir_op_unpack_32_2x16_split_y: {
1088 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1089 ctx->ac.v2i16,
1090 "");
1091 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1092 ctx->ac.i32_1, "");
1093 break;
1094 }
1095
1096 case nir_op_cube_face_coord: {
1097 src[0] = ac_to_float(&ctx->ac, src[0]);
1098 LLVMValueRef results[2];
1099 LLVMValueRef in[3];
1100 for (unsigned chan = 0; chan < 3; chan++)
1101 in[chan] = ac_llvm_extract_elem(&ctx->ac, src[0], chan);
1102 results[0] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubesc",
1103 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1104 results[1] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubetc",
1105 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1106 LLVMValueRef ma = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubema",
1107 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1108 results[0] = ac_build_fdiv(&ctx->ac, results[0], ma);
1109 results[1] = ac_build_fdiv(&ctx->ac, results[1], ma);
1110 LLVMValueRef offset = LLVMConstReal(ctx->ac.f32, 0.5);
1111 results[0] = LLVMBuildFAdd(ctx->ac.builder, results[0], offset, "");
1112 results[1] = LLVMBuildFAdd(ctx->ac.builder, results[1], offset, "");
1113 result = ac_build_gather_values(&ctx->ac, results, 2);
1114 break;
1115 }
1116
1117 case nir_op_cube_face_index: {
1118 src[0] = ac_to_float(&ctx->ac, src[0]);
1119 LLVMValueRef in[3];
1120 for (unsigned chan = 0; chan < 3; chan++)
1121 in[chan] = ac_llvm_extract_elem(&ctx->ac, src[0], chan);
1122 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubeid",
1123 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1124 break;
1125 }
1126
1127 case nir_op_fmin3:
1128 result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
1129 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
1130 result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
1131 ac_to_float_type(&ctx->ac, def_type), result, src[2]);
1132 break;
1133 case nir_op_umin3:
1134 result = ac_build_umin(&ctx->ac, src[0], src[1]);
1135 result = ac_build_umin(&ctx->ac, result, src[2]);
1136 break;
1137 case nir_op_imin3:
1138 result = ac_build_imin(&ctx->ac, src[0], src[1]);
1139 result = ac_build_imin(&ctx->ac, result, src[2]);
1140 break;
1141 case nir_op_fmax3:
1142 result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
1143 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
1144 result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
1145 ac_to_float_type(&ctx->ac, def_type), result, src[2]);
1146 break;
1147 case nir_op_umax3:
1148 result = ac_build_umax(&ctx->ac, src[0], src[1]);
1149 result = ac_build_umax(&ctx->ac, result, src[2]);
1150 break;
1151 case nir_op_imax3:
1152 result = ac_build_imax(&ctx->ac, src[0], src[1]);
1153 result = ac_build_imax(&ctx->ac, result, src[2]);
1154 break;
1155 case nir_op_fmed3: {
1156 src[0] = ac_to_float(&ctx->ac, src[0]);
1157 src[1] = ac_to_float(&ctx->ac, src[1]);
1158 src[2] = ac_to_float(&ctx->ac, src[2]);
1159 result = ac_build_fmed3(&ctx->ac, src[0], src[1], src[2],
1160 instr->dest.dest.ssa.bit_size);
1161 break;
1162 }
1163 case nir_op_imed3: {
1164 LLVMValueRef tmp1 = ac_build_imin(&ctx->ac, src[0], src[1]);
1165 LLVMValueRef tmp2 = ac_build_imax(&ctx->ac, src[0], src[1]);
1166 tmp2 = ac_build_imin(&ctx->ac, tmp2, src[2]);
1167 result = ac_build_imax(&ctx->ac, tmp1, tmp2);
1168 break;
1169 }
1170 case nir_op_umed3: {
1171 LLVMValueRef tmp1 = ac_build_umin(&ctx->ac, src[0], src[1]);
1172 LLVMValueRef tmp2 = ac_build_umax(&ctx->ac, src[0], src[1]);
1173 tmp2 = ac_build_umin(&ctx->ac, tmp2, src[2]);
1174 result = ac_build_umax(&ctx->ac, tmp1, tmp2);
1175 break;
1176 }
1177
1178 default:
1179 fprintf(stderr, "Unknown NIR alu instr: ");
1180 nir_print_instr(&instr->instr, stderr);
1181 fprintf(stderr, "\n");
1182 abort();
1183 }
1184
1185 if (result) {
1186 assert(instr->dest.dest.is_ssa);
1187 result = ac_to_integer_or_pointer(&ctx->ac, result);
1188 ctx->ssa_defs[instr->dest.dest.ssa.index] = result;
1189 }
1190
1191 if (instr->exact)
1192 ac_restore_inexact_math(ctx->ac.builder, saved_inexact);
1193 }
1194
1195 static void visit_load_const(struct ac_nir_context *ctx,
1196 const nir_load_const_instr *instr)
1197 {
1198 LLVMValueRef values[4], value = NULL;
1199 LLVMTypeRef element_type =
1200 LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
1201
1202 for (unsigned i = 0; i < instr->def.num_components; ++i) {
1203 switch (instr->def.bit_size) {
1204 case 8:
1205 values[i] = LLVMConstInt(element_type,
1206 instr->value[i].u8, false);
1207 break;
1208 case 16:
1209 values[i] = LLVMConstInt(element_type,
1210 instr->value[i].u16, false);
1211 break;
1212 case 32:
1213 values[i] = LLVMConstInt(element_type,
1214 instr->value[i].u32, false);
1215 break;
1216 case 64:
1217 values[i] = LLVMConstInt(element_type,
1218 instr->value[i].u64, false);
1219 break;
1220 default:
1221 fprintf(stderr,
1222 "unsupported nir load_const bit_size: %d\n",
1223 instr->def.bit_size);
1224 abort();
1225 }
1226 }
1227 if (instr->def.num_components > 1) {
1228 value = LLVMConstVector(values, instr->def.num_components);
1229 } else
1230 value = values[0];
1231
1232 ctx->ssa_defs[instr->def.index] = value;
1233 }
1234
1235 static LLVMValueRef
1236 get_buffer_size(struct ac_nir_context *ctx, LLVMValueRef descriptor, bool in_elements)
1237 {
1238 LLVMValueRef size =
1239 LLVMBuildExtractElement(ctx->ac.builder, descriptor,
1240 LLVMConstInt(ctx->ac.i32, 2, false), "");
1241
1242 /* GFX8 only */
1243 if (ctx->ac.chip_class == GFX8 && in_elements) {
1244 /* On GFX8, the descriptor contains the size in bytes,
1245 * but TXQ must return the size in elements.
1246 * The stride is always non-zero for resources using TXQ.
1247 */
1248 LLVMValueRef stride =
1249 LLVMBuildExtractElement(ctx->ac.builder, descriptor,
1250 ctx->ac.i32_1, "");
1251 stride = LLVMBuildLShr(ctx->ac.builder, stride,
1252 LLVMConstInt(ctx->ac.i32, 16, false), "");
1253 stride = LLVMBuildAnd(ctx->ac.builder, stride,
1254 LLVMConstInt(ctx->ac.i32, 0x3fff, false), "");
1255
1256 size = LLVMBuildUDiv(ctx->ac.builder, size, stride, "");
1257 }
1258 return size;
1259 }
1260
1261 /* Gather4 should follow the same rules as bilinear filtering, but the hardware
1262 * incorrectly forces nearest filtering if the texture format is integer.
1263 * The only effect it has on Gather4, which always returns 4 texels for
1264 * bilinear filtering, is that the final coordinates are off by 0.5 of
1265 * the texel size.
1266 *
1267 * The workaround is to subtract 0.5 from the unnormalized coordinates,
1268 * or (0.5 / size) from the normalized coordinates.
1269 *
1270 * However, cube textures with 8_8_8_8 data formats require a different
1271 * workaround of overriding the num format to USCALED/SSCALED. This would lose
1272 * precision in 32-bit data formats, so it needs to be applied dynamically at
1273 * runtime. In this case, return an i1 value that indicates whether the
1274 * descriptor was overridden (and hence a fixup of the sampler result is needed).
1275 */
1276 static LLVMValueRef lower_gather4_integer(struct ac_llvm_context *ctx,
1277 nir_variable *var,
1278 struct ac_image_args *args,
1279 const nir_tex_instr *instr)
1280 {
1281 const struct glsl_type *type = glsl_without_array(var->type);
1282 enum glsl_base_type stype = glsl_get_sampler_result_type(type);
1283 LLVMValueRef wa_8888 = NULL;
1284 LLVMValueRef half_texel[2];
1285 LLVMValueRef result;
1286
1287 assert(stype == GLSL_TYPE_INT || stype == GLSL_TYPE_UINT);
1288
1289 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
1290 LLVMValueRef formats;
1291 LLVMValueRef data_format;
1292 LLVMValueRef wa_formats;
1293
1294 formats = LLVMBuildExtractElement(ctx->builder, args->resource, ctx->i32_1, "");
1295
1296 data_format = LLVMBuildLShr(ctx->builder, formats,
1297 LLVMConstInt(ctx->i32, 20, false), "");
1298 data_format = LLVMBuildAnd(ctx->builder, data_format,
1299 LLVMConstInt(ctx->i32, (1u << 6) - 1, false), "");
1300 wa_8888 = LLVMBuildICmp(
1301 ctx->builder, LLVMIntEQ, data_format,
1302 LLVMConstInt(ctx->i32, V_008F14_IMG_DATA_FORMAT_8_8_8_8, false),
1303 "");
1304
1305 uint32_t wa_num_format =
1306 stype == GLSL_TYPE_UINT ?
1307 S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_USCALED) :
1308 S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_SSCALED);
1309 wa_formats = LLVMBuildAnd(ctx->builder, formats,
1310 LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT, false),
1311 "");
1312 wa_formats = LLVMBuildOr(ctx->builder, wa_formats,
1313 LLVMConstInt(ctx->i32, wa_num_format, false), "");
1314
1315 formats = LLVMBuildSelect(ctx->builder, wa_8888, wa_formats, formats, "");
1316 args->resource = LLVMBuildInsertElement(
1317 ctx->builder, args->resource, formats, ctx->i32_1, "");
1318 }
1319
1320 if (instr->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
1321 assert(!wa_8888);
1322 half_texel[0] = half_texel[1] = LLVMConstReal(ctx->f32, -0.5);
1323 } else {
1324 struct ac_image_args resinfo = {};
1325 LLVMBasicBlockRef bbs[2];
1326
1327 LLVMValueRef unnorm = NULL;
1328 LLVMValueRef default_offset = ctx->f32_0;
1329 if (instr->sampler_dim == GLSL_SAMPLER_DIM_2D &&
1330 !instr->is_array) {
1331 /* In vulkan, whether the sampler uses unnormalized
1332 * coordinates or not is a dynamic property of the
1333 * sampler. Hence, to figure out whether or not we
1334 * need to divide by the texture size, we need to test
1335 * the sampler at runtime. This tests the bit set by
1336 * radv_init_sampler().
1337 */
1338 LLVMValueRef sampler0 =
1339 LLVMBuildExtractElement(ctx->builder, args->sampler, ctx->i32_0, "");
1340 sampler0 = LLVMBuildLShr(ctx->builder, sampler0,
1341 LLVMConstInt(ctx->i32, 15, false), "");
1342 sampler0 = LLVMBuildAnd(ctx->builder, sampler0, ctx->i32_1, "");
1343 unnorm = LLVMBuildICmp(ctx->builder, LLVMIntEQ, sampler0, ctx->i32_1, "");
1344 default_offset = LLVMConstReal(ctx->f32, -0.5);
1345 }
1346
1347 bbs[0] = LLVMGetInsertBlock(ctx->builder);
1348 if (wa_8888 || unnorm) {
1349 assert(!(wa_8888 && unnorm));
1350 LLVMValueRef not_needed = wa_8888 ? wa_8888 : unnorm;
1351 /* Skip the texture size query entirely if we don't need it. */
1352 ac_build_ifcc(ctx, LLVMBuildNot(ctx->builder, not_needed, ""), 2000);
1353 bbs[1] = LLVMGetInsertBlock(ctx->builder);
1354 }
1355
1356 /* Query the texture size. */
1357 resinfo.dim = ac_get_sampler_dim(ctx->chip_class, instr->sampler_dim, instr->is_array);
1358 resinfo.opcode = ac_image_get_resinfo;
1359 resinfo.dmask = 0xf;
1360 resinfo.lod = ctx->i32_0;
1361 resinfo.resource = args->resource;
1362 resinfo.attributes = AC_FUNC_ATTR_READNONE;
1363 LLVMValueRef size = ac_build_image_opcode(ctx, &resinfo);
1364
1365 /* Compute -0.5 / size. */
1366 for (unsigned c = 0; c < 2; c++) {
1367 half_texel[c] =
1368 LLVMBuildExtractElement(ctx->builder, size,
1369 LLVMConstInt(ctx->i32, c, 0), "");
1370 half_texel[c] = LLVMBuildUIToFP(ctx->builder, half_texel[c], ctx->f32, "");
1371 half_texel[c] = ac_build_fdiv(ctx, ctx->f32_1, half_texel[c]);
1372 half_texel[c] = LLVMBuildFMul(ctx->builder, half_texel[c],
1373 LLVMConstReal(ctx->f32, -0.5), "");
1374 }
1375
1376 if (wa_8888 || unnorm) {
1377 ac_build_endif(ctx, 2000);
1378
1379 for (unsigned c = 0; c < 2; c++) {
1380 LLVMValueRef values[2] = { default_offset, half_texel[c] };
1381 half_texel[c] = ac_build_phi(ctx, ctx->f32, 2,
1382 values, bbs);
1383 }
1384 }
1385 }
1386
1387 for (unsigned c = 0; c < 2; c++) {
1388 LLVMValueRef tmp;
1389 tmp = LLVMBuildBitCast(ctx->builder, args->coords[c], ctx->f32, "");
1390 args->coords[c] = LLVMBuildFAdd(ctx->builder, tmp, half_texel[c], "");
1391 }
1392
1393 args->attributes = AC_FUNC_ATTR_READNONE;
1394 result = ac_build_image_opcode(ctx, args);
1395
1396 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
1397 LLVMValueRef tmp, tmp2;
1398
1399 /* if the cube workaround is in place, f2i the result. */
1400 for (unsigned c = 0; c < 4; c++) {
1401 tmp = LLVMBuildExtractElement(ctx->builder, result, LLVMConstInt(ctx->i32, c, false), "");
1402 if (stype == GLSL_TYPE_UINT)
1403 tmp2 = LLVMBuildFPToUI(ctx->builder, tmp, ctx->i32, "");
1404 else
1405 tmp2 = LLVMBuildFPToSI(ctx->builder, tmp, ctx->i32, "");
1406 tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->i32, "");
1407 tmp2 = LLVMBuildBitCast(ctx->builder, tmp2, ctx->i32, "");
1408 tmp = LLVMBuildSelect(ctx->builder, wa_8888, tmp2, tmp, "");
1409 tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->f32, "");
1410 result = LLVMBuildInsertElement(ctx->builder, result, tmp, LLVMConstInt(ctx->i32, c, false), "");
1411 }
1412 }
1413 return result;
1414 }
1415
1416 static nir_deref_instr *get_tex_texture_deref(const nir_tex_instr *instr)
1417 {
1418 nir_deref_instr *texture_deref_instr = NULL;
1419
1420 for (unsigned i = 0; i < instr->num_srcs; i++) {
1421 switch (instr->src[i].src_type) {
1422 case nir_tex_src_texture_deref:
1423 texture_deref_instr = nir_src_as_deref(instr->src[i].src);
1424 break;
1425 default:
1426 break;
1427 }
1428 }
1429 return texture_deref_instr;
1430 }
1431
1432 static LLVMValueRef build_tex_intrinsic(struct ac_nir_context *ctx,
1433 const nir_tex_instr *instr,
1434 struct ac_image_args *args)
1435 {
1436 if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
1437 unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
1438
1439 assert(instr->dest.is_ssa);
1440 return ac_build_buffer_load_format(&ctx->ac,
1441 args->resource,
1442 args->coords[0],
1443 ctx->ac.i32_0,
1444 util_last_bit(mask),
1445 0, true,
1446 instr->dest.ssa.bit_size == 16);
1447 }
1448
1449 args->opcode = ac_image_sample;
1450
1451 switch (instr->op) {
1452 case nir_texop_txf:
1453 case nir_texop_txf_ms:
1454 case nir_texop_samples_identical:
1455 args->opcode = args->level_zero ||
1456 instr->sampler_dim == GLSL_SAMPLER_DIM_MS ?
1457 ac_image_load : ac_image_load_mip;
1458 args->level_zero = false;
1459 break;
1460 case nir_texop_txs:
1461 case nir_texop_query_levels:
1462 args->opcode = ac_image_get_resinfo;
1463 if (!args->lod)
1464 args->lod = ctx->ac.i32_0;
1465 args->level_zero = false;
1466 break;
1467 case nir_texop_tex:
1468 if (ctx->stage != MESA_SHADER_FRAGMENT) {
1469 assert(!args->lod);
1470 args->level_zero = true;
1471 }
1472 break;
1473 case nir_texop_tg4:
1474 args->opcode = ac_image_gather4;
1475 if (!args->lod && !args->bias)
1476 args->level_zero = true;
1477 break;
1478 case nir_texop_lod:
1479 args->opcode = ac_image_get_lod;
1480 break;
1481 case nir_texop_fragment_fetch:
1482 case nir_texop_fragment_mask_fetch:
1483 args->opcode = ac_image_load;
1484 args->level_zero = false;
1485 break;
1486 default:
1487 break;
1488 }
1489
1490 if (instr->op == nir_texop_tg4 && ctx->ac.chip_class <= GFX8) {
1491 nir_deref_instr *texture_deref_instr = get_tex_texture_deref(instr);
1492 nir_variable *var = nir_deref_instr_get_variable(texture_deref_instr);
1493 const struct glsl_type *type = glsl_without_array(var->type);
1494 enum glsl_base_type stype = glsl_get_sampler_result_type(type);
1495 if (stype == GLSL_TYPE_UINT || stype == GLSL_TYPE_INT) {
1496 return lower_gather4_integer(&ctx->ac, var, args, instr);
1497 }
1498 }
1499
1500 /* Fixup for GFX9 which allocates 1D textures as 2D. */
1501 if (instr->op == nir_texop_lod && ctx->ac.chip_class == GFX9) {
1502 if ((args->dim == ac_image_2darray ||
1503 args->dim == ac_image_2d) && !args->coords[1]) {
1504 args->coords[1] = ctx->ac.i32_0;
1505 }
1506 }
1507
1508 args->attributes = AC_FUNC_ATTR_READNONE;
1509 bool cs_derivs = ctx->stage == MESA_SHADER_COMPUTE &&
1510 ctx->info->cs.derivative_group != DERIVATIVE_GROUP_NONE;
1511 if (ctx->stage == MESA_SHADER_FRAGMENT || cs_derivs) {
1512 /* Prevent texture instructions with implicit derivatives from being
1513 * sinked into branches. */
1514 switch (instr->op) {
1515 case nir_texop_tex:
1516 case nir_texop_txb:
1517 case nir_texop_lod:
1518 args->attributes |= AC_FUNC_ATTR_CONVERGENT;
1519 break;
1520 default:
1521 break;
1522 }
1523 }
1524
1525 return ac_build_image_opcode(&ctx->ac, args);
1526 }
1527
1528 static LLVMValueRef visit_vulkan_resource_reindex(struct ac_nir_context *ctx,
1529 nir_intrinsic_instr *instr)
1530 {
1531 LLVMValueRef ptr = get_src(ctx, instr->src[0]);
1532 LLVMValueRef index = get_src(ctx, instr->src[1]);
1533
1534 LLVMValueRef result = LLVMBuildGEP(ctx->ac.builder, ptr, &index, 1, "");
1535 LLVMSetMetadata(result, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
1536 return result;
1537 }
1538
1539 static LLVMValueRef visit_load_push_constant(struct ac_nir_context *ctx,
1540 nir_intrinsic_instr *instr)
1541 {
1542 LLVMValueRef ptr, addr;
1543 LLVMValueRef src0 = get_src(ctx, instr->src[0]);
1544 unsigned index = nir_intrinsic_base(instr);
1545
1546 addr = LLVMConstInt(ctx->ac.i32, index, 0);
1547 addr = LLVMBuildAdd(ctx->ac.builder, addr, src0, "");
1548
1549 /* Load constant values from user SGPRS when possible, otherwise
1550 * fallback to the default path that loads directly from memory.
1551 */
1552 if (LLVMIsConstant(src0) &&
1553 instr->dest.ssa.bit_size == 32) {
1554 unsigned count = instr->dest.ssa.num_components;
1555 unsigned offset = index;
1556
1557 offset += LLVMConstIntGetZExtValue(src0);
1558 offset /= 4;
1559
1560 offset -= ctx->args->base_inline_push_consts;
1561
1562 unsigned num_inline_push_consts = ctx->args->num_inline_push_consts;
1563 if (offset + count <= num_inline_push_consts) {
1564 LLVMValueRef push_constants[num_inline_push_consts];
1565 for (unsigned i = 0; i < num_inline_push_consts; i++)
1566 push_constants[i] = ac_get_arg(&ctx->ac,
1567 ctx->args->inline_push_consts[i]);
1568 return ac_build_gather_values(&ctx->ac,
1569 push_constants + offset,
1570 count);
1571 }
1572 }
1573
1574 ptr = LLVMBuildGEP(ctx->ac.builder,
1575 ac_get_arg(&ctx->ac, ctx->args->push_constants), &addr, 1, "");
1576
1577 if (instr->dest.ssa.bit_size == 8) {
1578 unsigned load_dwords = instr->dest.ssa.num_components > 1 ? 2 : 1;
1579 LLVMTypeRef vec_type = LLVMVectorType(ctx->ac.i8, 4 * load_dwords);
1580 ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
1581 LLVMValueRef res = LLVMBuildLoad(ctx->ac.builder, ptr, "");
1582
1583 LLVMValueRef params[3];
1584 if (load_dwords > 1) {
1585 LLVMValueRef res_vec = LLVMBuildBitCast(ctx->ac.builder, res, ctx->ac.v2i32, "");
1586 params[0] = LLVMBuildExtractElement(ctx->ac.builder, res_vec, LLVMConstInt(ctx->ac.i32, 1, false), "");
1587 params[1] = LLVMBuildExtractElement(ctx->ac.builder, res_vec, LLVMConstInt(ctx->ac.i32, 0, false), "");
1588 } else {
1589 res = LLVMBuildBitCast(ctx->ac.builder, res, ctx->ac.i32, "");
1590 params[0] = ctx->ac.i32_0;
1591 params[1] = res;
1592 }
1593 params[2] = addr;
1594 res = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.alignbyte", ctx->ac.i32, params, 3, 0);
1595
1596 res = LLVMBuildTrunc(ctx->ac.builder, res, LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.num_components * 8), "");
1597 if (instr->dest.ssa.num_components > 1)
1598 res = LLVMBuildBitCast(ctx->ac.builder, res, LLVMVectorType(ctx->ac.i8, instr->dest.ssa.num_components), "");
1599 return res;
1600 } else if (instr->dest.ssa.bit_size == 16) {
1601 unsigned load_dwords = instr->dest.ssa.num_components / 2 + 1;
1602 LLVMTypeRef vec_type = LLVMVectorType(ctx->ac.i16, 2 * load_dwords);
1603 ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
1604 LLVMValueRef res = LLVMBuildLoad(ctx->ac.builder, ptr, "");
1605 res = LLVMBuildBitCast(ctx->ac.builder, res, vec_type, "");
1606 LLVMValueRef cond = LLVMBuildLShr(ctx->ac.builder, addr, ctx->ac.i32_1, "");
1607 cond = LLVMBuildTrunc(ctx->ac.builder, cond, ctx->ac.i1, "");
1608 LLVMValueRef mask[] = { LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
1609 LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
1610 LLVMConstInt(ctx->ac.i32, 4, false)};
1611 LLVMValueRef swizzle_aligned = LLVMConstVector(&mask[0], instr->dest.ssa.num_components);
1612 LLVMValueRef swizzle_unaligned = LLVMConstVector(&mask[1], instr->dest.ssa.num_components);
1613 LLVMValueRef shuffle_aligned = LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_aligned, "");
1614 LLVMValueRef shuffle_unaligned = LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_unaligned, "");
1615 res = LLVMBuildSelect(ctx->ac.builder, cond, shuffle_unaligned, shuffle_aligned, "");
1616 return LLVMBuildBitCast(ctx->ac.builder, res, get_def_type(ctx, &instr->dest.ssa), "");
1617 }
1618
1619 ptr = ac_cast_ptr(&ctx->ac, ptr, get_def_type(ctx, &instr->dest.ssa));
1620
1621 return LLVMBuildLoad(ctx->ac.builder, ptr, "");
1622 }
1623
1624 static LLVMValueRef visit_get_buffer_size(struct ac_nir_context *ctx,
1625 const nir_intrinsic_instr *instr)
1626 {
1627 LLVMValueRef index = get_src(ctx, instr->src[0]);
1628
1629 return get_buffer_size(ctx, ctx->abi->load_ssbo(ctx->abi, index, false), false);
1630 }
1631
1632 static uint32_t widen_mask(uint32_t mask, unsigned multiplier)
1633 {
1634 uint32_t new_mask = 0;
1635 for(unsigned i = 0; i < 32 && (1u << i) <= mask; ++i)
1636 if (mask & (1u << i))
1637 new_mask |= ((1u << multiplier) - 1u) << (i * multiplier);
1638 return new_mask;
1639 }
1640
1641 static LLVMValueRef extract_vector_range(struct ac_llvm_context *ctx, LLVMValueRef src,
1642 unsigned start, unsigned count)
1643 {
1644 LLVMValueRef mask[] = {
1645 ctx->i32_0, ctx->i32_1,
1646 LLVMConstInt(ctx->i32, 2, false), LLVMConstInt(ctx->i32, 3, false) };
1647
1648 unsigned src_elements = ac_get_llvm_num_components(src);
1649
1650 if (count == src_elements) {
1651 assert(start == 0);
1652 return src;
1653 } else if (count == 1) {
1654 assert(start < src_elements);
1655 return LLVMBuildExtractElement(ctx->builder, src, mask[start], "");
1656 } else {
1657 assert(start + count <= src_elements);
1658 assert(count <= 4);
1659 LLVMValueRef swizzle = LLVMConstVector(&mask[start], count);
1660 return LLVMBuildShuffleVector(ctx->builder, src, src, swizzle, "");
1661 }
1662 }
1663
1664 static unsigned get_cache_policy(struct ac_nir_context *ctx,
1665 enum gl_access_qualifier access,
1666 bool may_store_unaligned,
1667 bool writeonly_memory)
1668 {
1669 unsigned cache_policy = 0;
1670
1671 /* GFX6 has a TC L1 bug causing corruption of 8bit/16bit stores. All
1672 * store opcodes not aligned to a dword are affected. The only way to
1673 * get unaligned stores is through shader images.
1674 */
1675 if (((may_store_unaligned && ctx->ac.chip_class == GFX6) ||
1676 /* If this is write-only, don't keep data in L1 to prevent
1677 * evicting L1 cache lines that may be needed by other
1678 * instructions.
1679 */
1680 writeonly_memory ||
1681 access & (ACCESS_COHERENT | ACCESS_VOLATILE))) {
1682 cache_policy |= ac_glc;
1683 }
1684
1685 if (access & ACCESS_STREAM_CACHE_POLICY)
1686 cache_policy |= ac_slc | ac_glc;
1687
1688 return cache_policy;
1689 }
1690
1691 static LLVMValueRef enter_waterfall_ssbo(struct ac_nir_context *ctx,
1692 struct waterfall_context *wctx,
1693 const nir_intrinsic_instr *instr,
1694 nir_src src)
1695 {
1696 return enter_waterfall(ctx, wctx, get_src(ctx, src),
1697 nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
1698 }
1699
1700 static void visit_store_ssbo(struct ac_nir_context *ctx,
1701 nir_intrinsic_instr *instr)
1702 {
1703 if (ctx->ac.postponed_kill) {
1704 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
1705 ctx->ac.postponed_kill, "");
1706 ac_build_ifcc(&ctx->ac, cond, 7000);
1707 }
1708
1709 LLVMValueRef src_data = get_src(ctx, instr->src[0]);
1710 int elem_size_bytes = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src_data)) / 8;
1711 unsigned writemask = nir_intrinsic_write_mask(instr);
1712 enum gl_access_qualifier access = nir_intrinsic_access(instr);
1713 bool writeonly_memory = access & ACCESS_NON_READABLE;
1714 unsigned cache_policy = get_cache_policy(ctx, access, false, writeonly_memory);
1715
1716 struct waterfall_context wctx;
1717 LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[1]);
1718
1719 LLVMValueRef rsrc = ctx->abi->load_ssbo(ctx->abi, rsrc_base, true);
1720 LLVMValueRef base_data = src_data;
1721 base_data = ac_trim_vector(&ctx->ac, base_data, instr->num_components);
1722 LLVMValueRef base_offset = get_src(ctx, instr->src[2]);
1723
1724 while (writemask) {
1725 int start, count;
1726 LLVMValueRef data, offset;
1727 LLVMTypeRef data_type;
1728
1729 u_bit_scan_consecutive_range(&writemask, &start, &count);
1730
1731 /* Due to an LLVM limitation with LLVM < 9, split 3-element
1732 * writes into a 2-element and a 1-element write. */
1733 if (count == 3 &&
1734 (elem_size_bytes != 4 || !ac_has_vec3_support(ctx->ac.chip_class, false))) {
1735 writemask |= 1 << (start + 2);
1736 count = 2;
1737 }
1738 int num_bytes = count * elem_size_bytes; /* count in bytes */
1739
1740 /* we can only store 4 DWords at the same time.
1741 * can only happen for 64 Bit vectors. */
1742 if (num_bytes > 16) {
1743 writemask |= ((1u << (count - 2)) - 1u) << (start + 2);
1744 count = 2;
1745 num_bytes = 16;
1746 }
1747
1748 /* check alignment of 16 Bit stores */
1749 if (elem_size_bytes == 2 && num_bytes > 2 && (start % 2) == 1) {
1750 writemask |= ((1u << (count - 1)) - 1u) << (start + 1);
1751 count = 1;
1752 num_bytes = 2;
1753 }
1754
1755 /* Due to alignment issues, split stores of 8-bit/16-bit
1756 * vectors.
1757 */
1758 if (ctx->ac.chip_class == GFX6 && count > 1 && elem_size_bytes < 4) {
1759 writemask |= ((1u << (count - 1)) - 1u) << (start + 1);
1760 count = 1;
1761 num_bytes = elem_size_bytes;
1762 }
1763
1764 data = extract_vector_range(&ctx->ac, base_data, start, count);
1765
1766 offset = LLVMBuildAdd(ctx->ac.builder, base_offset,
1767 LLVMConstInt(ctx->ac.i32, start * elem_size_bytes, false), "");
1768
1769 if (num_bytes == 1) {
1770 ac_build_tbuffer_store_byte(&ctx->ac, rsrc, data,
1771 offset, ctx->ac.i32_0,
1772 cache_policy);
1773 } else if (num_bytes == 2) {
1774 ac_build_tbuffer_store_short(&ctx->ac, rsrc, data,
1775 offset, ctx->ac.i32_0,
1776 cache_policy);
1777 } else {
1778 int num_channels = num_bytes / 4;
1779
1780 switch (num_bytes) {
1781 case 16: /* v4f32 */
1782 data_type = ctx->ac.v4f32;
1783 break;
1784 case 12: /* v3f32 */
1785 data_type = ctx->ac.v3f32;
1786 break;
1787 case 8: /* v2f32 */
1788 data_type = ctx->ac.v2f32;
1789 break;
1790 case 4: /* f32 */
1791 data_type = ctx->ac.f32;
1792 break;
1793 default:
1794 unreachable("Malformed vector store.");
1795 }
1796 data = LLVMBuildBitCast(ctx->ac.builder, data, data_type, "");
1797
1798 ac_build_buffer_store_dword(&ctx->ac, rsrc, data,
1799 num_channels, offset,
1800 ctx->ac.i32_0, 0,
1801 cache_policy);
1802 }
1803 }
1804
1805 exit_waterfall(ctx, &wctx, NULL);
1806
1807 if (ctx->ac.postponed_kill)
1808 ac_build_endif(&ctx->ac, 7000);
1809 }
1810
1811 static LLVMValueRef emit_ssbo_comp_swap_64(struct ac_nir_context *ctx,
1812 LLVMValueRef descriptor,
1813 LLVMValueRef offset,
1814 LLVMValueRef compare,
1815 LLVMValueRef exchange)
1816 {
1817 LLVMBasicBlockRef start_block = NULL, then_block = NULL;
1818 if (ctx->abi->robust_buffer_access) {
1819 LLVMValueRef size = ac_llvm_extract_elem(&ctx->ac, descriptor, 2);
1820
1821 LLVMValueRef cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, offset, size, "");
1822 start_block = LLVMGetInsertBlock(ctx->ac.builder);
1823
1824 ac_build_ifcc(&ctx->ac, cond, -1);
1825
1826 then_block = LLVMGetInsertBlock(ctx->ac.builder);
1827 }
1828
1829 LLVMValueRef ptr_parts[2] = {
1830 ac_llvm_extract_elem(&ctx->ac, descriptor, 0),
1831 LLVMBuildAnd(ctx->ac.builder,
1832 ac_llvm_extract_elem(&ctx->ac, descriptor, 1),
1833 LLVMConstInt(ctx->ac.i32, 65535, 0), "")
1834 };
1835
1836 ptr_parts[1] = LLVMBuildTrunc(ctx->ac.builder, ptr_parts[1], ctx->ac.i16, "");
1837 ptr_parts[1] = LLVMBuildSExt(ctx->ac.builder, ptr_parts[1], ctx->ac.i32, "");
1838
1839 offset = LLVMBuildZExt(ctx->ac.builder, offset, ctx->ac.i64, "");
1840
1841 LLVMValueRef ptr = ac_build_gather_values(&ctx->ac, ptr_parts, 2);
1842 ptr = LLVMBuildBitCast(ctx->ac.builder, ptr, ctx->ac.i64, "");
1843 ptr = LLVMBuildAdd(ctx->ac.builder, ptr, offset, "");
1844 ptr = LLVMBuildIntToPtr(ctx->ac.builder, ptr, LLVMPointerType(ctx->ac.i64, AC_ADDR_SPACE_GLOBAL), "");
1845
1846 LLVMValueRef result = ac_build_atomic_cmp_xchg(&ctx->ac, ptr, compare, exchange, "singlethread-one-as");
1847 result = LLVMBuildExtractValue(ctx->ac.builder, result, 0, "");
1848
1849 if (ctx->abi->robust_buffer_access) {
1850 ac_build_endif(&ctx->ac, -1);
1851
1852 LLVMBasicBlockRef incoming_blocks[2] = {
1853 start_block,
1854 then_block,
1855 };
1856
1857 LLVMValueRef incoming_values[2] = {
1858 LLVMConstInt(ctx->ac.i64, 0, 0),
1859 result,
1860 };
1861 LLVMValueRef ret = LLVMBuildPhi(ctx->ac.builder, ctx->ac.i64, "");
1862 LLVMAddIncoming(ret, incoming_values, incoming_blocks, 2);
1863 return ret;
1864 } else {
1865 return result;
1866 }
1867 }
1868
1869 static LLVMValueRef visit_atomic_ssbo(struct ac_nir_context *ctx,
1870 nir_intrinsic_instr *instr)
1871 {
1872 if (ctx->ac.postponed_kill) {
1873 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
1874 ctx->ac.postponed_kill, "");
1875 ac_build_ifcc(&ctx->ac, cond, 7001);
1876 }
1877
1878 LLVMTypeRef return_type = LLVMTypeOf(get_src(ctx, instr->src[2]));
1879 const char *op;
1880 char name[64], type[8];
1881 LLVMValueRef params[6], descriptor;
1882 LLVMValueRef result;
1883 int arg_count = 0;
1884
1885 struct waterfall_context wctx;
1886 LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[0]);
1887
1888 switch (instr->intrinsic) {
1889 case nir_intrinsic_ssbo_atomic_add:
1890 op = "add";
1891 break;
1892 case nir_intrinsic_ssbo_atomic_imin:
1893 op = "smin";
1894 break;
1895 case nir_intrinsic_ssbo_atomic_umin:
1896 op = "umin";
1897 break;
1898 case nir_intrinsic_ssbo_atomic_imax:
1899 op = "smax";
1900 break;
1901 case nir_intrinsic_ssbo_atomic_umax:
1902 op = "umax";
1903 break;
1904 case nir_intrinsic_ssbo_atomic_and:
1905 op = "and";
1906 break;
1907 case nir_intrinsic_ssbo_atomic_or:
1908 op = "or";
1909 break;
1910 case nir_intrinsic_ssbo_atomic_xor:
1911 op = "xor";
1912 break;
1913 case nir_intrinsic_ssbo_atomic_exchange:
1914 op = "swap";
1915 break;
1916 case nir_intrinsic_ssbo_atomic_comp_swap:
1917 op = "cmpswap";
1918 break;
1919 default:
1920 abort();
1921 }
1922
1923 descriptor = ctx->abi->load_ssbo(ctx->abi,
1924 rsrc_base,
1925 true);
1926
1927 if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap &&
1928 return_type == ctx->ac.i64) {
1929 result = emit_ssbo_comp_swap_64(ctx, descriptor,
1930 get_src(ctx, instr->src[1]),
1931 get_src(ctx, instr->src[2]),
1932 get_src(ctx, instr->src[3]));
1933 } else {
1934 if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap) {
1935 params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[3]), 0);
1936 }
1937 params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
1938 params[arg_count++] = descriptor;
1939
1940 if (LLVM_VERSION_MAJOR >= 9) {
1941 /* XXX: The new raw/struct atomic intrinsics are buggy with
1942 * LLVM 8, see r358579.
1943 */
1944 params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
1945 params[arg_count++] = ctx->ac.i32_0; /* soffset */
1946 params[arg_count++] = ctx->ac.i32_0; /* slc */
1947
1948 ac_build_type_name_for_intr(return_type, type, sizeof(type));
1949 snprintf(name, sizeof(name),
1950 "llvm.amdgcn.raw.buffer.atomic.%s.%s", op, type);
1951 } else {
1952 params[arg_count++] = ctx->ac.i32_0; /* vindex */
1953 params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
1954 params[arg_count++] = ctx->ac.i1false; /* slc */
1955
1956 assert(return_type == ctx->ac.i32);
1957 snprintf(name, sizeof(name),
1958 "llvm.amdgcn.buffer.atomic.%s", op);
1959 }
1960
1961 result = ac_build_intrinsic(&ctx->ac, name, return_type, params,
1962 arg_count, 0);
1963 }
1964
1965 result = exit_waterfall(ctx, &wctx, result);
1966 if (ctx->ac.postponed_kill)
1967 ac_build_endif(&ctx->ac, 7001);
1968 return result;
1969 }
1970
1971 static LLVMValueRef visit_load_buffer(struct ac_nir_context *ctx,
1972 nir_intrinsic_instr *instr)
1973 {
1974 struct waterfall_context wctx;
1975 LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[0]);
1976
1977 int elem_size_bytes = instr->dest.ssa.bit_size / 8;
1978 int num_components = instr->num_components;
1979 enum gl_access_qualifier access = nir_intrinsic_access(instr);
1980 unsigned cache_policy = get_cache_policy(ctx, access, false, false);
1981
1982 LLVMValueRef offset = get_src(ctx, instr->src[1]);
1983 LLVMValueRef rsrc = ctx->abi->load_ssbo(ctx->abi, rsrc_base, false);
1984 LLVMValueRef vindex = ctx->ac.i32_0;
1985
1986 LLVMTypeRef def_type = get_def_type(ctx, &instr->dest.ssa);
1987 LLVMTypeRef def_elem_type = num_components > 1 ? LLVMGetElementType(def_type) : def_type;
1988
1989 LLVMValueRef results[4];
1990 for (int i = 0; i < num_components;) {
1991 int num_elems = num_components - i;
1992 if (elem_size_bytes < 4 && nir_intrinsic_align(instr) % 4 != 0)
1993 num_elems = 1;
1994 if (num_elems * elem_size_bytes > 16)
1995 num_elems = 16 / elem_size_bytes;
1996 int load_bytes = num_elems * elem_size_bytes;
1997
1998 LLVMValueRef immoffset = LLVMConstInt(ctx->ac.i32, i * elem_size_bytes, false);
1999
2000 LLVMValueRef ret;
2001
2002 if (load_bytes == 1) {
2003 ret = ac_build_tbuffer_load_byte(&ctx->ac,
2004 rsrc,
2005 offset,
2006 ctx->ac.i32_0,
2007 immoffset,
2008 cache_policy);
2009 } else if (load_bytes == 2) {
2010 ret = ac_build_tbuffer_load_short(&ctx->ac,
2011 rsrc,
2012 offset,
2013 ctx->ac.i32_0,
2014 immoffset,
2015 cache_policy);
2016 } else {
2017 int num_channels = util_next_power_of_two(load_bytes) / 4;
2018 bool can_speculate = access & ACCESS_CAN_REORDER;
2019
2020 ret = ac_build_buffer_load(&ctx->ac, rsrc, num_channels,
2021 vindex, offset, immoffset, 0,
2022 cache_policy, can_speculate, false);
2023 }
2024
2025 LLVMTypeRef byte_vec = LLVMVectorType(ctx->ac.i8, ac_get_type_size(LLVMTypeOf(ret)));
2026 ret = LLVMBuildBitCast(ctx->ac.builder, ret, byte_vec, "");
2027 ret = ac_trim_vector(&ctx->ac, ret, load_bytes);
2028
2029 LLVMTypeRef ret_type = LLVMVectorType(def_elem_type, num_elems);
2030 ret = LLVMBuildBitCast(ctx->ac.builder, ret, ret_type, "");
2031
2032 for (unsigned j = 0; j < num_elems; j++) {
2033 results[i + j] = LLVMBuildExtractElement(ctx->ac.builder, ret, LLVMConstInt(ctx->ac.i32, j, false), "");
2034 }
2035 i += num_elems;
2036 }
2037
2038 LLVMValueRef ret = ac_build_gather_values(&ctx->ac, results, num_components);
2039 return exit_waterfall(ctx, &wctx, ret);
2040 }
2041
2042 static LLVMValueRef enter_waterfall_ubo(struct ac_nir_context *ctx,
2043 struct waterfall_context *wctx,
2044 const nir_intrinsic_instr *instr)
2045 {
2046 return enter_waterfall(ctx, wctx, get_src(ctx, instr->src[0]),
2047 nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
2048 }
2049
2050 static LLVMValueRef visit_load_ubo_buffer(struct ac_nir_context *ctx,
2051 nir_intrinsic_instr *instr)
2052 {
2053 struct waterfall_context wctx;
2054 LLVMValueRef rsrc_base = enter_waterfall_ubo(ctx, &wctx, instr);
2055
2056 LLVMValueRef ret;
2057 LLVMValueRef rsrc = rsrc_base;
2058 LLVMValueRef offset = get_src(ctx, instr->src[1]);
2059 int num_components = instr->num_components;
2060
2061 if (ctx->abi->load_ubo)
2062 rsrc = ctx->abi->load_ubo(ctx->abi, rsrc);
2063
2064 if (instr->dest.ssa.bit_size == 64)
2065 num_components *= 2;
2066
2067 if (instr->dest.ssa.bit_size == 16 || instr->dest.ssa.bit_size == 8) {
2068 unsigned load_bytes = instr->dest.ssa.bit_size / 8;
2069 LLVMValueRef results[num_components];
2070 for (unsigned i = 0; i < num_components; ++i) {
2071 LLVMValueRef immoffset = LLVMConstInt(ctx->ac.i32,
2072 load_bytes * i, 0);
2073
2074 if (load_bytes == 1) {
2075 results[i] = ac_build_tbuffer_load_byte(&ctx->ac,
2076 rsrc,
2077 offset,
2078 ctx->ac.i32_0,
2079 immoffset,
2080 0);
2081 } else {
2082 assert(load_bytes == 2);
2083 results[i] = ac_build_tbuffer_load_short(&ctx->ac,
2084 rsrc,
2085 offset,
2086 ctx->ac.i32_0,
2087 immoffset,
2088 0);
2089 }
2090 }
2091 ret = ac_build_gather_values(&ctx->ac, results, num_components);
2092 } else {
2093 ret = ac_build_buffer_load(&ctx->ac, rsrc, num_components, NULL, offset,
2094 NULL, 0, 0, true, true);
2095
2096 ret = ac_trim_vector(&ctx->ac, ret, num_components);
2097 }
2098
2099 ret = LLVMBuildBitCast(ctx->ac.builder, ret,
2100 get_def_type(ctx, &instr->dest.ssa), "");
2101
2102 return exit_waterfall(ctx, &wctx, ret);
2103 }
2104
2105 static void
2106 get_deref_offset(struct ac_nir_context *ctx, nir_deref_instr *instr,
2107 bool vs_in, unsigned *vertex_index_out,
2108 LLVMValueRef *vertex_index_ref,
2109 unsigned *const_out, LLVMValueRef *indir_out)
2110 {
2111 nir_variable *var = nir_deref_instr_get_variable(instr);
2112 nir_deref_path path;
2113 unsigned idx_lvl = 1;
2114
2115 nir_deref_path_init(&path, instr, NULL);
2116
2117 if (vertex_index_out != NULL || vertex_index_ref != NULL) {
2118 if (vertex_index_ref) {
2119 *vertex_index_ref = get_src(ctx, path.path[idx_lvl]->arr.index);
2120 if (vertex_index_out)
2121 *vertex_index_out = 0;
2122 } else {
2123 *vertex_index_out = nir_src_as_uint(path.path[idx_lvl]->arr.index);
2124 }
2125 ++idx_lvl;
2126 }
2127
2128 uint32_t const_offset = 0;
2129 LLVMValueRef offset = NULL;
2130
2131 if (var->data.compact) {
2132 assert(instr->deref_type == nir_deref_type_array);
2133 const_offset = nir_src_as_uint(instr->arr.index);
2134 goto out;
2135 }
2136
2137 for (; path.path[idx_lvl]; ++idx_lvl) {
2138 const struct glsl_type *parent_type = path.path[idx_lvl - 1]->type;
2139 if (path.path[idx_lvl]->deref_type == nir_deref_type_struct) {
2140 unsigned index = path.path[idx_lvl]->strct.index;
2141
2142 for (unsigned i = 0; i < index; i++) {
2143 const struct glsl_type *ft = glsl_get_struct_field(parent_type, i);
2144 const_offset += glsl_count_attribute_slots(ft, vs_in);
2145 }
2146 } else if(path.path[idx_lvl]->deref_type == nir_deref_type_array) {
2147 unsigned size = glsl_count_attribute_slots(path.path[idx_lvl]->type, vs_in);
2148 if (nir_src_is_const(path.path[idx_lvl]->arr.index)) {
2149 const_offset += size *
2150 nir_src_as_uint(path.path[idx_lvl]->arr.index);
2151 } else {
2152 LLVMValueRef array_off = LLVMBuildMul(ctx->ac.builder, LLVMConstInt(ctx->ac.i32, size, 0),
2153 get_src(ctx, path.path[idx_lvl]->arr.index), "");
2154 if (offset)
2155 offset = LLVMBuildAdd(ctx->ac.builder, offset, array_off, "");
2156 else
2157 offset = array_off;
2158 }
2159 } else
2160 unreachable("Uhandled deref type in get_deref_instr_offset");
2161 }
2162
2163 out:
2164 nir_deref_path_finish(&path);
2165
2166 if (const_offset && offset)
2167 offset = LLVMBuildAdd(ctx->ac.builder, offset,
2168 LLVMConstInt(ctx->ac.i32, const_offset, 0),
2169 "");
2170
2171 *const_out = const_offset;
2172 *indir_out = offset;
2173 }
2174
2175 static LLVMValueRef load_tess_varyings(struct ac_nir_context *ctx,
2176 nir_intrinsic_instr *instr,
2177 bool load_inputs)
2178 {
2179 LLVMValueRef result;
2180 LLVMValueRef vertex_index = NULL;
2181 LLVMValueRef indir_index = NULL;
2182 unsigned const_index = 0;
2183
2184 nir_variable *var = nir_deref_instr_get_variable(nir_instr_as_deref(instr->src[0].ssa->parent_instr));
2185
2186 unsigned location = var->data.location;
2187 unsigned driver_location = var->data.driver_location;
2188 const bool is_patch = var->data.patch ||
2189 var->data.location == VARYING_SLOT_TESS_LEVEL_INNER ||
2190 var->data.location == VARYING_SLOT_TESS_LEVEL_OUTER;
2191 const bool is_compact = var->data.compact;
2192
2193 get_deref_offset(ctx, nir_instr_as_deref(instr->src[0].ssa->parent_instr),
2194 false, NULL, is_patch ? NULL : &vertex_index,
2195 &const_index, &indir_index);
2196
2197 LLVMTypeRef dest_type = get_def_type(ctx, &instr->dest.ssa);
2198
2199 LLVMTypeRef src_component_type;
2200 if (LLVMGetTypeKind(dest_type) == LLVMVectorTypeKind)
2201 src_component_type = LLVMGetElementType(dest_type);
2202 else
2203 src_component_type = dest_type;
2204
2205 result = ctx->abi->load_tess_varyings(ctx->abi, src_component_type,
2206 vertex_index, indir_index,
2207 const_index, location, driver_location,
2208 var->data.location_frac,
2209 instr->num_components,
2210 is_patch, is_compact, load_inputs);
2211 if (instr->dest.ssa.bit_size == 16) {
2212 result = ac_to_integer(&ctx->ac, result);
2213 result = LLVMBuildTrunc(ctx->ac.builder, result, dest_type, "");
2214 }
2215 return LLVMBuildBitCast(ctx->ac.builder, result, dest_type, "");
2216 }
2217
2218 static unsigned
2219 type_scalar_size_bytes(const struct glsl_type *type)
2220 {
2221 assert(glsl_type_is_vector_or_scalar(type) ||
2222 glsl_type_is_matrix(type));
2223 return glsl_type_is_boolean(type) ? 4 : glsl_get_bit_size(type) / 8;
2224 }
2225
2226 static LLVMValueRef visit_load_var(struct ac_nir_context *ctx,
2227 nir_intrinsic_instr *instr)
2228 {
2229 nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2230 nir_variable *var = nir_deref_instr_get_variable(deref);
2231
2232 LLVMValueRef values[8];
2233 int idx = 0;
2234 int ve = instr->dest.ssa.num_components;
2235 unsigned comp = 0;
2236 LLVMValueRef indir_index;
2237 LLVMValueRef ret;
2238 unsigned const_index;
2239 unsigned stride = 4;
2240 int mode = deref->mode;
2241
2242 if (var) {
2243 bool vs_in = ctx->stage == MESA_SHADER_VERTEX &&
2244 var->data.mode == nir_var_shader_in;
2245 idx = var->data.driver_location;
2246 comp = var->data.location_frac;
2247 mode = var->data.mode;
2248
2249 get_deref_offset(ctx, deref, vs_in, NULL, NULL,
2250 &const_index, &indir_index);
2251
2252 if (var->data.compact) {
2253 stride = 1;
2254 const_index += comp;
2255 comp = 0;
2256 }
2257 }
2258
2259 if (instr->dest.ssa.bit_size == 64 &&
2260 (deref->mode == nir_var_shader_in ||
2261 deref->mode == nir_var_shader_out ||
2262 deref->mode == nir_var_function_temp))
2263 ve *= 2;
2264
2265 switch (mode) {
2266 case nir_var_shader_in:
2267 if (ctx->stage == MESA_SHADER_TESS_CTRL ||
2268 ctx->stage == MESA_SHADER_TESS_EVAL) {
2269 return load_tess_varyings(ctx, instr, true);
2270 }
2271
2272 if (ctx->stage == MESA_SHADER_GEOMETRY) {
2273 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
2274 LLVMValueRef indir_index;
2275 unsigned const_index, vertex_index;
2276 get_deref_offset(ctx, deref, false, &vertex_index, NULL,
2277 &const_index, &indir_index);
2278 assert(indir_index == NULL);
2279
2280 return ctx->abi->load_inputs(ctx->abi, var->data.location,
2281 var->data.driver_location,
2282 var->data.location_frac,
2283 instr->num_components, vertex_index, const_index, type);
2284 }
2285
2286 for (unsigned chan = comp; chan < ve + comp; chan++) {
2287 if (indir_index) {
2288 unsigned count = glsl_count_attribute_slots(
2289 var->type,
2290 ctx->stage == MESA_SHADER_VERTEX);
2291 count -= chan / 4;
2292 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2293 &ctx->ac, ctx->abi->inputs + idx + chan, count,
2294 stride, false, true);
2295
2296 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2297 tmp_vec,
2298 indir_index, "");
2299 } else
2300 values[chan] = ctx->abi->inputs[idx + chan + const_index * stride];
2301 }
2302 break;
2303 case nir_var_function_temp:
2304 for (unsigned chan = 0; chan < ve; chan++) {
2305 if (indir_index) {
2306 unsigned count = glsl_count_attribute_slots(
2307 var->type, false);
2308 count -= chan / 4;
2309 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2310 &ctx->ac, ctx->locals + idx + chan, count,
2311 stride, true, true);
2312
2313 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2314 tmp_vec,
2315 indir_index, "");
2316 } else {
2317 values[chan] = LLVMBuildLoad(ctx->ac.builder, ctx->locals[idx + chan + const_index * stride], "");
2318 }
2319 }
2320 break;
2321 case nir_var_shader_out:
2322 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
2323 return load_tess_varyings(ctx, instr, false);
2324 }
2325
2326 if (ctx->stage == MESA_SHADER_FRAGMENT &&
2327 var->data.fb_fetch_output &&
2328 ctx->abi->emit_fbfetch)
2329 return ctx->abi->emit_fbfetch(ctx->abi);
2330
2331 for (unsigned chan = comp; chan < ve + comp; chan++) {
2332 if (indir_index) {
2333 unsigned count = glsl_count_attribute_slots(
2334 var->type, false);
2335 count -= chan / 4;
2336 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2337 &ctx->ac, ctx->abi->outputs + idx + chan, count,
2338 stride, true, true);
2339
2340 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2341 tmp_vec,
2342 indir_index, "");
2343 } else {
2344 values[chan] = LLVMBuildLoad(ctx->ac.builder,
2345 ctx->abi->outputs[idx + chan + const_index * stride],
2346 "");
2347 }
2348 }
2349 break;
2350 case nir_var_mem_global: {
2351 LLVMValueRef address = get_src(ctx, instr->src[0]);
2352 LLVMTypeRef result_type = get_def_type(ctx, &instr->dest.ssa);
2353 unsigned explicit_stride = glsl_get_explicit_stride(deref->type);
2354 unsigned natural_stride = type_scalar_size_bytes(deref->type);
2355 unsigned stride = explicit_stride ? explicit_stride : natural_stride;
2356 int elem_size_bytes = ac_get_elem_bits(&ctx->ac, result_type) / 8;
2357 bool split_loads = ctx->ac.chip_class == GFX6 && elem_size_bytes < 4;
2358
2359 if (stride != natural_stride || split_loads) {
2360 if (LLVMGetTypeKind(result_type) == LLVMVectorTypeKind)
2361 result_type = LLVMGetElementType(result_type);
2362
2363 LLVMTypeRef ptr_type = LLVMPointerType(result_type,
2364 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2365 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2366
2367 for (unsigned i = 0; i < instr->dest.ssa.num_components; ++i) {
2368 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, i * stride / natural_stride, 0);
2369 values[i] = LLVMBuildLoad(ctx->ac.builder,
2370 ac_build_gep_ptr(&ctx->ac, address, offset), "");
2371 }
2372 return ac_build_gather_values(&ctx->ac, values, instr->dest.ssa.num_components);
2373 } else {
2374 LLVMTypeRef ptr_type = LLVMPointerType(result_type,
2375 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2376 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2377 LLVMValueRef val = LLVMBuildLoad(ctx->ac.builder, address, "");
2378 return val;
2379 }
2380 }
2381 default:
2382 unreachable("unhandle variable mode");
2383 }
2384 ret = ac_build_varying_gather_values(&ctx->ac, values, ve, comp);
2385 return LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->dest.ssa), "");
2386 }
2387
2388 static void
2389 visit_store_var(struct ac_nir_context *ctx,
2390 nir_intrinsic_instr *instr)
2391 {
2392 if (ctx->ac.postponed_kill) {
2393 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
2394 ctx->ac.postponed_kill, "");
2395 ac_build_ifcc(&ctx->ac, cond, 7002);
2396 }
2397
2398 nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2399 nir_variable *var = nir_deref_instr_get_variable(deref);
2400
2401 LLVMValueRef temp_ptr, value;
2402 int idx = 0;
2403 unsigned comp = 0;
2404 LLVMValueRef src = ac_to_float(&ctx->ac, get_src(ctx, instr->src[1]));
2405 int writemask = instr->const_index[0];
2406 LLVMValueRef indir_index;
2407 unsigned const_index;
2408
2409 if (var) {
2410 get_deref_offset(ctx, deref, false,
2411 NULL, NULL, &const_index, &indir_index);
2412 idx = var->data.driver_location;
2413 comp = var->data.location_frac;
2414
2415 if (var->data.compact) {
2416 const_index += comp;
2417 comp = 0;
2418 }
2419 }
2420
2421 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src)) == 64 &&
2422 (deref->mode == nir_var_shader_out ||
2423 deref->mode == nir_var_function_temp)) {
2424
2425 src = LLVMBuildBitCast(ctx->ac.builder, src,
2426 LLVMVectorType(ctx->ac.f32, ac_get_llvm_num_components(src) * 2),
2427 "");
2428
2429 writemask = widen_mask(writemask, 2);
2430 }
2431
2432 writemask = writemask << comp;
2433
2434 switch (deref->mode) {
2435 case nir_var_shader_out:
2436
2437 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
2438 LLVMValueRef vertex_index = NULL;
2439 LLVMValueRef indir_index = NULL;
2440 unsigned const_index = 0;
2441 const bool is_patch = var->data.patch ||
2442 var->data.location == VARYING_SLOT_TESS_LEVEL_INNER ||
2443 var->data.location == VARYING_SLOT_TESS_LEVEL_OUTER;
2444
2445 get_deref_offset(ctx, deref, false, NULL,
2446 is_patch ? NULL : &vertex_index,
2447 &const_index, &indir_index);
2448
2449 ctx->abi->store_tcs_outputs(ctx->abi, var,
2450 vertex_index, indir_index,
2451 const_index, src, writemask);
2452 break;
2453 }
2454
2455 for (unsigned chan = 0; chan < 8; chan++) {
2456 int stride = 4;
2457 if (!(writemask & (1 << chan)))
2458 continue;
2459
2460 value = ac_llvm_extract_elem(&ctx->ac, src, chan - comp);
2461
2462 if (var->data.compact)
2463 stride = 1;
2464 if (indir_index) {
2465 unsigned count = glsl_count_attribute_slots(
2466 var->type, false);
2467 count -= chan / 4;
2468 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2469 &ctx->ac, ctx->abi->outputs + idx + chan, count,
2470 stride, true, true);
2471
2472 tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
2473 value, indir_index, "");
2474 build_store_values_extended(&ctx->ac, ctx->abi->outputs + idx + chan,
2475 count, stride, tmp_vec);
2476
2477 } else {
2478 temp_ptr = ctx->abi->outputs[idx + chan + const_index * stride];
2479
2480 LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
2481 }
2482 }
2483 break;
2484 case nir_var_function_temp:
2485 for (unsigned chan = 0; chan < 8; chan++) {
2486 if (!(writemask & (1 << chan)))
2487 continue;
2488
2489 value = ac_llvm_extract_elem(&ctx->ac, src, chan);
2490 if (indir_index) {
2491 unsigned count = glsl_count_attribute_slots(
2492 var->type, false);
2493 count -= chan / 4;
2494 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2495 &ctx->ac, ctx->locals + idx + chan, count,
2496 4, true, true);
2497
2498 tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
2499 value, indir_index, "");
2500 build_store_values_extended(&ctx->ac, ctx->locals + idx + chan,
2501 count, 4, tmp_vec);
2502 } else {
2503 temp_ptr = ctx->locals[idx + chan + const_index * 4];
2504
2505 LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
2506 }
2507 }
2508 break;
2509
2510 case nir_var_mem_global: {
2511 int writemask = instr->const_index[0];
2512 LLVMValueRef address = get_src(ctx, instr->src[0]);
2513 LLVMValueRef val = get_src(ctx, instr->src[1]);
2514
2515 unsigned explicit_stride = glsl_get_explicit_stride(deref->type);
2516 unsigned natural_stride = type_scalar_size_bytes(deref->type);
2517 unsigned stride = explicit_stride ? explicit_stride : natural_stride;
2518 int elem_size_bytes = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(val)) / 8;
2519 bool split_stores = ctx->ac.chip_class == GFX6 && elem_size_bytes < 4;
2520
2521 LLVMTypeRef ptr_type = LLVMPointerType(LLVMTypeOf(val),
2522 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2523 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2524
2525 if (writemask == (1u << ac_get_llvm_num_components(val)) - 1 &&
2526 stride == natural_stride && !split_stores) {
2527 LLVMTypeRef ptr_type = LLVMPointerType(LLVMTypeOf(val),
2528 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2529 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2530
2531 val = LLVMBuildBitCast(ctx->ac.builder, val,
2532 LLVMGetElementType(LLVMTypeOf(address)), "");
2533 LLVMBuildStore(ctx->ac.builder, val, address);
2534 } else {
2535 LLVMTypeRef val_type = LLVMTypeOf(val);
2536 if (LLVMGetTypeKind(LLVMTypeOf(val)) == LLVMVectorTypeKind)
2537 val_type = LLVMGetElementType(val_type);
2538
2539 LLVMTypeRef ptr_type = LLVMPointerType(val_type,
2540 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2541 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2542 for (unsigned chan = 0; chan < 4; chan++) {
2543 if (!(writemask & (1 << chan)))
2544 continue;
2545
2546 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, chan * stride / natural_stride, 0);
2547
2548 LLVMValueRef ptr = ac_build_gep_ptr(&ctx->ac, address, offset);
2549 LLVMValueRef src = ac_llvm_extract_elem(&ctx->ac, val,
2550 chan);
2551 src = LLVMBuildBitCast(ctx->ac.builder, src,
2552 LLVMGetElementType(LLVMTypeOf(ptr)), "");
2553 LLVMBuildStore(ctx->ac.builder, src, ptr);
2554 }
2555 }
2556 break;
2557 }
2558 default:
2559 abort();
2560 break;
2561 }
2562
2563 if (ctx->ac.postponed_kill)
2564 ac_build_endif(&ctx->ac, 7002);
2565 }
2566
2567 static int image_type_to_components_count(enum glsl_sampler_dim dim, bool array)
2568 {
2569 switch (dim) {
2570 case GLSL_SAMPLER_DIM_BUF:
2571 return 1;
2572 case GLSL_SAMPLER_DIM_1D:
2573 return array ? 2 : 1;
2574 case GLSL_SAMPLER_DIM_2D:
2575 return array ? 3 : 2;
2576 case GLSL_SAMPLER_DIM_MS:
2577 return array ? 4 : 3;
2578 case GLSL_SAMPLER_DIM_3D:
2579 case GLSL_SAMPLER_DIM_CUBE:
2580 return 3;
2581 case GLSL_SAMPLER_DIM_RECT:
2582 case GLSL_SAMPLER_DIM_SUBPASS:
2583 return 2;
2584 case GLSL_SAMPLER_DIM_SUBPASS_MS:
2585 return 3;
2586 default:
2587 break;
2588 }
2589 return 0;
2590 }
2591
2592 static LLVMValueRef adjust_sample_index_using_fmask(struct ac_llvm_context *ctx,
2593 LLVMValueRef coord_x, LLVMValueRef coord_y,
2594 LLVMValueRef coord_z,
2595 LLVMValueRef sample_index,
2596 LLVMValueRef fmask_desc_ptr)
2597 {
2598 unsigned sample_chan = coord_z ? 3 : 2;
2599 LLVMValueRef addr[4] = {coord_x, coord_y, coord_z};
2600 addr[sample_chan] = sample_index;
2601
2602 ac_apply_fmask_to_sample(ctx, fmask_desc_ptr, addr, coord_z != NULL);
2603 return addr[sample_chan];
2604 }
2605
2606 static nir_deref_instr *get_image_deref(const nir_intrinsic_instr *instr)
2607 {
2608 assert(instr->src[0].is_ssa);
2609 return nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2610 }
2611
2612 static LLVMValueRef get_image_descriptor(struct ac_nir_context *ctx,
2613 const nir_intrinsic_instr *instr,
2614 LLVMValueRef dynamic_index,
2615 enum ac_descriptor_type desc_type,
2616 bool write)
2617 {
2618 nir_deref_instr *deref_instr =
2619 instr->src[0].ssa->parent_instr->type == nir_instr_type_deref ?
2620 nir_instr_as_deref(instr->src[0].ssa->parent_instr) : NULL;
2621
2622 return get_sampler_desc(ctx, deref_instr, desc_type, &instr->instr, dynamic_index, true, write);
2623 }
2624
2625 static void get_image_coords(struct ac_nir_context *ctx,
2626 const nir_intrinsic_instr *instr,
2627 LLVMValueRef dynamic_desc_index,
2628 struct ac_image_args *args,
2629 enum glsl_sampler_dim dim,
2630 bool is_array)
2631 {
2632 LLVMValueRef src0 = get_src(ctx, instr->src[1]);
2633 LLVMValueRef masks[] = {
2634 LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
2635 LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
2636 };
2637 LLVMValueRef sample_index = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
2638
2639 int count;
2640 ASSERTED bool add_frag_pos = (dim == GLSL_SAMPLER_DIM_SUBPASS ||
2641 dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2642 bool is_ms = (dim == GLSL_SAMPLER_DIM_MS ||
2643 dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2644 bool gfx9_1d = ctx->ac.chip_class == GFX9 && dim == GLSL_SAMPLER_DIM_1D;
2645 assert(!add_frag_pos && "Input attachments should be lowered by this point.");
2646 count = image_type_to_components_count(dim, is_array);
2647
2648 if (is_ms && (instr->intrinsic == nir_intrinsic_image_deref_load ||
2649 instr->intrinsic == nir_intrinsic_bindless_image_load)) {
2650 LLVMValueRef fmask_load_address[3];
2651
2652 fmask_load_address[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
2653 fmask_load_address[1] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[1], "");
2654 if (is_array)
2655 fmask_load_address[2] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[2], "");
2656 else
2657 fmask_load_address[2] = NULL;
2658
2659 sample_index = adjust_sample_index_using_fmask(&ctx->ac,
2660 fmask_load_address[0],
2661 fmask_load_address[1],
2662 fmask_load_address[2],
2663 sample_index,
2664 get_sampler_desc(ctx, nir_instr_as_deref(instr->src[0].ssa->parent_instr),
2665 AC_DESC_FMASK, &instr->instr, dynamic_desc_index, true, false));
2666 }
2667 if (count == 1 && !gfx9_1d) {
2668 if (instr->src[1].ssa->num_components)
2669 args->coords[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
2670 else
2671 args->coords[0] = src0;
2672 } else {
2673 int chan;
2674 if (is_ms)
2675 count--;
2676 for (chan = 0; chan < count; ++chan) {
2677 args->coords[chan] = ac_llvm_extract_elem(&ctx->ac, src0, chan);
2678 }
2679
2680 if (gfx9_1d) {
2681 if (is_array) {
2682 args->coords[2] = args->coords[1];
2683 args->coords[1] = ctx->ac.i32_0;
2684 } else
2685 args->coords[1] = ctx->ac.i32_0;
2686 count++;
2687 }
2688 if (ctx->ac.chip_class == GFX9 &&
2689 dim == GLSL_SAMPLER_DIM_2D &&
2690 !is_array) {
2691 /* The hw can't bind a slice of a 3D image as a 2D
2692 * image, because it ignores BASE_ARRAY if the target
2693 * is 3D. The workaround is to read BASE_ARRAY and set
2694 * it as the 3rd address operand for all 2D images.
2695 */
2696 LLVMValueRef first_layer, const5, mask;
2697
2698 const5 = LLVMConstInt(ctx->ac.i32, 5, 0);
2699 mask = LLVMConstInt(ctx->ac.i32, S_008F24_BASE_ARRAY(~0), 0);
2700 first_layer = LLVMBuildExtractElement(ctx->ac.builder, args->resource, const5, "");
2701 first_layer = LLVMBuildAnd(ctx->ac.builder, first_layer, mask, "");
2702
2703 args->coords[count] = first_layer;
2704 count++;
2705 }
2706
2707
2708 if (is_ms) {
2709 args->coords[count] = sample_index;
2710 count++;
2711 }
2712 }
2713 }
2714
2715 static LLVMValueRef get_image_buffer_descriptor(struct ac_nir_context *ctx,
2716 const nir_intrinsic_instr *instr,
2717 LLVMValueRef dynamic_index,
2718 bool write, bool atomic)
2719 {
2720 LLVMValueRef rsrc = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_BUFFER, write);
2721 if (ctx->ac.chip_class == GFX9 && LLVM_VERSION_MAJOR < 9 && atomic) {
2722 LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 2, 0), "");
2723 LLVMValueRef stride = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 1, 0), "");
2724 stride = LLVMBuildLShr(ctx->ac.builder, stride, LLVMConstInt(ctx->ac.i32, 16, 0), "");
2725
2726 LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->ac.builder,
2727 LLVMBuildICmp(ctx->ac.builder, LLVMIntUGT, elem_count, stride, ""),
2728 elem_count, stride, "");
2729
2730 rsrc = LLVMBuildInsertElement(ctx->ac.builder, rsrc, new_elem_count,
2731 LLVMConstInt(ctx->ac.i32, 2, 0), "");
2732 }
2733 return rsrc;
2734 }
2735
2736 static LLVMValueRef enter_waterfall_image(struct ac_nir_context *ctx,
2737 struct waterfall_context *wctx,
2738 const nir_intrinsic_instr *instr)
2739 {
2740 nir_deref_instr *deref_instr = NULL;
2741
2742 if (instr->src[0].ssa->parent_instr->type == nir_instr_type_deref)
2743 deref_instr = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2744
2745 LLVMValueRef value = get_sampler_desc_index(ctx, deref_instr, &instr->instr, true);
2746 return enter_waterfall(ctx, wctx, value, nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
2747 }
2748
2749 static LLVMValueRef visit_image_load(struct ac_nir_context *ctx,
2750 const nir_intrinsic_instr *instr,
2751 bool bindless)
2752 {
2753 LLVMValueRef res;
2754
2755 enum glsl_sampler_dim dim;
2756 enum gl_access_qualifier access;
2757 bool is_array;
2758 if (bindless) {
2759 dim = nir_intrinsic_image_dim(instr);
2760 access = nir_intrinsic_access(instr);
2761 is_array = nir_intrinsic_image_array(instr);
2762 } else {
2763 const nir_deref_instr *image_deref = get_image_deref(instr);
2764 const struct glsl_type *type = image_deref->type;
2765 const nir_variable *var = nir_deref_instr_get_variable(image_deref);
2766 dim = glsl_get_sampler_dim(type);
2767 access = var->data.access;
2768 is_array = glsl_sampler_type_is_array(type);
2769 }
2770
2771 struct waterfall_context wctx;
2772 LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
2773
2774 struct ac_image_args args = {};
2775
2776 args.cache_policy = get_cache_policy(ctx, access, false, false);
2777
2778 if (dim == GLSL_SAMPLER_DIM_BUF) {
2779 unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
2780 unsigned num_channels = util_last_bit(mask);
2781 LLVMValueRef rsrc, vindex;
2782
2783 rsrc = get_image_buffer_descriptor(ctx, instr, dynamic_index, false, false);
2784 vindex = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
2785 ctx->ac.i32_0, "");
2786
2787 assert(instr->dest.is_ssa);
2788 bool can_speculate = access & ACCESS_CAN_REORDER;
2789 res = ac_build_buffer_load_format(&ctx->ac, rsrc, vindex,
2790 ctx->ac.i32_0, num_channels,
2791 args.cache_policy,
2792 can_speculate,
2793 instr->dest.ssa.bit_size == 16);
2794 res = ac_build_expand_to_vec4(&ctx->ac, res, num_channels);
2795
2796 res = ac_trim_vector(&ctx->ac, res, instr->dest.ssa.num_components);
2797 res = ac_to_integer(&ctx->ac, res);
2798 } else {
2799 bool level_zero = nir_src_is_const(instr->src[3]) && nir_src_as_uint(instr->src[3]) == 0;
2800
2801 args.opcode = level_zero ? ac_image_load : ac_image_load_mip;
2802 args.resource = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_IMAGE, false);
2803 get_image_coords(ctx, instr, dynamic_index, &args, dim, is_array);
2804 args.dim = ac_get_image_dim(ctx->ac.chip_class, dim, is_array);
2805 if (!level_zero)
2806 args.lod = get_src(ctx, instr->src[3]);
2807 args.dmask = 15;
2808 args.attributes = AC_FUNC_ATTR_READONLY;
2809
2810 res = ac_build_image_opcode(&ctx->ac, &args);
2811 }
2812 return exit_waterfall(ctx, &wctx, res);
2813 }
2814
2815 static void visit_image_store(struct ac_nir_context *ctx,
2816 const nir_intrinsic_instr *instr,
2817 bool bindless)
2818 {
2819 if (ctx->ac.postponed_kill) {
2820 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
2821 ctx->ac.postponed_kill, "");
2822 ac_build_ifcc(&ctx->ac, cond, 7003);
2823 }
2824
2825 enum glsl_sampler_dim dim;
2826 enum gl_access_qualifier access;
2827 bool is_array;
2828
2829 if (bindless) {
2830 dim = nir_intrinsic_image_dim(instr);
2831 access = nir_intrinsic_access(instr);
2832 is_array = nir_intrinsic_image_array(instr);
2833 } else {
2834 const nir_deref_instr *image_deref = get_image_deref(instr);
2835 const struct glsl_type *type = image_deref->type;
2836 const nir_variable *var = nir_deref_instr_get_variable(image_deref);
2837 dim = glsl_get_sampler_dim(type);
2838 access = var->data.access;
2839 is_array = glsl_sampler_type_is_array(type);
2840 }
2841
2842 struct waterfall_context wctx;
2843 LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
2844
2845 bool writeonly_memory = access & ACCESS_NON_READABLE;
2846 struct ac_image_args args = {};
2847
2848 args.cache_policy = get_cache_policy(ctx, access, true, writeonly_memory);
2849
2850 if (dim == GLSL_SAMPLER_DIM_BUF) {
2851 LLVMValueRef rsrc = get_image_buffer_descriptor(ctx, instr, dynamic_index, true, false);
2852 LLVMValueRef src = ac_to_float(&ctx->ac, get_src(ctx, instr->src[3]));
2853 unsigned src_channels = ac_get_llvm_num_components(src);
2854 LLVMValueRef vindex;
2855
2856 if (src_channels == 3)
2857 src = ac_build_expand_to_vec4(&ctx->ac, src, 3);
2858
2859 vindex = LLVMBuildExtractElement(ctx->ac.builder,
2860 get_src(ctx, instr->src[1]),
2861 ctx->ac.i32_0, "");
2862
2863 ac_build_buffer_store_format(&ctx->ac, rsrc, src, vindex,
2864 ctx->ac.i32_0, args.cache_policy);
2865 } else {
2866 bool level_zero = nir_src_is_const(instr->src[4]) && nir_src_as_uint(instr->src[4]) == 0;
2867
2868 args.opcode = level_zero ? ac_image_store : ac_image_store_mip;
2869 args.data[0] = ac_to_float(&ctx->ac, get_src(ctx, instr->src[3]));
2870 args.resource = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_IMAGE, true);
2871 get_image_coords(ctx, instr, dynamic_index, &args, dim, is_array);
2872 args.dim = ac_get_image_dim(ctx->ac.chip_class, dim, is_array);
2873 if (!level_zero)
2874 args.lod = get_src(ctx, instr->src[4]);
2875 args.dmask = 15;
2876
2877 ac_build_image_opcode(&ctx->ac, &args);
2878 }
2879
2880 exit_waterfall(ctx, &wctx, NULL);
2881 if (ctx->ac.postponed_kill)
2882 ac_build_endif(&ctx->ac, 7003);
2883 }
2884
2885 static LLVMValueRef visit_image_atomic(struct ac_nir_context *ctx,
2886 const nir_intrinsic_instr *instr,
2887 bool bindless)
2888 {
2889 if (ctx->ac.postponed_kill) {
2890 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
2891 ctx->ac.postponed_kill, "");
2892 ac_build_ifcc(&ctx->ac, cond, 7004);
2893 }
2894
2895 LLVMValueRef params[7];
2896 int param_count = 0;
2897
2898 bool cmpswap = instr->intrinsic == nir_intrinsic_image_deref_atomic_comp_swap ||
2899 instr->intrinsic == nir_intrinsic_bindless_image_atomic_comp_swap;
2900 const char *atomic_name;
2901 char intrinsic_name[64];
2902 enum ac_atomic_op atomic_subop;
2903 ASSERTED int length;
2904
2905 enum glsl_sampler_dim dim;
2906 bool is_array;
2907 if (bindless) {
2908 if (instr->intrinsic == nir_intrinsic_bindless_image_atomic_imin ||
2909 instr->intrinsic == nir_intrinsic_bindless_image_atomic_umin ||
2910 instr->intrinsic == nir_intrinsic_bindless_image_atomic_imax ||
2911 instr->intrinsic == nir_intrinsic_bindless_image_atomic_umax) {
2912 ASSERTED const GLenum format = nir_intrinsic_format(instr);
2913 assert(format == GL_R32UI || format == GL_R32I);
2914 }
2915 dim = nir_intrinsic_image_dim(instr);
2916 is_array = nir_intrinsic_image_array(instr);
2917 } else {
2918 const struct glsl_type *type = get_image_deref(instr)->type;
2919 dim = glsl_get_sampler_dim(type);
2920 is_array = glsl_sampler_type_is_array(type);
2921 }
2922
2923 struct waterfall_context wctx;
2924 LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
2925
2926 switch (instr->intrinsic) {
2927 case nir_intrinsic_bindless_image_atomic_add:
2928 case nir_intrinsic_image_deref_atomic_add:
2929 atomic_name = "add";
2930 atomic_subop = ac_atomic_add;
2931 break;
2932 case nir_intrinsic_bindless_image_atomic_imin:
2933 case nir_intrinsic_image_deref_atomic_imin:
2934 atomic_name = "smin";
2935 atomic_subop = ac_atomic_smin;
2936 break;
2937 case nir_intrinsic_bindless_image_atomic_umin:
2938 case nir_intrinsic_image_deref_atomic_umin:
2939 atomic_name = "umin";
2940 atomic_subop = ac_atomic_umin;
2941 break;
2942 case nir_intrinsic_bindless_image_atomic_imax:
2943 case nir_intrinsic_image_deref_atomic_imax:
2944 atomic_name = "smax";
2945 atomic_subop = ac_atomic_smax;
2946 break;
2947 case nir_intrinsic_bindless_image_atomic_umax:
2948 case nir_intrinsic_image_deref_atomic_umax:
2949 atomic_name = "umax";
2950 atomic_subop = ac_atomic_umax;
2951 break;
2952 case nir_intrinsic_bindless_image_atomic_and:
2953 case nir_intrinsic_image_deref_atomic_and:
2954 atomic_name = "and";
2955 atomic_subop = ac_atomic_and;
2956 break;
2957 case nir_intrinsic_bindless_image_atomic_or:
2958 case nir_intrinsic_image_deref_atomic_or:
2959 atomic_name = "or";
2960 atomic_subop = ac_atomic_or;
2961 break;
2962 case nir_intrinsic_bindless_image_atomic_xor:
2963 case nir_intrinsic_image_deref_atomic_xor:
2964 atomic_name = "xor";
2965 atomic_subop = ac_atomic_xor;
2966 break;
2967 case nir_intrinsic_bindless_image_atomic_exchange:
2968 case nir_intrinsic_image_deref_atomic_exchange:
2969 atomic_name = "swap";
2970 atomic_subop = ac_atomic_swap;
2971 break;
2972 case nir_intrinsic_bindless_image_atomic_comp_swap:
2973 case nir_intrinsic_image_deref_atomic_comp_swap:
2974 atomic_name = "cmpswap";
2975 atomic_subop = 0; /* not used */
2976 break;
2977 case nir_intrinsic_bindless_image_atomic_inc_wrap:
2978 case nir_intrinsic_image_deref_atomic_inc_wrap: {
2979 atomic_name = "inc";
2980 atomic_subop = ac_atomic_inc_wrap;
2981 /* ATOMIC_INC instruction does:
2982 * value = (value + 1) % (data + 1)
2983 * but we want:
2984 * value = (value + 1) % data
2985 * So replace 'data' by 'data - 1'.
2986 */
2987 ctx->ssa_defs[instr->src[3].ssa->index] =
2988 LLVMBuildSub(ctx->ac.builder,
2989 ctx->ssa_defs[instr->src[3].ssa->index],
2990 ctx->ac.i32_1, "");
2991 break;
2992 }
2993 case nir_intrinsic_bindless_image_atomic_dec_wrap:
2994 case nir_intrinsic_image_deref_atomic_dec_wrap:
2995 atomic_name = "dec";
2996 atomic_subop = ac_atomic_dec_wrap;
2997 break;
2998 default:
2999 abort();
3000 }
3001
3002 if (cmpswap)
3003 params[param_count++] = get_src(ctx, instr->src[4]);
3004 params[param_count++] = get_src(ctx, instr->src[3]);
3005
3006 LLVMValueRef result;
3007 if (dim == GLSL_SAMPLER_DIM_BUF) {
3008 params[param_count++] = get_image_buffer_descriptor(ctx, instr, dynamic_index, true, true);
3009 params[param_count++] = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
3010 ctx->ac.i32_0, ""); /* vindex */
3011 params[param_count++] = ctx->ac.i32_0; /* voffset */
3012 if (LLVM_VERSION_MAJOR >= 9) {
3013 /* XXX: The new raw/struct atomic intrinsics are buggy
3014 * with LLVM 8, see r358579.
3015 */
3016 params[param_count++] = ctx->ac.i32_0; /* soffset */
3017 params[param_count++] = ctx->ac.i32_0; /* slc */
3018
3019 length = snprintf(intrinsic_name, sizeof(intrinsic_name),
3020 "llvm.amdgcn.struct.buffer.atomic.%s.i32", atomic_name);
3021 } else {
3022 params[param_count++] = ctx->ac.i1false; /* slc */
3023
3024 length = snprintf(intrinsic_name, sizeof(intrinsic_name),
3025 "llvm.amdgcn.buffer.atomic.%s", atomic_name);
3026 }
3027
3028 assert(length < sizeof(intrinsic_name));
3029 result = ac_build_intrinsic(&ctx->ac, intrinsic_name, ctx->ac.i32,
3030 params, param_count, 0);
3031 } else {
3032 struct ac_image_args args = {};
3033 args.opcode = cmpswap ? ac_image_atomic_cmpswap : ac_image_atomic;
3034 args.atomic = atomic_subop;
3035 args.data[0] = params[0];
3036 if (cmpswap)
3037 args.data[1] = params[1];
3038 args.resource = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_IMAGE, true);
3039 get_image_coords(ctx, instr, dynamic_index, &args, dim, is_array);
3040 args.dim = ac_get_image_dim(ctx->ac.chip_class, dim, is_array);
3041
3042 result = ac_build_image_opcode(&ctx->ac, &args);
3043 }
3044
3045 result = exit_waterfall(ctx, &wctx, result);
3046 if (ctx->ac.postponed_kill)
3047 ac_build_endif(&ctx->ac, 7004);
3048 return result;
3049 }
3050
3051 static LLVMValueRef visit_image_samples(struct ac_nir_context *ctx,
3052 nir_intrinsic_instr *instr)
3053 {
3054 struct waterfall_context wctx;
3055 LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
3056 LLVMValueRef rsrc = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_IMAGE, false);
3057
3058 LLVMValueRef ret = ac_build_image_get_sample_count(&ctx->ac, rsrc);
3059
3060 return exit_waterfall(ctx, &wctx, ret);
3061 }
3062
3063 static LLVMValueRef visit_image_size(struct ac_nir_context *ctx,
3064 const nir_intrinsic_instr *instr,
3065 bool bindless)
3066 {
3067 LLVMValueRef res;
3068
3069 enum glsl_sampler_dim dim;
3070 bool is_array;
3071 if (bindless) {
3072 dim = nir_intrinsic_image_dim(instr);
3073 is_array = nir_intrinsic_image_array(instr);
3074 } else {
3075 const struct glsl_type *type = get_image_deref(instr)->type;
3076 dim = glsl_get_sampler_dim(type);
3077 is_array = glsl_sampler_type_is_array(type);
3078 }
3079
3080 struct waterfall_context wctx;
3081 LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
3082
3083 if (dim == GLSL_SAMPLER_DIM_BUF) {
3084 res = get_buffer_size(ctx, get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_BUFFER, false), true);
3085 } else {
3086
3087 struct ac_image_args args = { 0 };
3088
3089 args.dim = ac_get_image_dim(ctx->ac.chip_class, dim, is_array);
3090 args.dmask = 0xf;
3091 args.resource = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_IMAGE, false);
3092 args.opcode = ac_image_get_resinfo;
3093 args.lod = ctx->ac.i32_0;
3094 args.attributes = AC_FUNC_ATTR_READNONE;
3095
3096 res = ac_build_image_opcode(&ctx->ac, &args);
3097
3098 LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
3099
3100 if (dim == GLSL_SAMPLER_DIM_CUBE && is_array) {
3101 LLVMValueRef six = LLVMConstInt(ctx->ac.i32, 6, false);
3102 LLVMValueRef z = LLVMBuildExtractElement(ctx->ac.builder, res, two, "");
3103 z = LLVMBuildSDiv(ctx->ac.builder, z, six, "");
3104 res = LLVMBuildInsertElement(ctx->ac.builder, res, z, two, "");
3105 }
3106
3107 if (ctx->ac.chip_class == GFX9 && dim == GLSL_SAMPLER_DIM_1D && is_array) {
3108 LLVMValueRef layers = LLVMBuildExtractElement(ctx->ac.builder, res, two, "");
3109 res = LLVMBuildInsertElement(ctx->ac.builder, res, layers,
3110 ctx->ac.i32_1, "");
3111 }
3112 }
3113 return exit_waterfall(ctx, &wctx, res);
3114 }
3115
3116 static void emit_membar(struct ac_llvm_context *ac,
3117 const nir_intrinsic_instr *instr)
3118 {
3119 unsigned wait_flags = 0;
3120
3121 switch (instr->intrinsic) {
3122 case nir_intrinsic_memory_barrier:
3123 case nir_intrinsic_group_memory_barrier:
3124 wait_flags = AC_WAIT_LGKM | AC_WAIT_VLOAD | AC_WAIT_VSTORE;
3125 break;
3126 case nir_intrinsic_memory_barrier_buffer:
3127 case nir_intrinsic_memory_barrier_image:
3128 wait_flags = AC_WAIT_VLOAD | AC_WAIT_VSTORE;
3129 break;
3130 case nir_intrinsic_memory_barrier_shared:
3131 wait_flags = AC_WAIT_LGKM;
3132 break;
3133 default:
3134 break;
3135 }
3136
3137 ac_build_waitcnt(ac, wait_flags);
3138 }
3139
3140 void ac_emit_barrier(struct ac_llvm_context *ac, gl_shader_stage stage)
3141 {
3142 /* GFX6 only (thanks to a hw bug workaround):
3143 * The real barrier instruction isn’t needed, because an entire patch
3144 * always fits into a single wave.
3145 */
3146 if (ac->chip_class == GFX6 && stage == MESA_SHADER_TESS_CTRL) {
3147 ac_build_waitcnt(ac, AC_WAIT_LGKM | AC_WAIT_VLOAD | AC_WAIT_VSTORE);
3148 return;
3149 }
3150 ac_build_s_barrier(ac);
3151 }
3152
3153 static void emit_discard(struct ac_nir_context *ctx,
3154 const nir_intrinsic_instr *instr)
3155 {
3156 LLVMValueRef cond;
3157
3158 if (instr->intrinsic == nir_intrinsic_discard_if) {
3159 cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
3160 get_src(ctx, instr->src[0]),
3161 ctx->ac.i32_0, "");
3162 } else {
3163 assert(instr->intrinsic == nir_intrinsic_discard);
3164 cond = ctx->ac.i1false;
3165 }
3166
3167 ac_build_kill_if_false(&ctx->ac, cond);
3168 }
3169
3170 static void emit_demote(struct ac_nir_context *ctx,
3171 const nir_intrinsic_instr *instr)
3172 {
3173 LLVMValueRef cond;
3174
3175 if (instr->intrinsic == nir_intrinsic_demote_if) {
3176 cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
3177 get_src(ctx, instr->src[0]),
3178 ctx->ac.i32_0, "");
3179 } else {
3180 assert(instr->intrinsic == nir_intrinsic_demote);
3181 cond = ctx->ac.i1false;
3182 }
3183
3184 /* Kill immediately while maintaining WQM. */
3185 ac_build_kill_if_false(&ctx->ac, ac_build_wqm_vote(&ctx->ac, cond));
3186
3187 LLVMValueRef mask = LLVMBuildLoad(ctx->ac.builder, ctx->ac.postponed_kill, "");
3188 mask = LLVMBuildAnd(ctx->ac.builder, mask, cond, "");
3189 LLVMBuildStore(ctx->ac.builder, mask, ctx->ac.postponed_kill);
3190 return;
3191 }
3192
3193 static LLVMValueRef
3194 visit_load_local_invocation_index(struct ac_nir_context *ctx)
3195 {
3196 LLVMValueRef result;
3197 LLVMValueRef thread_id = ac_get_thread_id(&ctx->ac);
3198 result = LLVMBuildAnd(ctx->ac.builder,
3199 ac_get_arg(&ctx->ac, ctx->args->tg_size),
3200 LLVMConstInt(ctx->ac.i32, 0xfc0, false), "");
3201
3202 if (ctx->ac.wave_size == 32)
3203 result = LLVMBuildLShr(ctx->ac.builder, result,
3204 LLVMConstInt(ctx->ac.i32, 1, false), "");
3205
3206 return LLVMBuildAdd(ctx->ac.builder, result, thread_id, "");
3207 }
3208
3209 static LLVMValueRef
3210 visit_load_subgroup_id(struct ac_nir_context *ctx)
3211 {
3212 if (ctx->stage == MESA_SHADER_COMPUTE) {
3213 LLVMValueRef result;
3214 result = LLVMBuildAnd(ctx->ac.builder,
3215 ac_get_arg(&ctx->ac, ctx->args->tg_size),
3216 LLVMConstInt(ctx->ac.i32, 0xfc0, false), "");
3217 return LLVMBuildLShr(ctx->ac.builder, result, LLVMConstInt(ctx->ac.i32, 6, false), "");
3218 } else {
3219 return LLVMConstInt(ctx->ac.i32, 0, false);
3220 }
3221 }
3222
3223 static LLVMValueRef
3224 visit_load_num_subgroups(struct ac_nir_context *ctx)
3225 {
3226 if (ctx->stage == MESA_SHADER_COMPUTE) {
3227 return LLVMBuildAnd(ctx->ac.builder,
3228 ac_get_arg(&ctx->ac, ctx->args->tg_size),
3229 LLVMConstInt(ctx->ac.i32, 0x3f, false), "");
3230 } else {
3231 return LLVMConstInt(ctx->ac.i32, 1, false);
3232 }
3233 }
3234
3235 static LLVMValueRef
3236 visit_first_invocation(struct ac_nir_context *ctx)
3237 {
3238 LLVMValueRef active_set = ac_build_ballot(&ctx->ac, ctx->ac.i32_1);
3239 const char *intr = ctx->ac.wave_size == 32 ? "llvm.cttz.i32" : "llvm.cttz.i64";
3240
3241 /* The second argument is whether cttz(0) should be defined, but we do not care. */
3242 LLVMValueRef args[] = {active_set, ctx->ac.i1false};
3243 LLVMValueRef result = ac_build_intrinsic(&ctx->ac, intr,
3244 ctx->ac.iN_wavemask, args, 2,
3245 AC_FUNC_ATTR_NOUNWIND |
3246 AC_FUNC_ATTR_READNONE);
3247
3248 return LLVMBuildTrunc(ctx->ac.builder, result, ctx->ac.i32, "");
3249 }
3250
3251 static LLVMValueRef
3252 visit_load_shared(struct ac_nir_context *ctx,
3253 const nir_intrinsic_instr *instr)
3254 {
3255 LLVMValueRef values[4], derived_ptr, index, ret;
3256
3257 LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[0],
3258 instr->dest.ssa.bit_size);
3259
3260 for (int chan = 0; chan < instr->num_components; chan++) {
3261 index = LLVMConstInt(ctx->ac.i32, chan, 0);
3262 derived_ptr = LLVMBuildGEP(ctx->ac.builder, ptr, &index, 1, "");
3263 values[chan] = LLVMBuildLoad(ctx->ac.builder, derived_ptr, "");
3264 }
3265
3266 ret = ac_build_gather_values(&ctx->ac, values, instr->num_components);
3267 return LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->dest.ssa), "");
3268 }
3269
3270 static void
3271 visit_store_shared(struct ac_nir_context *ctx,
3272 const nir_intrinsic_instr *instr)
3273 {
3274 LLVMValueRef derived_ptr, data,index;
3275 LLVMBuilderRef builder = ctx->ac.builder;
3276
3277 LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[1],
3278 instr->src[0].ssa->bit_size);
3279 LLVMValueRef src = get_src(ctx, instr->src[0]);
3280
3281 int writemask = nir_intrinsic_write_mask(instr);
3282 for (int chan = 0; chan < 4; chan++) {
3283 if (!(writemask & (1 << chan))) {
3284 continue;
3285 }
3286 data = ac_llvm_extract_elem(&ctx->ac, src, chan);
3287 index = LLVMConstInt(ctx->ac.i32, chan, 0);
3288 derived_ptr = LLVMBuildGEP(builder, ptr, &index, 1, "");
3289 LLVMBuildStore(builder, data, derived_ptr);
3290 }
3291 }
3292
3293 static LLVMValueRef visit_var_atomic(struct ac_nir_context *ctx,
3294 const nir_intrinsic_instr *instr,
3295 LLVMValueRef ptr, int src_idx)
3296 {
3297 if (ctx->ac.postponed_kill) {
3298 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
3299 ctx->ac.postponed_kill, "");
3300 ac_build_ifcc(&ctx->ac, cond, 7005);
3301 }
3302
3303 LLVMValueRef result;
3304 LLVMValueRef src = get_src(ctx, instr->src[src_idx]);
3305
3306 const char *sync_scope = LLVM_VERSION_MAJOR >= 9 ? "workgroup-one-as" : "workgroup";
3307
3308 if (instr->src[0].ssa->parent_instr->type == nir_instr_type_deref) {
3309 nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
3310 if (deref->mode == nir_var_mem_global) {
3311 /* use "singlethread" sync scope to implement relaxed ordering */
3312 sync_scope = LLVM_VERSION_MAJOR >= 9 ? "singlethread-one-as" : "singlethread";
3313
3314 LLVMTypeRef ptr_type = LLVMPointerType(LLVMTypeOf(src), LLVMGetPointerAddressSpace(LLVMTypeOf(ptr)));
3315 ptr = LLVMBuildBitCast(ctx->ac.builder, ptr, ptr_type , "");
3316 }
3317 }
3318
3319 if (instr->intrinsic == nir_intrinsic_shared_atomic_comp_swap ||
3320 instr->intrinsic == nir_intrinsic_deref_atomic_comp_swap) {
3321 LLVMValueRef src1 = get_src(ctx, instr->src[src_idx + 1]);
3322 result = ac_build_atomic_cmp_xchg(&ctx->ac, ptr, src, src1, sync_scope);
3323 result = LLVMBuildExtractValue(ctx->ac.builder, result, 0, "");
3324 } else {
3325 LLVMAtomicRMWBinOp op;
3326 switch (instr->intrinsic) {
3327 case nir_intrinsic_shared_atomic_add:
3328 case nir_intrinsic_deref_atomic_add:
3329 op = LLVMAtomicRMWBinOpAdd;
3330 break;
3331 case nir_intrinsic_shared_atomic_umin:
3332 case nir_intrinsic_deref_atomic_umin:
3333 op = LLVMAtomicRMWBinOpUMin;
3334 break;
3335 case nir_intrinsic_shared_atomic_umax:
3336 case nir_intrinsic_deref_atomic_umax:
3337 op = LLVMAtomicRMWBinOpUMax;
3338 break;
3339 case nir_intrinsic_shared_atomic_imin:
3340 case nir_intrinsic_deref_atomic_imin:
3341 op = LLVMAtomicRMWBinOpMin;
3342 break;
3343 case nir_intrinsic_shared_atomic_imax:
3344 case nir_intrinsic_deref_atomic_imax:
3345 op = LLVMAtomicRMWBinOpMax;
3346 break;
3347 case nir_intrinsic_shared_atomic_and:
3348 case nir_intrinsic_deref_atomic_and:
3349 op = LLVMAtomicRMWBinOpAnd;
3350 break;
3351 case nir_intrinsic_shared_atomic_or:
3352 case nir_intrinsic_deref_atomic_or:
3353 op = LLVMAtomicRMWBinOpOr;
3354 break;
3355 case nir_intrinsic_shared_atomic_xor:
3356 case nir_intrinsic_deref_atomic_xor:
3357 op = LLVMAtomicRMWBinOpXor;
3358 break;
3359 case nir_intrinsic_shared_atomic_exchange:
3360 case nir_intrinsic_deref_atomic_exchange:
3361 op = LLVMAtomicRMWBinOpXchg;
3362 break;
3363 default:
3364 return NULL;
3365 }
3366
3367 result = ac_build_atomic_rmw(&ctx->ac, op, ptr, ac_to_integer(&ctx->ac, src), sync_scope);
3368 }
3369
3370 if (ctx->ac.postponed_kill)
3371 ac_build_endif(&ctx->ac, 7005);
3372 return result;
3373 }
3374
3375 static LLVMValueRef load_sample_pos(struct ac_nir_context *ctx)
3376 {
3377 LLVMValueRef values[2];
3378 LLVMValueRef pos[2];
3379
3380 pos[0] = ac_to_float(&ctx->ac,
3381 ac_get_arg(&ctx->ac, ctx->args->frag_pos[0]));
3382 pos[1] = ac_to_float(&ctx->ac,
3383 ac_get_arg(&ctx->ac, ctx->args->frag_pos[1]));
3384
3385 values[0] = ac_build_fract(&ctx->ac, pos[0], 32);
3386 values[1] = ac_build_fract(&ctx->ac, pos[1], 32);
3387 return ac_build_gather_values(&ctx->ac, values, 2);
3388 }
3389
3390 static LLVMValueRef lookup_interp_param(struct ac_nir_context *ctx,
3391 enum glsl_interp_mode interp, unsigned location)
3392 {
3393 switch (interp) {
3394 case INTERP_MODE_FLAT:
3395 default:
3396 return NULL;
3397 case INTERP_MODE_SMOOTH:
3398 case INTERP_MODE_NONE:
3399 if (location == INTERP_CENTER)
3400 return ac_get_arg(&ctx->ac, ctx->args->persp_center);
3401 else if (location == INTERP_CENTROID)
3402 return ctx->abi->persp_centroid;
3403 else if (location == INTERP_SAMPLE)
3404 return ac_get_arg(&ctx->ac, ctx->args->persp_sample);
3405 break;
3406 case INTERP_MODE_NOPERSPECTIVE:
3407 if (location == INTERP_CENTER)
3408 return ac_get_arg(&ctx->ac, ctx->args->linear_center);
3409 else if (location == INTERP_CENTROID)
3410 return ctx->abi->linear_centroid;
3411 else if (location == INTERP_SAMPLE)
3412 return ac_get_arg(&ctx->ac, ctx->args->linear_sample);
3413 break;
3414 }
3415 return NULL;
3416 }
3417
3418 static LLVMValueRef barycentric_center(struct ac_nir_context *ctx,
3419 unsigned mode)
3420 {
3421 LLVMValueRef interp_param = lookup_interp_param(ctx, mode, INTERP_CENTER);
3422 return LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2i32, "");
3423 }
3424
3425 static LLVMValueRef barycentric_offset(struct ac_nir_context *ctx,
3426 unsigned mode,
3427 LLVMValueRef offset)
3428 {
3429 LLVMValueRef interp_param = lookup_interp_param(ctx, mode, INTERP_CENTER);
3430 LLVMValueRef src_c0 = ac_to_float(&ctx->ac, LLVMBuildExtractElement(ctx->ac.builder, offset, ctx->ac.i32_0, ""));
3431 LLVMValueRef src_c1 = ac_to_float(&ctx->ac, LLVMBuildExtractElement(ctx->ac.builder, offset, ctx->ac.i32_1, ""));
3432
3433 LLVMValueRef ij_out[2];
3434 LLVMValueRef ddxy_out = ac_build_ddxy_interp(&ctx->ac, interp_param);
3435
3436 /*
3437 * take the I then J parameters, and the DDX/Y for it, and
3438 * calculate the IJ inputs for the interpolator.
3439 * temp1 = ddx * offset/sample.x + I;
3440 * interp_param.I = ddy * offset/sample.y + temp1;
3441 * temp1 = ddx * offset/sample.x + J;
3442 * interp_param.J = ddy * offset/sample.y + temp1;
3443 */
3444 for (unsigned i = 0; i < 2; i++) {
3445 LLVMValueRef ix_ll = LLVMConstInt(ctx->ac.i32, i, false);
3446 LLVMValueRef iy_ll = LLVMConstInt(ctx->ac.i32, i + 2, false);
3447 LLVMValueRef ddx_el = LLVMBuildExtractElement(ctx->ac.builder,
3448 ddxy_out, ix_ll, "");
3449 LLVMValueRef ddy_el = LLVMBuildExtractElement(ctx->ac.builder,
3450 ddxy_out, iy_ll, "");
3451 LLVMValueRef interp_el = LLVMBuildExtractElement(ctx->ac.builder,
3452 interp_param, ix_ll, "");
3453 LLVMValueRef temp1, temp2;
3454
3455 interp_el = LLVMBuildBitCast(ctx->ac.builder, interp_el,
3456 ctx->ac.f32, "");
3457
3458 temp1 = ac_build_fmad(&ctx->ac, ddx_el, src_c0, interp_el);
3459 temp2 = ac_build_fmad(&ctx->ac, ddy_el, src_c1, temp1);
3460
3461 ij_out[i] = LLVMBuildBitCast(ctx->ac.builder,
3462 temp2, ctx->ac.i32, "");
3463 }
3464 interp_param = ac_build_gather_values(&ctx->ac, ij_out, 2);
3465 return LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2i32, "");
3466 }
3467
3468 static LLVMValueRef barycentric_centroid(struct ac_nir_context *ctx,
3469 unsigned mode)
3470 {
3471 LLVMValueRef interp_param = lookup_interp_param(ctx, mode, INTERP_CENTROID);
3472 return LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2i32, "");
3473 }
3474
3475 static LLVMValueRef barycentric_at_sample(struct ac_nir_context *ctx,
3476 unsigned mode,
3477 LLVMValueRef sample_id)
3478 {
3479 if (ctx->abi->interp_at_sample_force_center)
3480 return barycentric_center(ctx, mode);
3481
3482 LLVMValueRef halfval = LLVMConstReal(ctx->ac.f32, 0.5f);
3483
3484 /* fetch sample ID */
3485 LLVMValueRef sample_pos = ctx->abi->load_sample_position(ctx->abi, sample_id);
3486
3487 LLVMValueRef src_c0 = LLVMBuildExtractElement(ctx->ac.builder, sample_pos, ctx->ac.i32_0, "");
3488 src_c0 = LLVMBuildFSub(ctx->ac.builder, src_c0, halfval, "");
3489 LLVMValueRef src_c1 = LLVMBuildExtractElement(ctx->ac.builder, sample_pos, ctx->ac.i32_1, "");
3490 src_c1 = LLVMBuildFSub(ctx->ac.builder, src_c1, halfval, "");
3491 LLVMValueRef coords[] = { src_c0, src_c1 };
3492 LLVMValueRef offset = ac_build_gather_values(&ctx->ac, coords, 2);
3493
3494 return barycentric_offset(ctx, mode, offset);
3495 }
3496
3497
3498 static LLVMValueRef barycentric_sample(struct ac_nir_context *ctx,
3499 unsigned mode)
3500 {
3501 LLVMValueRef interp_param = lookup_interp_param(ctx, mode, INTERP_SAMPLE);
3502 return LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2i32, "");
3503 }
3504
3505 static LLVMValueRef barycentric_model(struct ac_nir_context *ctx)
3506 {
3507 return LLVMBuildBitCast(ctx->ac.builder,
3508 ac_get_arg(&ctx->ac, ctx->args->pull_model),
3509 ctx->ac.v3i32, "");
3510 }
3511
3512 static LLVMValueRef load_interpolated_input(struct ac_nir_context *ctx,
3513 LLVMValueRef interp_param,
3514 unsigned index, unsigned comp_start,
3515 unsigned num_components,
3516 unsigned bitsize)
3517 {
3518 LLVMValueRef attr_number = LLVMConstInt(ctx->ac.i32, index, false);
3519 LLVMValueRef interp_param_f;
3520
3521 interp_param_f = LLVMBuildBitCast(ctx->ac.builder,
3522 interp_param, ctx->ac.v2f32, "");
3523 LLVMValueRef i = LLVMBuildExtractElement(
3524 ctx->ac.builder, interp_param_f, ctx->ac.i32_0, "");
3525 LLVMValueRef j = LLVMBuildExtractElement(
3526 ctx->ac.builder, interp_param_f, ctx->ac.i32_1, "");
3527
3528 /* Workaround for issue 2647: kill threads with infinite interpolation coeffs */
3529 if (ctx->verified_interp &&
3530 !_mesa_hash_table_search(ctx->verified_interp, interp_param)) {
3531 LLVMValueRef args[2];
3532 args[0] = i;
3533 args[1] = LLVMConstInt(ctx->ac.i32, S_NAN | Q_NAN | N_INFINITY | P_INFINITY, false);
3534 LLVMValueRef cond = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.class.f32", ctx->ac.i1,
3535 args, 2, AC_FUNC_ATTR_READNONE);
3536 ac_build_kill_if_false(&ctx->ac, LLVMBuildNot(ctx->ac.builder, cond, ""));
3537 _mesa_hash_table_insert(ctx->verified_interp, interp_param, interp_param);
3538 }
3539
3540 LLVMValueRef values[4];
3541 assert(bitsize == 16 || bitsize == 32);
3542 for (unsigned comp = 0; comp < num_components; comp++) {
3543 LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, comp_start + comp, false);
3544 if (bitsize == 16) {
3545 values[comp] = ac_build_fs_interp_f16(&ctx->ac, llvm_chan, attr_number,
3546 ac_get_arg(&ctx->ac, ctx->args->prim_mask), i, j);
3547 } else {
3548 values[comp] = ac_build_fs_interp(&ctx->ac, llvm_chan, attr_number,
3549 ac_get_arg(&ctx->ac, ctx->args->prim_mask), i, j);
3550 }
3551 }
3552
3553 return ac_to_integer(&ctx->ac, ac_build_gather_values(&ctx->ac, values, num_components));
3554 }
3555
3556 static LLVMValueRef load_input(struct ac_nir_context *ctx,
3557 nir_intrinsic_instr *instr)
3558 {
3559 unsigned offset_idx = instr->intrinsic == nir_intrinsic_load_input ? 0 : 1;
3560
3561 /* We only lower inputs for fragment shaders ATM */
3562 ASSERTED nir_const_value *offset = nir_src_as_const_value(instr->src[offset_idx]);
3563 assert(offset);
3564 assert(offset[0].i32 == 0);
3565
3566 unsigned component = nir_intrinsic_component(instr);
3567 unsigned index = nir_intrinsic_base(instr);
3568 unsigned vertex_id = 2; /* P0 */
3569
3570 if (instr->intrinsic == nir_intrinsic_load_input_vertex) {
3571 nir_const_value *src0 = nir_src_as_const_value(instr->src[0]);
3572
3573 switch (src0[0].i32) {
3574 case 0:
3575 vertex_id = 2;
3576 break;
3577 case 1:
3578 vertex_id = 0;
3579 break;
3580 case 2:
3581 vertex_id = 1;
3582 break;
3583 default:
3584 unreachable("Invalid vertex index");
3585 }
3586 }
3587
3588 LLVMValueRef attr_number = LLVMConstInt(ctx->ac.i32, index, false);
3589 LLVMValueRef values[8];
3590
3591 /* Each component of a 64-bit value takes up two GL-level channels. */
3592 unsigned num_components = instr->dest.ssa.num_components;
3593 unsigned bit_size = instr->dest.ssa.bit_size;
3594 unsigned channels =
3595 bit_size == 64 ? num_components * 2 : num_components;
3596
3597 for (unsigned chan = 0; chan < channels; chan++) {
3598 if (component + chan > 4)
3599 attr_number = LLVMConstInt(ctx->ac.i32, index + 1, false);
3600 LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, (component + chan) % 4, false);
3601 values[chan] = ac_build_fs_interp_mov(&ctx->ac,
3602 LLVMConstInt(ctx->ac.i32, vertex_id, false),
3603 llvm_chan,
3604 attr_number,
3605 ac_get_arg(&ctx->ac, ctx->args->prim_mask));
3606 values[chan] = LLVMBuildBitCast(ctx->ac.builder, values[chan], ctx->ac.i32, "");
3607 values[chan] = LLVMBuildTruncOrBitCast(ctx->ac.builder, values[chan],
3608 bit_size == 16 ? ctx->ac.i16 : ctx->ac.i32, "");
3609 }
3610
3611 LLVMValueRef result = ac_build_gather_values(&ctx->ac, values, channels);
3612 if (bit_size == 64) {
3613 LLVMTypeRef type = num_components == 1 ? ctx->ac.i64 :
3614 LLVMVectorType(ctx->ac.i64, num_components);
3615 result = LLVMBuildBitCast(ctx->ac.builder, result, type, "");
3616 }
3617 return result;
3618 }
3619
3620 static void visit_intrinsic(struct ac_nir_context *ctx,
3621 nir_intrinsic_instr *instr)
3622 {
3623 LLVMValueRef result = NULL;
3624
3625 switch (instr->intrinsic) {
3626 case nir_intrinsic_ballot:
3627 result = ac_build_ballot(&ctx->ac, get_src(ctx, instr->src[0]));
3628 if (ctx->ac.ballot_mask_bits > ctx->ac.wave_size)
3629 result = LLVMBuildZExt(ctx->ac.builder, result, ctx->ac.iN_ballotmask, "");
3630 break;
3631 case nir_intrinsic_read_invocation:
3632 result = ac_build_readlane(&ctx->ac, get_src(ctx, instr->src[0]),
3633 get_src(ctx, instr->src[1]));
3634 break;
3635 case nir_intrinsic_read_first_invocation:
3636 result = ac_build_readlane(&ctx->ac, get_src(ctx, instr->src[0]), NULL);
3637 break;
3638 case nir_intrinsic_load_subgroup_invocation:
3639 result = ac_get_thread_id(&ctx->ac);
3640 break;
3641 case nir_intrinsic_load_work_group_id: {
3642 LLVMValueRef values[3];
3643
3644 for (int i = 0; i < 3; i++) {
3645 values[i] = ctx->args->workgroup_ids[i].used ?
3646 ac_get_arg(&ctx->ac, ctx->args->workgroup_ids[i]) : ctx->ac.i32_0;
3647 }
3648
3649 result = ac_build_gather_values(&ctx->ac, values, 3);
3650 break;
3651 }
3652 case nir_intrinsic_load_base_vertex:
3653 case nir_intrinsic_load_first_vertex:
3654 result = ctx->abi->load_base_vertex(ctx->abi);
3655 break;
3656 case nir_intrinsic_load_local_group_size:
3657 result = ctx->abi->load_local_group_size(ctx->abi);
3658 break;
3659 case nir_intrinsic_load_vertex_id:
3660 result = LLVMBuildAdd(ctx->ac.builder,
3661 ac_get_arg(&ctx->ac, ctx->args->vertex_id),
3662 ac_get_arg(&ctx->ac, ctx->args->base_vertex), "");
3663 break;
3664 case nir_intrinsic_load_vertex_id_zero_base: {
3665 result = ctx->abi->vertex_id;
3666 break;
3667 }
3668 case nir_intrinsic_load_local_invocation_id: {
3669 result = ac_get_arg(&ctx->ac, ctx->args->local_invocation_ids);
3670 break;
3671 }
3672 case nir_intrinsic_load_base_instance:
3673 result = ac_get_arg(&ctx->ac, ctx->args->start_instance);
3674 break;
3675 case nir_intrinsic_load_draw_id:
3676 result = ac_get_arg(&ctx->ac, ctx->args->draw_id);
3677 break;
3678 case nir_intrinsic_load_view_index:
3679 result = ac_get_arg(&ctx->ac, ctx->args->view_index);
3680 break;
3681 case nir_intrinsic_load_invocation_id:
3682 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
3683 result = ac_unpack_param(&ctx->ac,
3684 ac_get_arg(&ctx->ac, ctx->args->tcs_rel_ids),
3685 8, 5);
3686 } else {
3687 if (ctx->ac.chip_class >= GFX10) {
3688 result = LLVMBuildAnd(ctx->ac.builder,
3689 ac_get_arg(&ctx->ac, ctx->args->gs_invocation_id),
3690 LLVMConstInt(ctx->ac.i32, 127, 0), "");
3691 } else {
3692 result = ac_get_arg(&ctx->ac, ctx->args->gs_invocation_id);
3693 }
3694 }
3695 break;
3696 case nir_intrinsic_load_primitive_id:
3697 if (ctx->stage == MESA_SHADER_GEOMETRY) {
3698 result = ac_get_arg(&ctx->ac, ctx->args->gs_prim_id);
3699 } else if (ctx->stage == MESA_SHADER_TESS_CTRL) {
3700 result = ac_get_arg(&ctx->ac, ctx->args->tcs_patch_id);
3701 } else if (ctx->stage == MESA_SHADER_TESS_EVAL) {
3702 result = ac_get_arg(&ctx->ac, ctx->args->tes_patch_id);
3703 } else
3704 fprintf(stderr, "Unknown primitive id intrinsic: %d", ctx->stage);
3705 break;
3706 case nir_intrinsic_load_sample_id:
3707 result = ac_unpack_param(&ctx->ac,
3708 ac_get_arg(&ctx->ac, ctx->args->ancillary),
3709 8, 4);
3710 break;
3711 case nir_intrinsic_load_sample_pos:
3712 result = load_sample_pos(ctx);
3713 break;
3714 case nir_intrinsic_load_sample_mask_in:
3715 result = ctx->abi->load_sample_mask_in(ctx->abi);
3716 break;
3717 case nir_intrinsic_load_frag_coord: {
3718 LLVMValueRef values[4] = {
3719 ac_get_arg(&ctx->ac, ctx->args->frag_pos[0]),
3720 ac_get_arg(&ctx->ac, ctx->args->frag_pos[1]),
3721 ac_get_arg(&ctx->ac, ctx->args->frag_pos[2]),
3722 ac_build_fdiv(&ctx->ac, ctx->ac.f32_1,
3723 ac_get_arg(&ctx->ac, ctx->args->frag_pos[3]))
3724 };
3725 result = ac_to_integer(&ctx->ac,
3726 ac_build_gather_values(&ctx->ac, values, 4));
3727 break;
3728 }
3729 case nir_intrinsic_load_layer_id:
3730 result = ctx->abi->inputs[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER, 0)];
3731 break;
3732 case nir_intrinsic_load_front_face:
3733 result = ac_get_arg(&ctx->ac, ctx->args->front_face);
3734 break;
3735 case nir_intrinsic_load_helper_invocation:
3736 result = ac_build_load_helper_invocation(&ctx->ac);
3737 break;
3738 case nir_intrinsic_is_helper_invocation:
3739 result = ac_build_is_helper_invocation(&ctx->ac);
3740 break;
3741 case nir_intrinsic_load_color0:
3742 result = ctx->abi->color0;
3743 break;
3744 case nir_intrinsic_load_color1:
3745 result = ctx->abi->color1;
3746 break;
3747 case nir_intrinsic_load_user_data_amd:
3748 assert(LLVMTypeOf(ctx->abi->user_data) == ctx->ac.v4i32);
3749 result = ctx->abi->user_data;
3750 break;
3751 case nir_intrinsic_load_instance_id:
3752 result = ctx->abi->instance_id;
3753 break;
3754 case nir_intrinsic_load_num_work_groups:
3755 result = ac_get_arg(&ctx->ac, ctx->args->num_work_groups);
3756 break;
3757 case nir_intrinsic_load_local_invocation_index:
3758 result = visit_load_local_invocation_index(ctx);
3759 break;
3760 case nir_intrinsic_load_subgroup_id:
3761 result = visit_load_subgroup_id(ctx);
3762 break;
3763 case nir_intrinsic_load_num_subgroups:
3764 result = visit_load_num_subgroups(ctx);
3765 break;
3766 case nir_intrinsic_first_invocation:
3767 result = visit_first_invocation(ctx);
3768 break;
3769 case nir_intrinsic_load_push_constant:
3770 result = visit_load_push_constant(ctx, instr);
3771 break;
3772 case nir_intrinsic_vulkan_resource_index: {
3773 LLVMValueRef index = get_src(ctx, instr->src[0]);
3774 unsigned desc_set = nir_intrinsic_desc_set(instr);
3775 unsigned binding = nir_intrinsic_binding(instr);
3776
3777 result = ctx->abi->load_resource(ctx->abi, index, desc_set,
3778 binding);
3779 break;
3780 }
3781 case nir_intrinsic_vulkan_resource_reindex:
3782 result = visit_vulkan_resource_reindex(ctx, instr);
3783 break;
3784 case nir_intrinsic_store_ssbo:
3785 visit_store_ssbo(ctx, instr);
3786 break;
3787 case nir_intrinsic_load_ssbo:
3788 result = visit_load_buffer(ctx, instr);
3789 break;
3790 case nir_intrinsic_ssbo_atomic_add:
3791 case nir_intrinsic_ssbo_atomic_imin:
3792 case nir_intrinsic_ssbo_atomic_umin:
3793 case nir_intrinsic_ssbo_atomic_imax:
3794 case nir_intrinsic_ssbo_atomic_umax:
3795 case nir_intrinsic_ssbo_atomic_and:
3796 case nir_intrinsic_ssbo_atomic_or:
3797 case nir_intrinsic_ssbo_atomic_xor:
3798 case nir_intrinsic_ssbo_atomic_exchange:
3799 case nir_intrinsic_ssbo_atomic_comp_swap:
3800 result = visit_atomic_ssbo(ctx, instr);
3801 break;
3802 case nir_intrinsic_load_ubo:
3803 result = visit_load_ubo_buffer(ctx, instr);
3804 break;
3805 case nir_intrinsic_get_buffer_size:
3806 result = visit_get_buffer_size(ctx, instr);
3807 break;
3808 case nir_intrinsic_load_deref:
3809 result = visit_load_var(ctx, instr);
3810 break;
3811 case nir_intrinsic_store_deref:
3812 visit_store_var(ctx, instr);
3813 break;
3814 case nir_intrinsic_load_shared:
3815 result = visit_load_shared(ctx, instr);
3816 break;
3817 case nir_intrinsic_store_shared:
3818 visit_store_shared(ctx, instr);
3819 break;
3820 case nir_intrinsic_bindless_image_samples:
3821 case nir_intrinsic_image_deref_samples:
3822 result = visit_image_samples(ctx, instr);
3823 break;
3824 case nir_intrinsic_bindless_image_load:
3825 result = visit_image_load(ctx, instr, true);
3826 break;
3827 case nir_intrinsic_image_deref_load:
3828 result = visit_image_load(ctx, instr, false);
3829 break;
3830 case nir_intrinsic_bindless_image_store:
3831 visit_image_store(ctx, instr, true);
3832 break;
3833 case nir_intrinsic_image_deref_store:
3834 visit_image_store(ctx, instr, false);
3835 break;
3836 case nir_intrinsic_bindless_image_atomic_add:
3837 case nir_intrinsic_bindless_image_atomic_imin:
3838 case nir_intrinsic_bindless_image_atomic_umin:
3839 case nir_intrinsic_bindless_image_atomic_imax:
3840 case nir_intrinsic_bindless_image_atomic_umax:
3841 case nir_intrinsic_bindless_image_atomic_and:
3842 case nir_intrinsic_bindless_image_atomic_or:
3843 case nir_intrinsic_bindless_image_atomic_xor:
3844 case nir_intrinsic_bindless_image_atomic_exchange:
3845 case nir_intrinsic_bindless_image_atomic_comp_swap:
3846 case nir_intrinsic_bindless_image_atomic_inc_wrap:
3847 case nir_intrinsic_bindless_image_atomic_dec_wrap:
3848 result = visit_image_atomic(ctx, instr, true);
3849 break;
3850 case nir_intrinsic_image_deref_atomic_add:
3851 case nir_intrinsic_image_deref_atomic_imin:
3852 case nir_intrinsic_image_deref_atomic_umin:
3853 case nir_intrinsic_image_deref_atomic_imax:
3854 case nir_intrinsic_image_deref_atomic_umax:
3855 case nir_intrinsic_image_deref_atomic_and:
3856 case nir_intrinsic_image_deref_atomic_or:
3857 case nir_intrinsic_image_deref_atomic_xor:
3858 case nir_intrinsic_image_deref_atomic_exchange:
3859 case nir_intrinsic_image_deref_atomic_comp_swap:
3860 case nir_intrinsic_image_deref_atomic_inc_wrap:
3861 case nir_intrinsic_image_deref_atomic_dec_wrap:
3862 result = visit_image_atomic(ctx, instr, false);
3863 break;
3864 case nir_intrinsic_bindless_image_size:
3865 result = visit_image_size(ctx, instr, true);
3866 break;
3867 case nir_intrinsic_image_deref_size:
3868 result = visit_image_size(ctx, instr, false);
3869 break;
3870 case nir_intrinsic_shader_clock:
3871 result = ac_build_shader_clock(&ctx->ac,
3872 nir_intrinsic_memory_scope(instr));
3873 break;
3874 case nir_intrinsic_discard:
3875 case nir_intrinsic_discard_if:
3876 emit_discard(ctx, instr);
3877 break;
3878 case nir_intrinsic_demote:
3879 case nir_intrinsic_demote_if:
3880 emit_demote(ctx, instr);
3881 break;
3882 case nir_intrinsic_memory_barrier:
3883 case nir_intrinsic_group_memory_barrier:
3884 case nir_intrinsic_memory_barrier_buffer:
3885 case nir_intrinsic_memory_barrier_image:
3886 case nir_intrinsic_memory_barrier_shared:
3887 emit_membar(&ctx->ac, instr);
3888 break;
3889 case nir_intrinsic_memory_barrier_tcs_patch:
3890 break;
3891 case nir_intrinsic_control_barrier:
3892 ac_emit_barrier(&ctx->ac, ctx->stage);
3893 break;
3894 case nir_intrinsic_shared_atomic_add:
3895 case nir_intrinsic_shared_atomic_imin:
3896 case nir_intrinsic_shared_atomic_umin:
3897 case nir_intrinsic_shared_atomic_imax:
3898 case nir_intrinsic_shared_atomic_umax:
3899 case nir_intrinsic_shared_atomic_and:
3900 case nir_intrinsic_shared_atomic_or:
3901 case nir_intrinsic_shared_atomic_xor:
3902 case nir_intrinsic_shared_atomic_exchange:
3903 case nir_intrinsic_shared_atomic_comp_swap: {
3904 LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[0],
3905 instr->src[1].ssa->bit_size);
3906 result = visit_var_atomic(ctx, instr, ptr, 1);
3907 break;
3908 }
3909 case nir_intrinsic_deref_atomic_add:
3910 case nir_intrinsic_deref_atomic_imin:
3911 case nir_intrinsic_deref_atomic_umin:
3912 case nir_intrinsic_deref_atomic_imax:
3913 case nir_intrinsic_deref_atomic_umax:
3914 case nir_intrinsic_deref_atomic_and:
3915 case nir_intrinsic_deref_atomic_or:
3916 case nir_intrinsic_deref_atomic_xor:
3917 case nir_intrinsic_deref_atomic_exchange:
3918 case nir_intrinsic_deref_atomic_comp_swap: {
3919 LLVMValueRef ptr = get_src(ctx, instr->src[0]);
3920 result = visit_var_atomic(ctx, instr, ptr, 1);
3921 break;
3922 }
3923 case nir_intrinsic_load_barycentric_pixel:
3924 result = barycentric_center(ctx, nir_intrinsic_interp_mode(instr));
3925 break;
3926 case nir_intrinsic_load_barycentric_centroid:
3927 result = barycentric_centroid(ctx, nir_intrinsic_interp_mode(instr));
3928 break;
3929 case nir_intrinsic_load_barycentric_sample:
3930 result = barycentric_sample(ctx, nir_intrinsic_interp_mode(instr));
3931 break;
3932 case nir_intrinsic_load_barycentric_model:
3933 result = barycentric_model(ctx);
3934 break;
3935 case nir_intrinsic_load_barycentric_at_offset: {
3936 LLVMValueRef offset = ac_to_float(&ctx->ac, get_src(ctx, instr->src[0]));
3937 result = barycentric_offset(ctx, nir_intrinsic_interp_mode(instr), offset);
3938 break;
3939 }
3940 case nir_intrinsic_load_barycentric_at_sample: {
3941 LLVMValueRef sample_id = get_src(ctx, instr->src[0]);
3942 result = barycentric_at_sample(ctx, nir_intrinsic_interp_mode(instr), sample_id);
3943 break;
3944 }
3945 case nir_intrinsic_load_interpolated_input: {
3946 /* We assume any indirect loads have been lowered away */
3947 ASSERTED nir_const_value *offset = nir_src_as_const_value(instr->src[1]);
3948 assert(offset);
3949 assert(offset[0].i32 == 0);
3950
3951 LLVMValueRef interp_param = get_src(ctx, instr->src[0]);
3952 unsigned index = nir_intrinsic_base(instr);
3953 unsigned component = nir_intrinsic_component(instr);
3954 result = load_interpolated_input(ctx, interp_param, index,
3955 component,
3956 instr->dest.ssa.num_components,
3957 instr->dest.ssa.bit_size);
3958 break;
3959 }
3960 case nir_intrinsic_load_input:
3961 case nir_intrinsic_load_input_vertex:
3962 result = load_input(ctx, instr);
3963 break;
3964 case nir_intrinsic_emit_vertex:
3965 ctx->abi->emit_vertex(ctx->abi, nir_intrinsic_stream_id(instr), ctx->abi->outputs);
3966 break;
3967 case nir_intrinsic_emit_vertex_with_counter: {
3968 unsigned stream = nir_intrinsic_stream_id(instr);
3969 LLVMValueRef next_vertex = get_src(ctx, instr->src[0]);
3970 ctx->abi->emit_vertex_with_counter(ctx->abi, stream,
3971 next_vertex,
3972 ctx->abi->outputs);
3973 break;
3974 }
3975 case nir_intrinsic_end_primitive:
3976 case nir_intrinsic_end_primitive_with_counter:
3977 ctx->abi->emit_primitive(ctx->abi, nir_intrinsic_stream_id(instr));
3978 break;
3979 case nir_intrinsic_load_tess_coord:
3980 result = ctx->abi->load_tess_coord(ctx->abi);
3981 break;
3982 case nir_intrinsic_load_tess_level_outer:
3983 result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_OUTER, false);
3984 break;
3985 case nir_intrinsic_load_tess_level_inner:
3986 result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_INNER, false);
3987 break;
3988 case nir_intrinsic_load_tess_level_outer_default:
3989 result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_OUTER, true);
3990 break;
3991 case nir_intrinsic_load_tess_level_inner_default:
3992 result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_INNER, true);
3993 break;
3994 case nir_intrinsic_load_patch_vertices_in:
3995 result = ctx->abi->load_patch_vertices_in(ctx->abi);
3996 break;
3997 case nir_intrinsic_vote_all: {
3998 LLVMValueRef tmp = ac_build_vote_all(&ctx->ac, get_src(ctx, instr->src[0]));
3999 result = LLVMBuildSExt(ctx->ac.builder, tmp, ctx->ac.i32, "");
4000 break;
4001 }
4002 case nir_intrinsic_vote_any: {
4003 LLVMValueRef tmp = ac_build_vote_any(&ctx->ac, get_src(ctx, instr->src[0]));
4004 result = LLVMBuildSExt(ctx->ac.builder, tmp, ctx->ac.i32, "");
4005 break;
4006 }
4007 case nir_intrinsic_shuffle:
4008 if (ctx->ac.chip_class == GFX8 ||
4009 ctx->ac.chip_class == GFX9 ||
4010 (ctx->ac.chip_class == GFX10 && ctx->ac.wave_size == 32)) {
4011 result = ac_build_shuffle(&ctx->ac, get_src(ctx, instr->src[0]),
4012 get_src(ctx, instr->src[1]));
4013 } else {
4014 LLVMValueRef src = get_src(ctx, instr->src[0]);
4015 LLVMValueRef index = get_src(ctx, instr->src[1]);
4016 LLVMTypeRef type = LLVMTypeOf(src);
4017 struct waterfall_context wctx;
4018 LLVMValueRef index_val;
4019
4020 index_val = enter_waterfall(ctx, &wctx, index, true);
4021
4022 src = LLVMBuildZExt(ctx->ac.builder, src,
4023 ctx->ac.i32, "");
4024
4025 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.readlane",
4026 ctx->ac.i32,
4027 (LLVMValueRef []) { src, index_val }, 2,
4028 AC_FUNC_ATTR_READNONE |
4029 AC_FUNC_ATTR_CONVERGENT);
4030
4031 result = LLVMBuildTrunc(ctx->ac.builder, result, type, "");
4032
4033 result = exit_waterfall(ctx, &wctx, result);
4034 }
4035 break;
4036 case nir_intrinsic_reduce:
4037 result = ac_build_reduce(&ctx->ac,
4038 get_src(ctx, instr->src[0]),
4039 instr->const_index[0],
4040 instr->const_index[1]);
4041 break;
4042 case nir_intrinsic_inclusive_scan:
4043 result = ac_build_inclusive_scan(&ctx->ac,
4044 get_src(ctx, instr->src[0]),
4045 instr->const_index[0]);
4046 break;
4047 case nir_intrinsic_exclusive_scan:
4048 result = ac_build_exclusive_scan(&ctx->ac,
4049 get_src(ctx, instr->src[0]),
4050 instr->const_index[0]);
4051 break;
4052 case nir_intrinsic_quad_broadcast: {
4053 unsigned lane = nir_src_as_uint(instr->src[1]);
4054 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]),
4055 lane, lane, lane, lane);
4056 break;
4057 }
4058 case nir_intrinsic_quad_swap_horizontal:
4059 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 1, 0, 3 ,2);
4060 break;
4061 case nir_intrinsic_quad_swap_vertical:
4062 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 2, 3, 0 ,1);
4063 break;
4064 case nir_intrinsic_quad_swap_diagonal:
4065 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 3, 2, 1 ,0);
4066 break;
4067 case nir_intrinsic_quad_swizzle_amd: {
4068 uint32_t mask = nir_intrinsic_swizzle_mask(instr);
4069 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]),
4070 mask & 0x3, (mask >> 2) & 0x3,
4071 (mask >> 4) & 0x3, (mask >> 6) & 0x3);
4072 break;
4073 }
4074 case nir_intrinsic_masked_swizzle_amd: {
4075 uint32_t mask = nir_intrinsic_swizzle_mask(instr);
4076 result = ac_build_ds_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), mask);
4077 break;
4078 }
4079 case nir_intrinsic_write_invocation_amd:
4080 result = ac_build_writelane(&ctx->ac, get_src(ctx, instr->src[0]),
4081 get_src(ctx, instr->src[1]),
4082 get_src(ctx, instr->src[2]));
4083 break;
4084 case nir_intrinsic_mbcnt_amd:
4085 result = ac_build_mbcnt(&ctx->ac, get_src(ctx, instr->src[0]));
4086 break;
4087 case nir_intrinsic_load_scratch: {
4088 LLVMValueRef offset = get_src(ctx, instr->src[0]);
4089 LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->scratch,
4090 offset);
4091 LLVMTypeRef comp_type =
4092 LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
4093 LLVMTypeRef vec_type =
4094 instr->dest.ssa.num_components == 1 ? comp_type :
4095 LLVMVectorType(comp_type, instr->dest.ssa.num_components);
4096 unsigned addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
4097 ptr = LLVMBuildBitCast(ctx->ac.builder, ptr,
4098 LLVMPointerType(vec_type, addr_space), "");
4099 result = LLVMBuildLoad(ctx->ac.builder, ptr, "");
4100 break;
4101 }
4102 case nir_intrinsic_store_scratch: {
4103 LLVMValueRef offset = get_src(ctx, instr->src[1]);
4104 LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->scratch,
4105 offset);
4106 LLVMTypeRef comp_type =
4107 LLVMIntTypeInContext(ctx->ac.context, instr->src[0].ssa->bit_size);
4108 unsigned addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
4109 ptr = LLVMBuildBitCast(ctx->ac.builder, ptr,
4110 LLVMPointerType(comp_type, addr_space), "");
4111 LLVMValueRef src = get_src(ctx, instr->src[0]);
4112 unsigned wrmask = nir_intrinsic_write_mask(instr);
4113 while (wrmask) {
4114 int start, count;
4115 u_bit_scan_consecutive_range(&wrmask, &start, &count);
4116
4117 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, start, false);
4118 LLVMValueRef offset_ptr = LLVMBuildGEP(ctx->ac.builder, ptr, &offset, 1, "");
4119 LLVMTypeRef vec_type =
4120 count == 1 ? comp_type : LLVMVectorType(comp_type, count);
4121 offset_ptr = LLVMBuildBitCast(ctx->ac.builder,
4122 offset_ptr,
4123 LLVMPointerType(vec_type, addr_space),
4124 "");
4125 LLVMValueRef offset_src =
4126 ac_extract_components(&ctx->ac, src, start, count);
4127 LLVMBuildStore(ctx->ac.builder, offset_src, offset_ptr);
4128 }
4129 break;
4130 }
4131 case nir_intrinsic_load_constant: {
4132 unsigned base = nir_intrinsic_base(instr);
4133 unsigned range = nir_intrinsic_range(instr);
4134
4135 LLVMValueRef offset = get_src(ctx, instr->src[0]);
4136 offset = LLVMBuildAdd(ctx->ac.builder, offset,
4137 LLVMConstInt(ctx->ac.i32, base, false), "");
4138
4139 /* Clamp the offset to avoid out-of-bound access because global
4140 * instructions can't handle them.
4141 */
4142 LLVMValueRef size = LLVMConstInt(ctx->ac.i32, base + range, false);
4143 LLVMValueRef cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT,
4144 offset, size, "");
4145 offset = LLVMBuildSelect(ctx->ac.builder, cond, offset, size, "");
4146
4147 LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->constant_data,
4148 offset);
4149 LLVMTypeRef comp_type =
4150 LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
4151 LLVMTypeRef vec_type =
4152 instr->dest.ssa.num_components == 1 ? comp_type :
4153 LLVMVectorType(comp_type, instr->dest.ssa.num_components);
4154 unsigned addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
4155 ptr = LLVMBuildBitCast(ctx->ac.builder, ptr,
4156 LLVMPointerType(vec_type, addr_space), "");
4157 result = LLVMBuildLoad(ctx->ac.builder, ptr, "");
4158 break;
4159 }
4160 default:
4161 fprintf(stderr, "Unknown intrinsic: ");
4162 nir_print_instr(&instr->instr, stderr);
4163 fprintf(stderr, "\n");
4164 break;
4165 }
4166 if (result) {
4167 ctx->ssa_defs[instr->dest.ssa.index] = result;
4168 }
4169 }
4170
4171 static LLVMValueRef get_bindless_index_from_uniform(struct ac_nir_context *ctx,
4172 unsigned base_index,
4173 unsigned constant_index,
4174 LLVMValueRef dynamic_index)
4175 {
4176 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, base_index * 4, 0);
4177 LLVMValueRef index = LLVMBuildAdd(ctx->ac.builder, dynamic_index,
4178 LLVMConstInt(ctx->ac.i32, constant_index, 0), "");
4179
4180 /* Bindless uniforms are 64bit so multiple index by 8 */
4181 index = LLVMBuildMul(ctx->ac.builder, index, LLVMConstInt(ctx->ac.i32, 8, 0), "");
4182 offset = LLVMBuildAdd(ctx->ac.builder, offset, index, "");
4183
4184 LLVMValueRef ubo_index = ctx->abi->load_ubo(ctx->abi, ctx->ac.i32_0);
4185
4186 LLVMValueRef ret = ac_build_buffer_load(&ctx->ac, ubo_index, 1, NULL, offset,
4187 NULL, 0, 0, true, true);
4188
4189 return LLVMBuildBitCast(ctx->ac.builder, ret, ctx->ac.i32, "");
4190 }
4191
4192 struct sampler_desc_address {
4193 unsigned descriptor_set;
4194 unsigned base_index; /* binding in vulkan */
4195 unsigned constant_index;
4196 LLVMValueRef dynamic_index;
4197 bool image;
4198 bool bindless;
4199 };
4200
4201 static struct sampler_desc_address
4202 get_sampler_desc_internal(struct ac_nir_context *ctx,
4203 nir_deref_instr *deref_instr,
4204 const nir_instr *instr,
4205 bool image)
4206 {
4207 LLVMValueRef index = NULL;
4208 unsigned constant_index = 0;
4209 unsigned descriptor_set;
4210 unsigned base_index;
4211 bool bindless = false;
4212
4213 if (!deref_instr) {
4214 descriptor_set = 0;
4215 if (image) {
4216 nir_intrinsic_instr *img_instr = nir_instr_as_intrinsic(instr);
4217 base_index = 0;
4218 bindless = true;
4219 index = get_src(ctx, img_instr->src[0]);
4220 } else {
4221 nir_tex_instr *tex_instr = nir_instr_as_tex(instr);
4222 int sampSrcIdx = nir_tex_instr_src_index(tex_instr,
4223 nir_tex_src_sampler_handle);
4224 if (sampSrcIdx != -1) {
4225 base_index = 0;
4226 bindless = true;
4227 index = get_src(ctx, tex_instr->src[sampSrcIdx].src);
4228 } else {
4229 assert(tex_instr && !image);
4230 base_index = tex_instr->sampler_index;
4231 }
4232 }
4233 } else {
4234 while(deref_instr->deref_type != nir_deref_type_var) {
4235 if (deref_instr->deref_type == nir_deref_type_array) {
4236 unsigned array_size = glsl_get_aoa_size(deref_instr->type);
4237 if (!array_size)
4238 array_size = 1;
4239
4240 if (nir_src_is_const(deref_instr->arr.index)) {
4241 constant_index += array_size * nir_src_as_uint(deref_instr->arr.index);
4242 } else {
4243 LLVMValueRef indirect = get_src(ctx, deref_instr->arr.index);
4244
4245 indirect = LLVMBuildMul(ctx->ac.builder, indirect,
4246 LLVMConstInt(ctx->ac.i32, array_size, false), "");
4247
4248 if (!index)
4249 index = indirect;
4250 else
4251 index = LLVMBuildAdd(ctx->ac.builder, index, indirect, "");
4252 }
4253
4254 deref_instr = nir_src_as_deref(deref_instr->parent);
4255 } else if (deref_instr->deref_type == nir_deref_type_struct) {
4256 unsigned sidx = deref_instr->strct.index;
4257 deref_instr = nir_src_as_deref(deref_instr->parent);
4258 constant_index += glsl_get_struct_location_offset(deref_instr->type, sidx);
4259 } else {
4260 unreachable("Unsupported deref type");
4261 }
4262 }
4263 descriptor_set = deref_instr->var->data.descriptor_set;
4264
4265 if (deref_instr->var->data.bindless) {
4266 /* For now just assert on unhandled variable types */
4267 assert(deref_instr->var->data.mode == nir_var_uniform);
4268
4269 base_index = deref_instr->var->data.driver_location;
4270 bindless = true;
4271
4272 index = index ? index : ctx->ac.i32_0;
4273 index = get_bindless_index_from_uniform(ctx, base_index,
4274 constant_index, index);
4275 } else
4276 base_index = deref_instr->var->data.binding;
4277 }
4278 return (struct sampler_desc_address) {
4279 .descriptor_set = descriptor_set,
4280 .base_index = base_index,
4281 .constant_index = constant_index,
4282 .dynamic_index = index,
4283 .image = image,
4284 .bindless = bindless,
4285 };
4286 }
4287
4288 /* Extract any possibly divergent index into a separate value that can be fed
4289 * into get_sampler_desc with the same arguments. */
4290 static LLVMValueRef get_sampler_desc_index(struct ac_nir_context *ctx,
4291 nir_deref_instr *deref_instr,
4292 const nir_instr *instr,
4293 bool image)
4294 {
4295 struct sampler_desc_address addr = get_sampler_desc_internal(ctx, deref_instr, instr, image);
4296 return addr.dynamic_index;
4297 }
4298
4299 static LLVMValueRef get_sampler_desc(struct ac_nir_context *ctx,
4300 nir_deref_instr *deref_instr,
4301 enum ac_descriptor_type desc_type,
4302 const nir_instr *instr,
4303 LLVMValueRef index,
4304 bool image, bool write)
4305 {
4306 struct sampler_desc_address addr = get_sampler_desc_internal(ctx, deref_instr, instr, image);
4307 return ctx->abi->load_sampler_desc(ctx->abi,
4308 addr.descriptor_set,
4309 addr.base_index,
4310 addr.constant_index, index,
4311 desc_type, addr.image, write, addr.bindless);
4312 }
4313
4314 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
4315 *
4316 * GFX6-GFX7:
4317 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
4318 * filtering manually. The driver sets img7 to a mask clearing
4319 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
4320 * s_and_b32 samp0, samp0, img7
4321 *
4322 * GFX8:
4323 * The ANISO_OVERRIDE sampler field enables this fix in TA.
4324 */
4325 static LLVMValueRef sici_fix_sampler_aniso(struct ac_nir_context *ctx,
4326 LLVMValueRef res, LLVMValueRef samp)
4327 {
4328 LLVMBuilderRef builder = ctx->ac.builder;
4329 LLVMValueRef img7, samp0;
4330
4331 if (ctx->ac.chip_class >= GFX8)
4332 return samp;
4333
4334 img7 = LLVMBuildExtractElement(builder, res,
4335 LLVMConstInt(ctx->ac.i32, 7, 0), "");
4336 samp0 = LLVMBuildExtractElement(builder, samp,
4337 LLVMConstInt(ctx->ac.i32, 0, 0), "");
4338 samp0 = LLVMBuildAnd(builder, samp0, img7, "");
4339 return LLVMBuildInsertElement(builder, samp, samp0,
4340 LLVMConstInt(ctx->ac.i32, 0, 0), "");
4341 }
4342
4343 static void tex_fetch_ptrs(struct ac_nir_context *ctx,
4344 nir_tex_instr *instr,
4345 struct waterfall_context *wctx,
4346 LLVMValueRef *res_ptr, LLVMValueRef *samp_ptr,
4347 LLVMValueRef *fmask_ptr)
4348 {
4349 nir_deref_instr *texture_deref_instr = NULL;
4350 nir_deref_instr *sampler_deref_instr = NULL;
4351 int plane = -1;
4352
4353 for (unsigned i = 0; i < instr->num_srcs; i++) {
4354 switch (instr->src[i].src_type) {
4355 case nir_tex_src_texture_deref:
4356 texture_deref_instr = nir_src_as_deref(instr->src[i].src);
4357 break;
4358 case nir_tex_src_sampler_deref:
4359 sampler_deref_instr = nir_src_as_deref(instr->src[i].src);
4360 break;
4361 case nir_tex_src_plane:
4362 plane = nir_src_as_int(instr->src[i].src);
4363 break;
4364 default:
4365 break;
4366 }
4367 }
4368
4369 LLVMValueRef texture_dynamic_index = get_sampler_desc_index(ctx, texture_deref_instr,
4370 &instr->instr, false);
4371 if (!sampler_deref_instr)
4372 sampler_deref_instr = texture_deref_instr;
4373
4374 LLVMValueRef sampler_dynamic_index = get_sampler_desc_index(ctx, sampler_deref_instr,
4375 &instr->instr, false);
4376 if (instr->texture_non_uniform)
4377 texture_dynamic_index = enter_waterfall(ctx, wctx + 0, texture_dynamic_index, true);
4378
4379 if (instr->sampler_non_uniform)
4380 sampler_dynamic_index = enter_waterfall(ctx, wctx + 1, sampler_dynamic_index, true);
4381
4382 enum ac_descriptor_type main_descriptor = instr->sampler_dim == GLSL_SAMPLER_DIM_BUF ? AC_DESC_BUFFER : AC_DESC_IMAGE;
4383
4384 if (plane >= 0) {
4385 assert(instr->op != nir_texop_txf_ms &&
4386 instr->op != nir_texop_samples_identical);
4387 assert(instr->sampler_dim != GLSL_SAMPLER_DIM_BUF);
4388
4389 main_descriptor = AC_DESC_PLANE_0 + plane;
4390 }
4391
4392 if (instr->op == nir_texop_fragment_mask_fetch) {
4393 /* The fragment mask is fetched from the compressed
4394 * multisampled surface.
4395 */
4396 main_descriptor = AC_DESC_FMASK;
4397 }
4398
4399 *res_ptr = get_sampler_desc(ctx, texture_deref_instr, main_descriptor, &instr->instr,
4400 texture_dynamic_index, false, false);
4401
4402 if (samp_ptr) {
4403 *samp_ptr = get_sampler_desc(ctx, sampler_deref_instr, AC_DESC_SAMPLER, &instr->instr,
4404 sampler_dynamic_index, false, false);
4405 if (instr->sampler_dim < GLSL_SAMPLER_DIM_RECT)
4406 *samp_ptr = sici_fix_sampler_aniso(ctx, *res_ptr, *samp_ptr);
4407 }
4408 if (fmask_ptr && (instr->op == nir_texop_txf_ms ||
4409 instr->op == nir_texop_samples_identical))
4410 *fmask_ptr = get_sampler_desc(ctx, texture_deref_instr, AC_DESC_FMASK,
4411 &instr->instr, texture_dynamic_index, false, false);
4412 }
4413
4414 static LLVMValueRef apply_round_slice(struct ac_llvm_context *ctx,
4415 LLVMValueRef coord)
4416 {
4417 coord = ac_to_float(ctx, coord);
4418 coord = ac_build_round(ctx, coord);
4419 coord = ac_to_integer(ctx, coord);
4420 return coord;
4421 }
4422
4423 static void visit_tex(struct ac_nir_context *ctx, nir_tex_instr *instr)
4424 {
4425 LLVMValueRef result = NULL;
4426 struct ac_image_args args = { 0 };
4427 LLVMValueRef fmask_ptr = NULL, sample_index = NULL;
4428 LLVMValueRef ddx = NULL, ddy = NULL;
4429 unsigned offset_src = 0;
4430 struct waterfall_context wctx[2] = {{{0}}};
4431
4432 tex_fetch_ptrs(ctx, instr, wctx, &args.resource, &args.sampler, &fmask_ptr);
4433
4434 for (unsigned i = 0; i < instr->num_srcs; i++) {
4435 switch (instr->src[i].src_type) {
4436 case nir_tex_src_coord: {
4437 LLVMValueRef coord = get_src(ctx, instr->src[i].src);
4438 for (unsigned chan = 0; chan < instr->coord_components; ++chan)
4439 args.coords[chan] = ac_llvm_extract_elem(&ctx->ac, coord, chan);
4440 break;
4441 }
4442 case nir_tex_src_projector:
4443 break;
4444 case nir_tex_src_comparator:
4445 if (instr->is_shadow) {
4446 args.compare = get_src(ctx, instr->src[i].src);
4447 args.compare = ac_to_float(&ctx->ac, args.compare);
4448 }
4449 break;
4450 case nir_tex_src_offset:
4451 args.offset = get_src(ctx, instr->src[i].src);
4452 offset_src = i;
4453 break;
4454 case nir_tex_src_bias:
4455 args.bias = get_src(ctx, instr->src[i].src);
4456 break;
4457 case nir_tex_src_lod: {
4458 if (nir_src_is_const(instr->src[i].src) && nir_src_as_uint(instr->src[i].src) == 0)
4459 args.level_zero = true;
4460 else
4461 args.lod = get_src(ctx, instr->src[i].src);
4462 break;
4463 }
4464 case nir_tex_src_ms_index:
4465 sample_index = get_src(ctx, instr->src[i].src);
4466 break;
4467 case nir_tex_src_ms_mcs:
4468 break;
4469 case nir_tex_src_ddx:
4470 ddx = get_src(ctx, instr->src[i].src);
4471 break;
4472 case nir_tex_src_ddy:
4473 ddy = get_src(ctx, instr->src[i].src);
4474 break;
4475 case nir_tex_src_min_lod:
4476 args.min_lod = get_src(ctx, instr->src[i].src);
4477 break;
4478 case nir_tex_src_texture_offset:
4479 case nir_tex_src_sampler_offset:
4480 case nir_tex_src_plane:
4481 default:
4482 break;
4483 }
4484 }
4485
4486 if (instr->op == nir_texop_txs && instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
4487 result = get_buffer_size(ctx, args.resource, true);
4488 goto write_result;
4489 }
4490
4491 if (instr->op == nir_texop_texture_samples) {
4492 LLVMValueRef res, samples, is_msaa;
4493 LLVMValueRef default_sample;
4494
4495 res = LLVMBuildBitCast(ctx->ac.builder, args.resource, ctx->ac.v8i32, "");
4496 samples = LLVMBuildExtractElement(ctx->ac.builder, res,
4497 LLVMConstInt(ctx->ac.i32, 3, false), "");
4498 is_msaa = LLVMBuildLShr(ctx->ac.builder, samples,
4499 LLVMConstInt(ctx->ac.i32, 28, false), "");
4500 is_msaa = LLVMBuildAnd(ctx->ac.builder, is_msaa,
4501 LLVMConstInt(ctx->ac.i32, 0xe, false), "");
4502 is_msaa = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, is_msaa,
4503 LLVMConstInt(ctx->ac.i32, 0xe, false), "");
4504
4505 samples = LLVMBuildLShr(ctx->ac.builder, samples,
4506 LLVMConstInt(ctx->ac.i32, 16, false), "");
4507 samples = LLVMBuildAnd(ctx->ac.builder, samples,
4508 LLVMConstInt(ctx->ac.i32, 0xf, false), "");
4509 samples = LLVMBuildShl(ctx->ac.builder, ctx->ac.i32_1,
4510 samples, "");
4511
4512 if (ctx->abi->robust_buffer_access) {
4513 LLVMValueRef dword1, is_null_descriptor;
4514
4515 /* Extract the second dword of the descriptor, if it's
4516 * all zero, then it's a null descriptor.
4517 */
4518 dword1 = LLVMBuildExtractElement(ctx->ac.builder, res,
4519 LLVMConstInt(ctx->ac.i32, 1, false), "");
4520 is_null_descriptor =
4521 LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, dword1,
4522 LLVMConstInt(ctx->ac.i32, 0, false), "");
4523 default_sample =
4524 LLVMBuildSelect(ctx->ac.builder, is_null_descriptor,
4525 ctx->ac.i32_0, ctx->ac.i32_1, "");
4526 } else {
4527 default_sample = ctx->ac.i32_1;
4528 }
4529
4530 samples = LLVMBuildSelect(ctx->ac.builder, is_msaa, samples,
4531 default_sample, "");
4532 result = samples;
4533 goto write_result;
4534 }
4535
4536 if (args.offset && instr->op != nir_texop_txf && instr->op != nir_texop_txf_ms) {
4537 LLVMValueRef offset[3], pack;
4538 for (unsigned chan = 0; chan < 3; ++chan)
4539 offset[chan] = ctx->ac.i32_0;
4540
4541 unsigned num_components = ac_get_llvm_num_components(args.offset);
4542 for (unsigned chan = 0; chan < num_components; chan++) {
4543 offset[chan] = ac_llvm_extract_elem(&ctx->ac, args.offset, chan);
4544 offset[chan] = LLVMBuildAnd(ctx->ac.builder, offset[chan],
4545 LLVMConstInt(ctx->ac.i32, 0x3f, false), "");
4546 if (chan)
4547 offset[chan] = LLVMBuildShl(ctx->ac.builder, offset[chan],
4548 LLVMConstInt(ctx->ac.i32, chan * 8, false), "");
4549 }
4550 pack = LLVMBuildOr(ctx->ac.builder, offset[0], offset[1], "");
4551 pack = LLVMBuildOr(ctx->ac.builder, pack, offset[2], "");
4552 args.offset = pack;
4553 }
4554
4555 /* Section 8.23.1 (Depth Texture Comparison Mode) of the
4556 * OpenGL 4.5 spec says:
4557 *
4558 * "If the texture’s internal format indicates a fixed-point
4559 * depth texture, then D_t and D_ref are clamped to the
4560 * range [0, 1]; otherwise no clamping is performed."
4561 *
4562 * TC-compatible HTILE promotes Z16 and Z24 to Z32_FLOAT,
4563 * so the depth comparison value isn't clamped for Z16 and
4564 * Z24 anymore. Do it manually here for GFX8-9; GFX10 has
4565 * an explicitly clamped 32-bit float format.
4566 */
4567 if (args.compare &&
4568 ctx->ac.chip_class >= GFX8 &&
4569 ctx->ac.chip_class <= GFX9 &&
4570 ctx->abi->clamp_shadow_reference) {
4571 LLVMValueRef upgraded, clamped;
4572
4573 upgraded = LLVMBuildExtractElement(ctx->ac.builder, args.sampler,
4574 LLVMConstInt(ctx->ac.i32, 3, false), "");
4575 upgraded = LLVMBuildLShr(ctx->ac.builder, upgraded,
4576 LLVMConstInt(ctx->ac.i32, 29, false), "");
4577 upgraded = LLVMBuildTrunc(ctx->ac.builder, upgraded, ctx->ac.i1, "");
4578 clamped = ac_build_clamp(&ctx->ac, args.compare);
4579 args.compare = LLVMBuildSelect(ctx->ac.builder, upgraded, clamped,
4580 args.compare, "");
4581 }
4582
4583 /* pack derivatives */
4584 if (ddx || ddy) {
4585 int num_src_deriv_channels, num_dest_deriv_channels;
4586 switch (instr->sampler_dim) {
4587 case GLSL_SAMPLER_DIM_3D:
4588 case GLSL_SAMPLER_DIM_CUBE:
4589 num_src_deriv_channels = 3;
4590 num_dest_deriv_channels = 3;
4591 break;
4592 case GLSL_SAMPLER_DIM_2D:
4593 default:
4594 num_src_deriv_channels = 2;
4595 num_dest_deriv_channels = 2;
4596 break;
4597 case GLSL_SAMPLER_DIM_1D:
4598 num_src_deriv_channels = 1;
4599 if (ctx->ac.chip_class == GFX9) {
4600 num_dest_deriv_channels = 2;
4601 } else {
4602 num_dest_deriv_channels = 1;
4603 }
4604 break;
4605 }
4606
4607 for (unsigned i = 0; i < num_src_deriv_channels; i++) {
4608 args.derivs[i] = ac_to_float(&ctx->ac,
4609 ac_llvm_extract_elem(&ctx->ac, ddx, i));
4610 args.derivs[num_dest_deriv_channels + i] = ac_to_float(&ctx->ac,
4611 ac_llvm_extract_elem(&ctx->ac, ddy, i));
4612 }
4613 for (unsigned i = num_src_deriv_channels; i < num_dest_deriv_channels; i++) {
4614 args.derivs[i] = ctx->ac.f32_0;
4615 args.derivs[num_dest_deriv_channels + i] = ctx->ac.f32_0;
4616 }
4617 }
4618
4619 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE && args.coords[0]) {
4620 for (unsigned chan = 0; chan < instr->coord_components; chan++)
4621 args.coords[chan] = ac_to_float(&ctx->ac, args.coords[chan]);
4622 if (instr->coord_components == 3)
4623 args.coords[3] = LLVMGetUndef(ctx->ac.f32);
4624 ac_prepare_cube_coords(&ctx->ac,
4625 instr->op == nir_texop_txd, instr->is_array,
4626 instr->op == nir_texop_lod, args.coords, args.derivs);
4627 }
4628
4629 /* Texture coordinates fixups */
4630 if (instr->coord_components > 1 &&
4631 instr->sampler_dim == GLSL_SAMPLER_DIM_1D &&
4632 instr->is_array &&
4633 instr->op != nir_texop_txf) {
4634 args.coords[1] = apply_round_slice(&ctx->ac, args.coords[1]);
4635 }
4636
4637 if (instr->coord_components > 2 &&
4638 (instr->sampler_dim == GLSL_SAMPLER_DIM_2D ||
4639 instr->sampler_dim == GLSL_SAMPLER_DIM_MS ||
4640 instr->sampler_dim == GLSL_SAMPLER_DIM_SUBPASS ||
4641 instr->sampler_dim == GLSL_SAMPLER_DIM_SUBPASS_MS) &&
4642 instr->is_array &&
4643 instr->op != nir_texop_txf &&
4644 instr->op != nir_texop_txf_ms &&
4645 instr->op != nir_texop_fragment_fetch &&
4646 instr->op != nir_texop_fragment_mask_fetch) {
4647 args.coords[2] = apply_round_slice(&ctx->ac, args.coords[2]);
4648 }
4649
4650 if (ctx->ac.chip_class == GFX9 &&
4651 instr->sampler_dim == GLSL_SAMPLER_DIM_1D &&
4652 instr->op != nir_texop_lod) {
4653 LLVMValueRef filler;
4654 if (instr->op == nir_texop_txf)
4655 filler = ctx->ac.i32_0;
4656 else
4657 filler = LLVMConstReal(ctx->ac.f32, 0.5);
4658
4659 if (instr->is_array)
4660 args.coords[2] = args.coords[1];
4661 args.coords[1] = filler;
4662 }
4663
4664 /* Pack sample index */
4665 if (sample_index && (instr->op == nir_texop_txf_ms ||
4666 instr->op == nir_texop_fragment_fetch))
4667 args.coords[instr->coord_components] = sample_index;
4668
4669 if (instr->op == nir_texop_samples_identical) {
4670 struct ac_image_args txf_args = { 0 };
4671 memcpy(txf_args.coords, args.coords, sizeof(txf_args.coords));
4672
4673 txf_args.dmask = 0xf;
4674 txf_args.resource = fmask_ptr;
4675 txf_args.dim = instr->is_array ? ac_image_2darray : ac_image_2d;
4676 result = build_tex_intrinsic(ctx, instr, &txf_args);
4677
4678 result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
4679 result = emit_int_cmp(&ctx->ac, LLVMIntEQ, result, ctx->ac.i32_0);
4680 goto write_result;
4681 }
4682
4683 if ((instr->sampler_dim == GLSL_SAMPLER_DIM_SUBPASS_MS ||
4684 instr->sampler_dim == GLSL_SAMPLER_DIM_MS) &&
4685 instr->op != nir_texop_txs &&
4686 instr->op != nir_texop_fragment_fetch &&
4687 instr->op != nir_texop_fragment_mask_fetch) {
4688 unsigned sample_chan = instr->is_array ? 3 : 2;
4689 args.coords[sample_chan] = adjust_sample_index_using_fmask(
4690 &ctx->ac, args.coords[0], args.coords[1],
4691 instr->is_array ? args.coords[2] : NULL,
4692 args.coords[sample_chan], fmask_ptr);
4693 }
4694
4695 if (args.offset && (instr->op == nir_texop_txf || instr->op == nir_texop_txf_ms)) {
4696 int num_offsets = instr->src[offset_src].src.ssa->num_components;
4697 num_offsets = MIN2(num_offsets, instr->coord_components);
4698 for (unsigned i = 0; i < num_offsets; ++i) {
4699 args.coords[i] = LLVMBuildAdd(
4700 ctx->ac.builder, args.coords[i],
4701 LLVMConstInt(ctx->ac.i32, nir_src_comp_as_uint(instr->src[offset_src].src, i), false), "");
4702 }
4703 args.offset = NULL;
4704 }
4705
4706 /* DMASK was repurposed for GATHER4. 4 components are always
4707 * returned and DMASK works like a swizzle - it selects
4708 * the component to fetch. The only valid DMASK values are
4709 * 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
4710 * (red,red,red,red) etc.) The ISA document doesn't mention
4711 * this.
4712 */
4713 args.dmask = 0xf;
4714 if (instr->op == nir_texop_tg4) {
4715 if (instr->is_shadow)
4716 args.dmask = 1;
4717 else
4718 args.dmask = 1 << instr->component;
4719 }
4720
4721 if (instr->sampler_dim != GLSL_SAMPLER_DIM_BUF) {
4722 args.dim = ac_get_sampler_dim(ctx->ac.chip_class, instr->sampler_dim, instr->is_array);
4723 args.unorm = instr->sampler_dim == GLSL_SAMPLER_DIM_RECT;
4724 }
4725
4726 /* Adjust the number of coordinates because we only need (x,y) for 2D
4727 * multisampled images and (x,y,layer) for 2D multisampled layered
4728 * images or for multisampled input attachments.
4729 */
4730 if (instr->op == nir_texop_fragment_mask_fetch) {
4731 if (args.dim == ac_image_2dmsaa) {
4732 args.dim = ac_image_2d;
4733 } else {
4734 assert(args.dim == ac_image_2darraymsaa);
4735 args.dim = ac_image_2darray;
4736 }
4737 }
4738
4739 result = build_tex_intrinsic(ctx, instr, &args);
4740
4741 if (instr->op == nir_texop_query_levels)
4742 result = LLVMBuildExtractElement(ctx->ac.builder, result, LLVMConstInt(ctx->ac.i32, 3, false), "");
4743 else if (instr->is_shadow && instr->is_new_style_shadow &&
4744 instr->op != nir_texop_txs && instr->op != nir_texop_lod &&
4745 instr->op != nir_texop_tg4)
4746 result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
4747 else if (instr->op == nir_texop_txs &&
4748 instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE &&
4749 instr->is_array) {
4750 LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
4751 LLVMValueRef six = LLVMConstInt(ctx->ac.i32, 6, false);
4752 LLVMValueRef z = LLVMBuildExtractElement(ctx->ac.builder, result, two, "");
4753 z = LLVMBuildSDiv(ctx->ac.builder, z, six, "");
4754 result = LLVMBuildInsertElement(ctx->ac.builder, result, z, two, "");
4755 } else if (ctx->ac.chip_class == GFX9 &&
4756 instr->op == nir_texop_txs &&
4757 instr->sampler_dim == GLSL_SAMPLER_DIM_1D &&
4758 instr->is_array) {
4759 LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
4760 LLVMValueRef layers = LLVMBuildExtractElement(ctx->ac.builder, result, two, "");
4761 result = LLVMBuildInsertElement(ctx->ac.builder, result, layers,
4762 ctx->ac.i32_1, "");
4763 } else if (instr->dest.ssa.num_components != 4)
4764 result = ac_trim_vector(&ctx->ac, result, instr->dest.ssa.num_components);
4765
4766 write_result:
4767 if (result) {
4768 assert(instr->dest.is_ssa);
4769 result = ac_to_integer(&ctx->ac, result);
4770
4771 for (int i = ARRAY_SIZE(wctx); --i >= 0;) {
4772 result = exit_waterfall(ctx, wctx + i, result);
4773 }
4774
4775 ctx->ssa_defs[instr->dest.ssa.index] = result;
4776 }
4777 }
4778
4779 static void visit_phi(struct ac_nir_context *ctx, nir_phi_instr *instr)
4780 {
4781 LLVMTypeRef type = get_def_type(ctx, &instr->dest.ssa);
4782 LLVMValueRef result = LLVMBuildPhi(ctx->ac.builder, type, "");
4783
4784 ctx->ssa_defs[instr->dest.ssa.index] = result;
4785 _mesa_hash_table_insert(ctx->phis, instr, result);
4786 }
4787
4788 static void visit_post_phi(struct ac_nir_context *ctx,
4789 nir_phi_instr *instr,
4790 LLVMValueRef llvm_phi)
4791 {
4792 nir_foreach_phi_src(src, instr) {
4793 LLVMBasicBlockRef block = get_block(ctx, src->pred);
4794 LLVMValueRef llvm_src = get_src(ctx, src->src);
4795
4796 LLVMAddIncoming(llvm_phi, &llvm_src, &block, 1);
4797 }
4798 }
4799
4800 static void phi_post_pass(struct ac_nir_context *ctx)
4801 {
4802 hash_table_foreach(ctx->phis, entry) {
4803 visit_post_phi(ctx, (nir_phi_instr*)entry->key,
4804 (LLVMValueRef)entry->data);
4805 }
4806 }
4807
4808
4809 static bool is_def_used_in_an_export(const nir_ssa_def* def) {
4810 nir_foreach_use(use_src, def) {
4811 if (use_src->parent_instr->type == nir_instr_type_intrinsic) {
4812 nir_intrinsic_instr *instr = nir_instr_as_intrinsic(use_src->parent_instr);
4813 if (instr->intrinsic == nir_intrinsic_store_deref)
4814 return true;
4815 } else if (use_src->parent_instr->type == nir_instr_type_alu) {
4816 nir_alu_instr *instr = nir_instr_as_alu(use_src->parent_instr);
4817 if (instr->op == nir_op_vec4 &&
4818 is_def_used_in_an_export(&instr->dest.dest.ssa)) {
4819 return true;
4820 }
4821 }
4822 }
4823 return false;
4824 }
4825
4826 static void visit_ssa_undef(struct ac_nir_context *ctx,
4827 const nir_ssa_undef_instr *instr)
4828 {
4829 unsigned num_components = instr->def.num_components;
4830 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
4831
4832 if (!ctx->abi->convert_undef_to_zero || is_def_used_in_an_export(&instr->def)) {
4833 LLVMValueRef undef;
4834
4835 if (num_components == 1)
4836 undef = LLVMGetUndef(type);
4837 else {
4838 undef = LLVMGetUndef(LLVMVectorType(type, num_components));
4839 }
4840 ctx->ssa_defs[instr->def.index] = undef;
4841 } else {
4842 LLVMValueRef zero = LLVMConstInt(type, 0, false);
4843 if (num_components > 1) {
4844 zero = ac_build_gather_values_extended(
4845 &ctx->ac, &zero, 4, 0, false, false);
4846 }
4847 ctx->ssa_defs[instr->def.index] = zero;
4848 }
4849 }
4850
4851 static void visit_jump(struct ac_llvm_context *ctx,
4852 const nir_jump_instr *instr)
4853 {
4854 switch (instr->type) {
4855 case nir_jump_break:
4856 ac_build_break(ctx);
4857 break;
4858 case nir_jump_continue:
4859 ac_build_continue(ctx);
4860 break;
4861 default:
4862 fprintf(stderr, "Unknown NIR jump instr: ");
4863 nir_print_instr(&instr->instr, stderr);
4864 fprintf(stderr, "\n");
4865 abort();
4866 }
4867 }
4868
4869 static LLVMTypeRef
4870 glsl_base_to_llvm_type(struct ac_llvm_context *ac,
4871 enum glsl_base_type type)
4872 {
4873 switch (type) {
4874 case GLSL_TYPE_INT:
4875 case GLSL_TYPE_UINT:
4876 case GLSL_TYPE_BOOL:
4877 case GLSL_TYPE_SUBROUTINE:
4878 return ac->i32;
4879 case GLSL_TYPE_INT8:
4880 case GLSL_TYPE_UINT8:
4881 return ac->i8;
4882 case GLSL_TYPE_INT16:
4883 case GLSL_TYPE_UINT16:
4884 return ac->i16;
4885 case GLSL_TYPE_FLOAT:
4886 return ac->f32;
4887 case GLSL_TYPE_FLOAT16:
4888 return ac->f16;
4889 case GLSL_TYPE_INT64:
4890 case GLSL_TYPE_UINT64:
4891 return ac->i64;
4892 case GLSL_TYPE_DOUBLE:
4893 return ac->f64;
4894 default:
4895 unreachable("unknown GLSL type");
4896 }
4897 }
4898
4899 static LLVMTypeRef
4900 glsl_to_llvm_type(struct ac_llvm_context *ac,
4901 const struct glsl_type *type)
4902 {
4903 if (glsl_type_is_scalar(type)) {
4904 return glsl_base_to_llvm_type(ac, glsl_get_base_type(type));
4905 }
4906
4907 if (glsl_type_is_vector(type)) {
4908 return LLVMVectorType(
4909 glsl_base_to_llvm_type(ac, glsl_get_base_type(type)),
4910 glsl_get_vector_elements(type));
4911 }
4912
4913 if (glsl_type_is_matrix(type)) {
4914 return LLVMArrayType(
4915 glsl_to_llvm_type(ac, glsl_get_column_type(type)),
4916 glsl_get_matrix_columns(type));
4917 }
4918
4919 if (glsl_type_is_array(type)) {
4920 return LLVMArrayType(
4921 glsl_to_llvm_type(ac, glsl_get_array_element(type)),
4922 glsl_get_length(type));
4923 }
4924
4925 assert(glsl_type_is_struct_or_ifc(type));
4926
4927 LLVMTypeRef member_types[glsl_get_length(type)];
4928
4929 for (unsigned i = 0; i < glsl_get_length(type); i++) {
4930 member_types[i] =
4931 glsl_to_llvm_type(ac,
4932 glsl_get_struct_field(type, i));
4933 }
4934
4935 return LLVMStructTypeInContext(ac->context, member_types,
4936 glsl_get_length(type), false);
4937 }
4938
4939 static void visit_deref(struct ac_nir_context *ctx,
4940 nir_deref_instr *instr)
4941 {
4942 if (instr->mode != nir_var_mem_shared &&
4943 instr->mode != nir_var_mem_global)
4944 return;
4945
4946 LLVMValueRef result = NULL;
4947 switch(instr->deref_type) {
4948 case nir_deref_type_var: {
4949 struct hash_entry *entry = _mesa_hash_table_search(ctx->vars, instr->var);
4950 result = entry->data;
4951 break;
4952 }
4953 case nir_deref_type_struct:
4954 if (instr->mode == nir_var_mem_global) {
4955 nir_deref_instr *parent = nir_deref_instr_parent(instr);
4956 uint64_t offset = glsl_get_struct_field_offset(parent->type,
4957 instr->strct.index);
4958 result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent),
4959 LLVMConstInt(ctx->ac.i32, offset, 0));
4960 } else {
4961 result = ac_build_gep0(&ctx->ac, get_src(ctx, instr->parent),
4962 LLVMConstInt(ctx->ac.i32, instr->strct.index, 0));
4963 }
4964 break;
4965 case nir_deref_type_array:
4966 if (instr->mode == nir_var_mem_global) {
4967 nir_deref_instr *parent = nir_deref_instr_parent(instr);
4968 unsigned stride = glsl_get_explicit_stride(parent->type);
4969
4970 if ((glsl_type_is_matrix(parent->type) &&
4971 glsl_matrix_type_is_row_major(parent->type)) ||
4972 (glsl_type_is_vector(parent->type) && stride == 0))
4973 stride = type_scalar_size_bytes(parent->type);
4974
4975 assert(stride > 0);
4976 LLVMValueRef index = get_src(ctx, instr->arr.index);
4977 if (LLVMTypeOf(index) != ctx->ac.i64)
4978 index = LLVMBuildZExt(ctx->ac.builder, index, ctx->ac.i64, "");
4979
4980 LLVMValueRef offset = LLVMBuildMul(ctx->ac.builder, index, LLVMConstInt(ctx->ac.i64, stride, 0), "");
4981
4982 result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent), offset);
4983 } else {
4984 result = ac_build_gep0(&ctx->ac, get_src(ctx, instr->parent),
4985 get_src(ctx, instr->arr.index));
4986 }
4987 break;
4988 case nir_deref_type_ptr_as_array:
4989 if (instr->mode == nir_var_mem_global) {
4990 unsigned stride = nir_deref_instr_ptr_as_array_stride(instr);
4991
4992 LLVMValueRef index = get_src(ctx, instr->arr.index);
4993 if (LLVMTypeOf(index) != ctx->ac.i64)
4994 index = LLVMBuildZExt(ctx->ac.builder, index, ctx->ac.i64, "");
4995
4996 LLVMValueRef offset = LLVMBuildMul(ctx->ac.builder, index, LLVMConstInt(ctx->ac.i64, stride, 0), "");
4997
4998 result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent), offset);
4999 } else {
5000 result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent),
5001 get_src(ctx, instr->arr.index));
5002 }
5003 break;
5004 case nir_deref_type_cast: {
5005 result = get_src(ctx, instr->parent);
5006
5007 /* We can't use the structs from LLVM because the shader
5008 * specifies its own offsets. */
5009 LLVMTypeRef pointee_type = ctx->ac.i8;
5010 if (instr->mode == nir_var_mem_shared)
5011 pointee_type = glsl_to_llvm_type(&ctx->ac, instr->type);
5012
5013 unsigned address_space;
5014
5015 switch(instr->mode) {
5016 case nir_var_mem_shared:
5017 address_space = AC_ADDR_SPACE_LDS;
5018 break;
5019 case nir_var_mem_global:
5020 address_space = AC_ADDR_SPACE_GLOBAL;
5021 break;
5022 default:
5023 unreachable("Unhandled address space");
5024 }
5025
5026 LLVMTypeRef type = LLVMPointerType(pointee_type, address_space);
5027
5028 if (LLVMTypeOf(result) != type) {
5029 if (LLVMGetTypeKind(LLVMTypeOf(result)) == LLVMVectorTypeKind) {
5030 result = LLVMBuildBitCast(ctx->ac.builder, result,
5031 type, "");
5032 } else {
5033 result = LLVMBuildIntToPtr(ctx->ac.builder, result,
5034 type, "");
5035 }
5036 }
5037 break;
5038 }
5039 default:
5040 unreachable("Unhandled deref_instr deref type");
5041 }
5042
5043 ctx->ssa_defs[instr->dest.ssa.index] = result;
5044 }
5045
5046 static void visit_cf_list(struct ac_nir_context *ctx,
5047 struct exec_list *list);
5048
5049 static void visit_block(struct ac_nir_context *ctx, nir_block *block)
5050 {
5051 nir_foreach_instr(instr, block)
5052 {
5053 switch (instr->type) {
5054 case nir_instr_type_alu:
5055 visit_alu(ctx, nir_instr_as_alu(instr));
5056 break;
5057 case nir_instr_type_load_const:
5058 visit_load_const(ctx, nir_instr_as_load_const(instr));
5059 break;
5060 case nir_instr_type_intrinsic:
5061 visit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
5062 break;
5063 case nir_instr_type_tex:
5064 visit_tex(ctx, nir_instr_as_tex(instr));
5065 break;
5066 case nir_instr_type_phi:
5067 visit_phi(ctx, nir_instr_as_phi(instr));
5068 break;
5069 case nir_instr_type_ssa_undef:
5070 visit_ssa_undef(ctx, nir_instr_as_ssa_undef(instr));
5071 break;
5072 case nir_instr_type_jump:
5073 visit_jump(&ctx->ac, nir_instr_as_jump(instr));
5074 break;
5075 case nir_instr_type_deref:
5076 visit_deref(ctx, nir_instr_as_deref(instr));
5077 break;
5078 default:
5079 fprintf(stderr, "Unknown NIR instr type: ");
5080 nir_print_instr(instr, stderr);
5081 fprintf(stderr, "\n");
5082 abort();
5083 }
5084 }
5085
5086 _mesa_hash_table_insert(ctx->defs, block,
5087 LLVMGetInsertBlock(ctx->ac.builder));
5088 }
5089
5090 static void visit_if(struct ac_nir_context *ctx, nir_if *if_stmt)
5091 {
5092 LLVMValueRef value = get_src(ctx, if_stmt->condition);
5093
5094 nir_block *then_block =
5095 (nir_block *) exec_list_get_head(&if_stmt->then_list);
5096
5097 ac_build_uif(&ctx->ac, value, then_block->index);
5098
5099 visit_cf_list(ctx, &if_stmt->then_list);
5100
5101 if (!exec_list_is_empty(&if_stmt->else_list)) {
5102 nir_block *else_block =
5103 (nir_block *) exec_list_get_head(&if_stmt->else_list);
5104
5105 ac_build_else(&ctx->ac, else_block->index);
5106 visit_cf_list(ctx, &if_stmt->else_list);
5107 }
5108
5109 ac_build_endif(&ctx->ac, then_block->index);
5110 }
5111
5112 static void visit_loop(struct ac_nir_context *ctx, nir_loop *loop)
5113 {
5114 nir_block *first_loop_block =
5115 (nir_block *) exec_list_get_head(&loop->body);
5116
5117 ac_build_bgnloop(&ctx->ac, first_loop_block->index);
5118
5119 visit_cf_list(ctx, &loop->body);
5120
5121 ac_build_endloop(&ctx->ac, first_loop_block->index);
5122 }
5123
5124 static void visit_cf_list(struct ac_nir_context *ctx,
5125 struct exec_list *list)
5126 {
5127 foreach_list_typed(nir_cf_node, node, node, list)
5128 {
5129 switch (node->type) {
5130 case nir_cf_node_block:
5131 visit_block(ctx, nir_cf_node_as_block(node));
5132 break;
5133
5134 case nir_cf_node_if:
5135 visit_if(ctx, nir_cf_node_as_if(node));
5136 break;
5137
5138 case nir_cf_node_loop:
5139 visit_loop(ctx, nir_cf_node_as_loop(node));
5140 break;
5141
5142 default:
5143 assert(0);
5144 }
5145 }
5146 }
5147
5148 void
5149 ac_handle_shader_output_decl(struct ac_llvm_context *ctx,
5150 struct ac_shader_abi *abi,
5151 struct nir_shader *nir,
5152 struct nir_variable *variable,
5153 gl_shader_stage stage)
5154 {
5155 unsigned output_loc = variable->data.driver_location / 4;
5156 unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
5157
5158 /* tess ctrl has it's own load/store paths for outputs */
5159 if (stage == MESA_SHADER_TESS_CTRL)
5160 return;
5161
5162 if (stage == MESA_SHADER_VERTEX ||
5163 stage == MESA_SHADER_TESS_EVAL ||
5164 stage == MESA_SHADER_GEOMETRY) {
5165 int idx = variable->data.location + variable->data.index;
5166 if (idx == VARYING_SLOT_CLIP_DIST0) {
5167 int length = nir->info.clip_distance_array_size +
5168 nir->info.cull_distance_array_size;
5169
5170 if (length > 4)
5171 attrib_count = 2;
5172 else
5173 attrib_count = 1;
5174 }
5175 }
5176
5177 bool is_16bit = glsl_type_is_16bit(glsl_without_array(variable->type));
5178 LLVMTypeRef type = is_16bit ? ctx->f16 : ctx->f32;
5179 for (unsigned i = 0; i < attrib_count; ++i) {
5180 for (unsigned chan = 0; chan < 4; chan++) {
5181 abi->outputs[ac_llvm_reg_index_soa(output_loc + i, chan)] =
5182 ac_build_alloca_undef(ctx, type, "");
5183 }
5184 }
5185 }
5186
5187 static void
5188 setup_locals(struct ac_nir_context *ctx,
5189 struct nir_function *func)
5190 {
5191 int i, j;
5192 ctx->num_locals = 0;
5193 nir_foreach_variable(variable, &func->impl->locals) {
5194 unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
5195 variable->data.driver_location = ctx->num_locals * 4;
5196 variable->data.location_frac = 0;
5197 ctx->num_locals += attrib_count;
5198 }
5199 ctx->locals = malloc(4 * ctx->num_locals * sizeof(LLVMValueRef));
5200 if (!ctx->locals)
5201 return;
5202
5203 for (i = 0; i < ctx->num_locals; i++) {
5204 for (j = 0; j < 4; j++) {
5205 ctx->locals[i * 4 + j] =
5206 ac_build_alloca_undef(&ctx->ac, ctx->ac.f32, "temp");
5207 }
5208 }
5209 }
5210
5211 static void
5212 setup_scratch(struct ac_nir_context *ctx,
5213 struct nir_shader *shader)
5214 {
5215 if (shader->scratch_size == 0)
5216 return;
5217
5218 ctx->scratch = ac_build_alloca_undef(&ctx->ac,
5219 LLVMArrayType(ctx->ac.i8, shader->scratch_size),
5220 "scratch");
5221 }
5222
5223 static void
5224 setup_constant_data(struct ac_nir_context *ctx,
5225 struct nir_shader *shader)
5226 {
5227 if (!shader->constant_data)
5228 return;
5229
5230 LLVMValueRef data =
5231 LLVMConstStringInContext(ctx->ac.context,
5232 shader->constant_data,
5233 shader->constant_data_size,
5234 true);
5235 LLVMTypeRef type = LLVMArrayType(ctx->ac.i8, shader->constant_data_size);
5236
5237 /* We want to put the constant data in the CONST address space so that
5238 * we can use scalar loads. However, LLVM versions before 10 put these
5239 * variables in the same section as the code, which is unacceptable
5240 * for RadeonSI as it needs to relocate all the data sections after
5241 * the code sections. See https://reviews.llvm.org/D65813.
5242 */
5243 unsigned address_space =
5244 LLVM_VERSION_MAJOR < 10 ? AC_ADDR_SPACE_GLOBAL : AC_ADDR_SPACE_CONST;
5245
5246 LLVMValueRef global =
5247 LLVMAddGlobalInAddressSpace(ctx->ac.module, type,
5248 "const_data",
5249 address_space);
5250
5251 LLVMSetInitializer(global, data);
5252 LLVMSetGlobalConstant(global, true);
5253 LLVMSetVisibility(global, LLVMHiddenVisibility);
5254 ctx->constant_data = global;
5255 }
5256
5257 static void
5258 setup_shared(struct ac_nir_context *ctx,
5259 struct nir_shader *nir)
5260 {
5261 if (ctx->ac.lds)
5262 return;
5263
5264 LLVMTypeRef type = LLVMArrayType(ctx->ac.i8,
5265 nir->info.cs.shared_size);
5266
5267 LLVMValueRef lds =
5268 LLVMAddGlobalInAddressSpace(ctx->ac.module, type,
5269 "compute_lds",
5270 AC_ADDR_SPACE_LDS);
5271 LLVMSetAlignment(lds, 64 * 1024);
5272
5273 ctx->ac.lds = LLVMBuildBitCast(ctx->ac.builder, lds,
5274 LLVMPointerType(ctx->ac.i8,
5275 AC_ADDR_SPACE_LDS), "");
5276 }
5277
5278 void ac_nir_translate(struct ac_llvm_context *ac, struct ac_shader_abi *abi,
5279 const struct ac_shader_args *args, struct nir_shader *nir)
5280 {
5281 struct ac_nir_context ctx = {};
5282 struct nir_function *func;
5283
5284 ctx.ac = *ac;
5285 ctx.abi = abi;
5286 ctx.args = args;
5287
5288 ctx.stage = nir->info.stage;
5289 ctx.info = &nir->info;
5290
5291 ctx.main_function = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx.ac.builder));
5292
5293 nir_foreach_variable(variable, &nir->outputs)
5294 ac_handle_shader_output_decl(&ctx.ac, ctx.abi, nir, variable,
5295 ctx.stage);
5296
5297 ctx.defs = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
5298 _mesa_key_pointer_equal);
5299 ctx.phis = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
5300 _mesa_key_pointer_equal);
5301 ctx.vars = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
5302 _mesa_key_pointer_equal);
5303
5304 if (ctx.abi->kill_ps_if_inf_interp)
5305 ctx.verified_interp = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
5306 _mesa_key_pointer_equal);
5307
5308 func = (struct nir_function *)exec_list_get_head(&nir->functions);
5309
5310 nir_index_ssa_defs(func->impl);
5311 ctx.ssa_defs = calloc(func->impl->ssa_alloc, sizeof(LLVMValueRef));
5312
5313 setup_locals(&ctx, func);
5314 setup_scratch(&ctx, nir);
5315 setup_constant_data(&ctx, nir);
5316
5317 if (gl_shader_stage_is_compute(nir->info.stage))
5318 setup_shared(&ctx, nir);
5319
5320 if (nir->info.stage == MESA_SHADER_FRAGMENT && nir->info.fs.uses_demote) {
5321 ctx.ac.postponed_kill = ac_build_alloca_undef(&ctx.ac, ac->i1, "");
5322 /* true = don't kill. */
5323 LLVMBuildStore(ctx.ac.builder, ctx.ac.i1true, ctx.ac.postponed_kill);
5324 }
5325
5326 visit_cf_list(&ctx, &func->impl->body);
5327 phi_post_pass(&ctx);
5328
5329 if (ctx.ac.postponed_kill)
5330 ac_build_kill_if_false(&ctx.ac, LLVMBuildLoad(ctx.ac.builder,
5331 ctx.ac.postponed_kill, ""));
5332
5333 if (!gl_shader_stage_is_compute(nir->info.stage))
5334 ctx.abi->emit_outputs(ctx.abi, AC_LLVM_MAX_OUTPUTS,
5335 ctx.abi->outputs);
5336
5337 free(ctx.locals);
5338 free(ctx.ssa_defs);
5339 ralloc_free(ctx.defs);
5340 ralloc_free(ctx.phis);
5341 ralloc_free(ctx.vars);
5342 if (ctx.abi->kill_ps_if_inf_interp)
5343 ralloc_free(ctx.verified_interp);
5344 }
5345
5346 bool
5347 ac_lower_indirect_derefs(struct nir_shader *nir, enum chip_class chip_class)
5348 {
5349 bool progress = false;
5350
5351 /* Lower large variables to scratch first so that we won't bloat the
5352 * shader by generating large if ladders for them. We later lower
5353 * scratch to alloca's, assuming LLVM won't generate VGPR indexing.
5354 */
5355 NIR_PASS(progress, nir, nir_lower_vars_to_scratch,
5356 nir_var_function_temp,
5357 256,
5358 glsl_get_natural_size_align_bytes);
5359
5360 /* While it would be nice not to have this flag, we are constrained
5361 * by the reality that LLVM 9.0 has buggy VGPR indexing on GFX9.
5362 */
5363 bool llvm_has_working_vgpr_indexing = chip_class != GFX9;
5364
5365 /* TODO: Indirect indexing of GS inputs is unimplemented.
5366 *
5367 * TCS and TES load inputs directly from LDS or offchip memory, so
5368 * indirect indexing is trivial.
5369 */
5370 nir_variable_mode indirect_mask = 0;
5371 if (nir->info.stage == MESA_SHADER_GEOMETRY ||
5372 (nir->info.stage != MESA_SHADER_TESS_CTRL &&
5373 nir->info.stage != MESA_SHADER_TESS_EVAL &&
5374 !llvm_has_working_vgpr_indexing)) {
5375 indirect_mask |= nir_var_shader_in;
5376 }
5377 if (!llvm_has_working_vgpr_indexing &&
5378 nir->info.stage != MESA_SHADER_TESS_CTRL)
5379 indirect_mask |= nir_var_shader_out;
5380
5381 /* TODO: We shouldn't need to do this, however LLVM isn't currently
5382 * smart enough to handle indirects without causing excess spilling
5383 * causing the gpu to hang.
5384 *
5385 * See the following thread for more details of the problem:
5386 * https://lists.freedesktop.org/archives/mesa-dev/2017-July/162106.html
5387 */
5388 indirect_mask |= nir_var_function_temp;
5389
5390 progress |= nir_lower_indirect_derefs(nir, indirect_mask);
5391 return progress;
5392 }
5393
5394 static unsigned
5395 get_inst_tessfactor_writemask(nir_intrinsic_instr *intrin)
5396 {
5397 if (intrin->intrinsic != nir_intrinsic_store_deref)
5398 return 0;
5399
5400 nir_variable *var =
5401 nir_deref_instr_get_variable(nir_src_as_deref(intrin->src[0]));
5402
5403 if (var->data.mode != nir_var_shader_out)
5404 return 0;
5405
5406 unsigned writemask = 0;
5407 const int location = var->data.location;
5408 unsigned first_component = var->data.location_frac;
5409 unsigned num_comps = intrin->dest.ssa.num_components;
5410
5411 if (location == VARYING_SLOT_TESS_LEVEL_INNER)
5412 writemask = ((1 << (num_comps + 1)) - 1) << first_component;
5413 else if (location == VARYING_SLOT_TESS_LEVEL_OUTER)
5414 writemask = (((1 << (num_comps + 1)) - 1) << first_component) << 4;
5415
5416 return writemask;
5417 }
5418
5419 static void
5420 scan_tess_ctrl(nir_cf_node *cf_node, unsigned *upper_block_tf_writemask,
5421 unsigned *cond_block_tf_writemask,
5422 bool *tessfactors_are_def_in_all_invocs, bool is_nested_cf)
5423 {
5424 switch (cf_node->type) {
5425 case nir_cf_node_block: {
5426 nir_block *block = nir_cf_node_as_block(cf_node);
5427 nir_foreach_instr(instr, block) {
5428 if (instr->type != nir_instr_type_intrinsic)
5429 continue;
5430
5431 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
5432 if (intrin->intrinsic == nir_intrinsic_control_barrier) {
5433
5434 /* If we find a barrier in nested control flow put this in the
5435 * too hard basket. In GLSL this is not possible but it is in
5436 * SPIR-V.
5437 */
5438 if (is_nested_cf) {
5439 *tessfactors_are_def_in_all_invocs = false;
5440 return;
5441 }
5442
5443 /* The following case must be prevented:
5444 * gl_TessLevelInner = ...;
5445 * barrier();
5446 * if (gl_InvocationID == 1)
5447 * gl_TessLevelInner = ...;
5448 *
5449 * If you consider disjoint code segments separated by barriers, each
5450 * such segment that writes tess factor channels should write the same
5451 * channels in all codepaths within that segment.
5452 */
5453 if (upper_block_tf_writemask || cond_block_tf_writemask) {
5454 /* Accumulate the result: */
5455 *tessfactors_are_def_in_all_invocs &=
5456 !(*cond_block_tf_writemask & ~(*upper_block_tf_writemask));
5457
5458 /* Analyze the next code segment from scratch. */
5459 *upper_block_tf_writemask = 0;
5460 *cond_block_tf_writemask = 0;
5461 }
5462 } else
5463 *upper_block_tf_writemask |= get_inst_tessfactor_writemask(intrin);
5464 }
5465
5466 break;
5467 }
5468 case nir_cf_node_if: {
5469 unsigned then_tessfactor_writemask = 0;
5470 unsigned else_tessfactor_writemask = 0;
5471
5472 nir_if *if_stmt = nir_cf_node_as_if(cf_node);
5473 foreach_list_typed(nir_cf_node, nested_node, node, &if_stmt->then_list) {
5474 scan_tess_ctrl(nested_node, &then_tessfactor_writemask,
5475 cond_block_tf_writemask,
5476 tessfactors_are_def_in_all_invocs, true);
5477 }
5478
5479 foreach_list_typed(nir_cf_node, nested_node, node, &if_stmt->else_list) {
5480 scan_tess_ctrl(nested_node, &else_tessfactor_writemask,
5481 cond_block_tf_writemask,
5482 tessfactors_are_def_in_all_invocs, true);
5483 }
5484
5485 if (then_tessfactor_writemask || else_tessfactor_writemask) {
5486 /* If both statements write the same tess factor channels,
5487 * we can say that the upper block writes them too.
5488 */
5489 *upper_block_tf_writemask |= then_tessfactor_writemask &
5490 else_tessfactor_writemask;
5491 *cond_block_tf_writemask |= then_tessfactor_writemask |
5492 else_tessfactor_writemask;
5493 }
5494
5495 break;
5496 }
5497 case nir_cf_node_loop: {
5498 nir_loop *loop = nir_cf_node_as_loop(cf_node);
5499 foreach_list_typed(nir_cf_node, nested_node, node, &loop->body) {
5500 scan_tess_ctrl(nested_node, cond_block_tf_writemask,
5501 cond_block_tf_writemask,
5502 tessfactors_are_def_in_all_invocs, true);
5503 }
5504
5505 break;
5506 }
5507 default:
5508 unreachable("unknown cf node type");
5509 }
5510 }
5511
5512 bool
5513 ac_are_tessfactors_def_in_all_invocs(const struct nir_shader *nir)
5514 {
5515 assert(nir->info.stage == MESA_SHADER_TESS_CTRL);
5516
5517 /* The pass works as follows:
5518 * If all codepaths write tess factors, we can say that all
5519 * invocations define tess factors.
5520 *
5521 * Each tess factor channel is tracked separately.
5522 */
5523 unsigned main_block_tf_writemask = 0; /* if main block writes tess factors */
5524 unsigned cond_block_tf_writemask = 0; /* if cond block writes tess factors */
5525
5526 /* Initial value = true. Here the pass will accumulate results from
5527 * multiple segments surrounded by barriers. If tess factors aren't
5528 * written at all, it's a shader bug and we don't care if this will be
5529 * true.
5530 */
5531 bool tessfactors_are_def_in_all_invocs = true;
5532
5533 nir_foreach_function(function, nir) {
5534 if (function->impl) {
5535 foreach_list_typed(nir_cf_node, node, node, &function->impl->body) {
5536 scan_tess_ctrl(node, &main_block_tf_writemask,
5537 &cond_block_tf_writemask,
5538 &tessfactors_are_def_in_all_invocs,
5539 false);
5540 }
5541 }
5542 }
5543
5544 /* Accumulate the result for the last code segment separated by a
5545 * barrier.
5546 */
5547 if (main_block_tf_writemask || cond_block_tf_writemask) {
5548 tessfactors_are_def_in_all_invocs &=
5549 !(cond_block_tf_writemask & ~main_block_tf_writemask);
5550 }
5551
5552 return tessfactors_are_def_in_all_invocs;
5553 }