projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
ac/nir: handle nir_op_[fiu]2[fiu]mp 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 case nir_op_f2fmp:
958 src[0] = ac_to_float(&ctx->ac, src[0]);
959 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
960 result = LLVMBuildFPExt(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
961 else
962 result = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
963 break;
964 case nir_op_u2u8:
965 case nir_op_u2u16:
966 case nir_op_u2ump:
967 case nir_op_u2u32:
968 case nir_op_u2u64:
969 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
970 result = LLVMBuildZExt(ctx->ac.builder, src[0], def_type, "");
971 else
972 result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
973 break;
974 case nir_op_i2i8:
975 case nir_op_i2i16:
976 case nir_op_i2imp:
977 case nir_op_i2i32:
978 case nir_op_i2i64:
979 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
980 result = LLVMBuildSExt(ctx->ac.builder, src[0], def_type, "");
981 else
982 result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
983 break;
984 case nir_op_b32csel:
985 result = emit_bcsel(&ctx->ac, src[0], src[1], src[2]);
986 break;
987 case nir_op_find_lsb:
988 result = ac_find_lsb(&ctx->ac, ctx->ac.i32, src[0]);
989 break;
990 case nir_op_ufind_msb:
991 result = ac_build_umsb(&ctx->ac, src[0], ctx->ac.i32);
992 break;
993 case nir_op_ifind_msb:
994 result = ac_build_imsb(&ctx->ac, src[0], ctx->ac.i32);
995 break;
996 case nir_op_uadd_carry:
997 result = emit_uint_carry(&ctx->ac, "llvm.uadd.with.overflow.i32", src[0], src[1]);
998 break;
999 case nir_op_usub_borrow:
1000 result = emit_uint_carry(&ctx->ac, "llvm.usub.with.overflow.i32", src[0], src[1]);
1001 break;
1002 case nir_op_b2f16:
1003 case nir_op_b2f32:
1004 case nir_op_b2f64:
1005 result = emit_b2f(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
1006 break;
1007 case nir_op_f2b32:
1008 result = emit_f2b(&ctx->ac, src[0]);
1009 break;
1010 case nir_op_b2i8:
1011 case nir_op_b2i16:
1012 case nir_op_b2i32:
1013 case nir_op_b2i64:
1014 result = emit_b2i(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
1015 break;
1016 case nir_op_i2b32:
1017 result = emit_i2b(&ctx->ac, src[0]);
1018 break;
1019 case nir_op_fquantize2f16:
1020 result = emit_f2f16(&ctx->ac, src[0]);
1021 break;
1022 case nir_op_umul_high:
1023 result = emit_umul_high(&ctx->ac, src[0], src[1]);
1024 break;
1025 case nir_op_imul_high:
1026 result = emit_imul_high(&ctx->ac, src[0], src[1]);
1027 break;
1028 case nir_op_pack_half_2x16:
1029 result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pkrtz_f16);
1030 break;
1031 case nir_op_pack_snorm_2x16:
1032 result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pknorm_i16);
1033 break;
1034 case nir_op_pack_unorm_2x16:
1035 result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pknorm_u16);
1036 break;
1037 case nir_op_unpack_half_2x16:
1038 result = emit_unpack_half_2x16(&ctx->ac, src[0]);
1039 break;
1040 case nir_op_fddx:
1041 case nir_op_fddy:
1042 case nir_op_fddx_fine:
1043 case nir_op_fddy_fine:
1044 case nir_op_fddx_coarse:
1045 case nir_op_fddy_coarse:
1046 result = emit_ddxy(ctx, instr->op, src[0]);
1047 break;
1048
1049 case nir_op_unpack_64_2x32_split_x: {
1050 assert(ac_get_llvm_num_components(src[0]) == 1);
1051 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1052 ctx->ac.v2i32,
1053 "");
1054 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1055 ctx->ac.i32_0, "");
1056 break;
1057 }
1058
1059 case nir_op_unpack_64_2x32_split_y: {
1060 assert(ac_get_llvm_num_components(src[0]) == 1);
1061 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1062 ctx->ac.v2i32,
1063 "");
1064 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1065 ctx->ac.i32_1, "");
1066 break;
1067 }
1068
1069 case nir_op_pack_64_2x32_split: {
1070 LLVMValueRef tmp = ac_build_gather_values(&ctx->ac, src, 2);
1071 result = LLVMBuildBitCast(ctx->ac.builder, tmp, ctx->ac.i64, "");
1072 break;
1073 }
1074
1075 case nir_op_pack_32_2x16_split: {
1076 LLVMValueRef tmp = ac_build_gather_values(&ctx->ac, src, 2);
1077 result = LLVMBuildBitCast(ctx->ac.builder, tmp, ctx->ac.i32, "");
1078 break;
1079 }
1080
1081 case nir_op_unpack_32_2x16_split_x: {
1082 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1083 ctx->ac.v2i16,
1084 "");
1085 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1086 ctx->ac.i32_0, "");
1087 break;
1088 }
1089
1090 case nir_op_unpack_32_2x16_split_y: {
1091 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1092 ctx->ac.v2i16,
1093 "");
1094 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1095 ctx->ac.i32_1, "");
1096 break;
1097 }
1098
1099 case nir_op_cube_face_coord: {
1100 src[0] = ac_to_float(&ctx->ac, src[0]);
1101 LLVMValueRef results[2];
1102 LLVMValueRef in[3];
1103 for (unsigned chan = 0; chan < 3; chan++)
1104 in[chan] = ac_llvm_extract_elem(&ctx->ac, src[0], chan);
1105 results[0] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubesc",
1106 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1107 results[1] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubetc",
1108 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1109 LLVMValueRef ma = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubema",
1110 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1111 results[0] = ac_build_fdiv(&ctx->ac, results[0], ma);
1112 results[1] = ac_build_fdiv(&ctx->ac, results[1], ma);
1113 LLVMValueRef offset = LLVMConstReal(ctx->ac.f32, 0.5);
1114 results[0] = LLVMBuildFAdd(ctx->ac.builder, results[0], offset, "");
1115 results[1] = LLVMBuildFAdd(ctx->ac.builder, results[1], offset, "");
1116 result = ac_build_gather_values(&ctx->ac, results, 2);
1117 break;
1118 }
1119
1120 case nir_op_cube_face_index: {
1121 src[0] = ac_to_float(&ctx->ac, src[0]);
1122 LLVMValueRef in[3];
1123 for (unsigned chan = 0; chan < 3; chan++)
1124 in[chan] = ac_llvm_extract_elem(&ctx->ac, src[0], chan);
1125 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubeid",
1126 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1127 break;
1128 }
1129
1130 case nir_op_fmin3:
1131 result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
1132 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
1133 result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
1134 ac_to_float_type(&ctx->ac, def_type), result, src[2]);
1135 break;
1136 case nir_op_umin3:
1137 result = ac_build_umin(&ctx->ac, src[0], src[1]);
1138 result = ac_build_umin(&ctx->ac, result, src[2]);
1139 break;
1140 case nir_op_imin3:
1141 result = ac_build_imin(&ctx->ac, src[0], src[1]);
1142 result = ac_build_imin(&ctx->ac, result, src[2]);
1143 break;
1144 case nir_op_fmax3:
1145 result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
1146 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
1147 result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
1148 ac_to_float_type(&ctx->ac, def_type), result, src[2]);
1149 break;
1150 case nir_op_umax3:
1151 result = ac_build_umax(&ctx->ac, src[0], src[1]);
1152 result = ac_build_umax(&ctx->ac, result, src[2]);
1153 break;
1154 case nir_op_imax3:
1155 result = ac_build_imax(&ctx->ac, src[0], src[1]);
1156 result = ac_build_imax(&ctx->ac, result, src[2]);
1157 break;
1158 case nir_op_fmed3: {
1159 src[0] = ac_to_float(&ctx->ac, src[0]);
1160 src[1] = ac_to_float(&ctx->ac, src[1]);
1161 src[2] = ac_to_float(&ctx->ac, src[2]);
1162 result = ac_build_fmed3(&ctx->ac, src[0], src[1], src[2],
1163 instr->dest.dest.ssa.bit_size);
1164 break;
1165 }
1166 case nir_op_imed3: {
1167 LLVMValueRef tmp1 = ac_build_imin(&ctx->ac, src[0], src[1]);
1168 LLVMValueRef tmp2 = ac_build_imax(&ctx->ac, src[0], src[1]);
1169 tmp2 = ac_build_imin(&ctx->ac, tmp2, src[2]);
1170 result = ac_build_imax(&ctx->ac, tmp1, tmp2);
1171 break;
1172 }
1173 case nir_op_umed3: {
1174 LLVMValueRef tmp1 = ac_build_umin(&ctx->ac, src[0], src[1]);
1175 LLVMValueRef tmp2 = ac_build_umax(&ctx->ac, src[0], src[1]);
1176 tmp2 = ac_build_umin(&ctx->ac, tmp2, src[2]);
1177 result = ac_build_umax(&ctx->ac, tmp1, tmp2);
1178 break;
1179 }
1180
1181 default:
1182 fprintf(stderr, "Unknown NIR alu instr: ");
1183 nir_print_instr(&instr->instr, stderr);
1184 fprintf(stderr, "\n");
1185 abort();
1186 }
1187
1188 if (result) {
1189 assert(instr->dest.dest.is_ssa);
1190 result = ac_to_integer_or_pointer(&ctx->ac, result);
1191 ctx->ssa_defs[instr->dest.dest.ssa.index] = result;
1192 }
1193
1194 if (instr->exact)
1195 ac_restore_inexact_math(ctx->ac.builder, saved_inexact);
1196 }
1197
1198 static void visit_load_const(struct ac_nir_context *ctx,
1199 const nir_load_const_instr *instr)
1200 {
1201 LLVMValueRef values[4], value = NULL;
1202 LLVMTypeRef element_type =
1203 LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
1204
1205 for (unsigned i = 0; i < instr->def.num_components; ++i) {
1206 switch (instr->def.bit_size) {
1207 case 8:
1208 values[i] = LLVMConstInt(element_type,
1209 instr->value[i].u8, false);
1210 break;
1211 case 16:
1212 values[i] = LLVMConstInt(element_type,
1213 instr->value[i].u16, false);
1214 break;
1215 case 32:
1216 values[i] = LLVMConstInt(element_type,
1217 instr->value[i].u32, false);
1218 break;
1219 case 64:
1220 values[i] = LLVMConstInt(element_type,
1221 instr->value[i].u64, false);
1222 break;
1223 default:
1224 fprintf(stderr,
1225 "unsupported nir load_const bit_size: %d\n",
1226 instr->def.bit_size);
1227 abort();
1228 }
1229 }
1230 if (instr->def.num_components > 1) {
1231 value = LLVMConstVector(values, instr->def.num_components);
1232 } else
1233 value = values[0];
1234
1235 ctx->ssa_defs[instr->def.index] = value;
1236 }
1237
1238 static LLVMValueRef
1239 get_buffer_size(struct ac_nir_context *ctx, LLVMValueRef descriptor, bool in_elements)
1240 {
1241 LLVMValueRef size =
1242 LLVMBuildExtractElement(ctx->ac.builder, descriptor,
1243 LLVMConstInt(ctx->ac.i32, 2, false), "");
1244
1245 /* GFX8 only */
1246 if (ctx->ac.chip_class == GFX8 && in_elements) {
1247 /* On GFX8, the descriptor contains the size in bytes,
1248 * but TXQ must return the size in elements.
1249 * The stride is always non-zero for resources using TXQ.
1250 */
1251 LLVMValueRef stride =
1252 LLVMBuildExtractElement(ctx->ac.builder, descriptor,
1253 ctx->ac.i32_1, "");
1254 stride = LLVMBuildLShr(ctx->ac.builder, stride,
1255 LLVMConstInt(ctx->ac.i32, 16, false), "");
1256 stride = LLVMBuildAnd(ctx->ac.builder, stride,
1257 LLVMConstInt(ctx->ac.i32, 0x3fff, false), "");
1258
1259 size = LLVMBuildUDiv(ctx->ac.builder, size, stride, "");
1260 }
1261 return size;
1262 }
1263
1264 /* Gather4 should follow the same rules as bilinear filtering, but the hardware
1265 * incorrectly forces nearest filtering if the texture format is integer.
1266 * The only effect it has on Gather4, which always returns 4 texels for
1267 * bilinear filtering, is that the final coordinates are off by 0.5 of
1268 * the texel size.
1269 *
1270 * The workaround is to subtract 0.5 from the unnormalized coordinates,
1271 * or (0.5 / size) from the normalized coordinates.
1272 *
1273 * However, cube textures with 8_8_8_8 data formats require a different
1274 * workaround of overriding the num format to USCALED/SSCALED. This would lose
1275 * precision in 32-bit data formats, so it needs to be applied dynamically at
1276 * runtime. In this case, return an i1 value that indicates whether the
1277 * descriptor was overridden (and hence a fixup of the sampler result is needed).
1278 */
1279 static LLVMValueRef lower_gather4_integer(struct ac_llvm_context *ctx,
1280 nir_variable *var,
1281 struct ac_image_args *args,
1282 const nir_tex_instr *instr)
1283 {
1284 const struct glsl_type *type = glsl_without_array(var->type);
1285 enum glsl_base_type stype = glsl_get_sampler_result_type(type);
1286 LLVMValueRef wa_8888 = NULL;
1287 LLVMValueRef half_texel[2];
1288 LLVMValueRef result;
1289
1290 assert(stype == GLSL_TYPE_INT || stype == GLSL_TYPE_UINT);
1291
1292 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
1293 LLVMValueRef formats;
1294 LLVMValueRef data_format;
1295 LLVMValueRef wa_formats;
1296
1297 formats = LLVMBuildExtractElement(ctx->builder, args->resource, ctx->i32_1, "");
1298
1299 data_format = LLVMBuildLShr(ctx->builder, formats,
1300 LLVMConstInt(ctx->i32, 20, false), "");
1301 data_format = LLVMBuildAnd(ctx->builder, data_format,
1302 LLVMConstInt(ctx->i32, (1u << 6) - 1, false), "");
1303 wa_8888 = LLVMBuildICmp(
1304 ctx->builder, LLVMIntEQ, data_format,
1305 LLVMConstInt(ctx->i32, V_008F14_IMG_DATA_FORMAT_8_8_8_8, false),
1306 "");
1307
1308 uint32_t wa_num_format =
1309 stype == GLSL_TYPE_UINT ?
1310 S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_USCALED) :
1311 S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_SSCALED);
1312 wa_formats = LLVMBuildAnd(ctx->builder, formats,
1313 LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT, false),
1314 "");
1315 wa_formats = LLVMBuildOr(ctx->builder, wa_formats,
1316 LLVMConstInt(ctx->i32, wa_num_format, false), "");
1317
1318 formats = LLVMBuildSelect(ctx->builder, wa_8888, wa_formats, formats, "");
1319 args->resource = LLVMBuildInsertElement(
1320 ctx->builder, args->resource, formats, ctx->i32_1, "");
1321 }
1322
1323 if (instr->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
1324 assert(!wa_8888);
1325 half_texel[0] = half_texel[1] = LLVMConstReal(ctx->f32, -0.5);
1326 } else {
1327 struct ac_image_args resinfo = {};
1328 LLVMBasicBlockRef bbs[2];
1329
1330 LLVMValueRef unnorm = NULL;
1331 LLVMValueRef default_offset = ctx->f32_0;
1332 if (instr->sampler_dim == GLSL_SAMPLER_DIM_2D &&
1333 !instr->is_array) {
1334 /* In vulkan, whether the sampler uses unnormalized
1335 * coordinates or not is a dynamic property of the
1336 * sampler. Hence, to figure out whether or not we
1337 * need to divide by the texture size, we need to test
1338 * the sampler at runtime. This tests the bit set by
1339 * radv_init_sampler().
1340 */
1341 LLVMValueRef sampler0 =
1342 LLVMBuildExtractElement(ctx->builder, args->sampler, ctx->i32_0, "");
1343 sampler0 = LLVMBuildLShr(ctx->builder, sampler0,
1344 LLVMConstInt(ctx->i32, 15, false), "");
1345 sampler0 = LLVMBuildAnd(ctx->builder, sampler0, ctx->i32_1, "");
1346 unnorm = LLVMBuildICmp(ctx->builder, LLVMIntEQ, sampler0, ctx->i32_1, "");
1347 default_offset = LLVMConstReal(ctx->f32, -0.5);
1348 }
1349
1350 bbs[0] = LLVMGetInsertBlock(ctx->builder);
1351 if (wa_8888 || unnorm) {
1352 assert(!(wa_8888 && unnorm));
1353 LLVMValueRef not_needed = wa_8888 ? wa_8888 : unnorm;
1354 /* Skip the texture size query entirely if we don't need it. */
1355 ac_build_ifcc(ctx, LLVMBuildNot(ctx->builder, not_needed, ""), 2000);
1356 bbs[1] = LLVMGetInsertBlock(ctx->builder);
1357 }
1358
1359 /* Query the texture size. */
1360 resinfo.dim = ac_get_sampler_dim(ctx->chip_class, instr->sampler_dim, instr->is_array);
1361 resinfo.opcode = ac_image_get_resinfo;
1362 resinfo.dmask = 0xf;
1363 resinfo.lod = ctx->i32_0;
1364 resinfo.resource = args->resource;
1365 resinfo.attributes = AC_FUNC_ATTR_READNONE;
1366 LLVMValueRef size = ac_build_image_opcode(ctx, &resinfo);
1367
1368 /* Compute -0.5 / size. */
1369 for (unsigned c = 0; c < 2; c++) {
1370 half_texel[c] =
1371 LLVMBuildExtractElement(ctx->builder, size,
1372 LLVMConstInt(ctx->i32, c, 0), "");
1373 half_texel[c] = LLVMBuildUIToFP(ctx->builder, half_texel[c], ctx->f32, "");
1374 half_texel[c] = ac_build_fdiv(ctx, ctx->f32_1, half_texel[c]);
1375 half_texel[c] = LLVMBuildFMul(ctx->builder, half_texel[c],
1376 LLVMConstReal(ctx->f32, -0.5), "");
1377 }
1378
1379 if (wa_8888 || unnorm) {
1380 ac_build_endif(ctx, 2000);
1381
1382 for (unsigned c = 0; c < 2; c++) {
1383 LLVMValueRef values[2] = { default_offset, half_texel[c] };
1384 half_texel[c] = ac_build_phi(ctx, ctx->f32, 2,
1385 values, bbs);
1386 }
1387 }
1388 }
1389
1390 for (unsigned c = 0; c < 2; c++) {
1391 LLVMValueRef tmp;
1392 tmp = LLVMBuildBitCast(ctx->builder, args->coords[c], ctx->f32, "");
1393 args->coords[c] = LLVMBuildFAdd(ctx->builder, tmp, half_texel[c], "");
1394 }
1395
1396 args->attributes = AC_FUNC_ATTR_READNONE;
1397 result = ac_build_image_opcode(ctx, args);
1398
1399 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
1400 LLVMValueRef tmp, tmp2;
1401
1402 /* if the cube workaround is in place, f2i the result. */
1403 for (unsigned c = 0; c < 4; c++) {
1404 tmp = LLVMBuildExtractElement(ctx->builder, result, LLVMConstInt(ctx->i32, c, false), "");
1405 if (stype == GLSL_TYPE_UINT)
1406 tmp2 = LLVMBuildFPToUI(ctx->builder, tmp, ctx->i32, "");
1407 else
1408 tmp2 = LLVMBuildFPToSI(ctx->builder, tmp, ctx->i32, "");
1409 tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->i32, "");
1410 tmp2 = LLVMBuildBitCast(ctx->builder, tmp2, ctx->i32, "");
1411 tmp = LLVMBuildSelect(ctx->builder, wa_8888, tmp2, tmp, "");
1412 tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->f32, "");
1413 result = LLVMBuildInsertElement(ctx->builder, result, tmp, LLVMConstInt(ctx->i32, c, false), "");
1414 }
1415 }
1416 return result;
1417 }
1418
1419 static nir_deref_instr *get_tex_texture_deref(const nir_tex_instr *instr)
1420 {
1421 nir_deref_instr *texture_deref_instr = NULL;
1422
1423 for (unsigned i = 0; i < instr->num_srcs; i++) {
1424 switch (instr->src[i].src_type) {
1425 case nir_tex_src_texture_deref:
1426 texture_deref_instr = nir_src_as_deref(instr->src[i].src);
1427 break;
1428 default:
1429 break;
1430 }
1431 }
1432 return texture_deref_instr;
1433 }
1434
1435 static LLVMValueRef build_tex_intrinsic(struct ac_nir_context *ctx,
1436 const nir_tex_instr *instr,
1437 struct ac_image_args *args)
1438 {
1439 if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
1440 unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
1441
1442 assert(instr->dest.is_ssa);
1443 return ac_build_buffer_load_format(&ctx->ac,
1444 args->resource,
1445 args->coords[0],
1446 ctx->ac.i32_0,
1447 util_last_bit(mask),
1448 0, true,
1449 instr->dest.ssa.bit_size == 16);
1450 }
1451
1452 args->opcode = ac_image_sample;
1453
1454 switch (instr->op) {
1455 case nir_texop_txf:
1456 case nir_texop_txf_ms:
1457 case nir_texop_samples_identical:
1458 args->opcode = args->level_zero ||
1459 instr->sampler_dim == GLSL_SAMPLER_DIM_MS ?
1460 ac_image_load : ac_image_load_mip;
1461 args->level_zero = false;
1462 break;
1463 case nir_texop_txs:
1464 case nir_texop_query_levels:
1465 args->opcode = ac_image_get_resinfo;
1466 if (!args->lod)
1467 args->lod = ctx->ac.i32_0;
1468 args->level_zero = false;
1469 break;
1470 case nir_texop_tex:
1471 if (ctx->stage != MESA_SHADER_FRAGMENT) {
1472 assert(!args->lod);
1473 args->level_zero = true;
1474 }
1475 break;
1476 case nir_texop_tg4:
1477 args->opcode = ac_image_gather4;
1478 if (!args->lod && !args->bias)
1479 args->level_zero = true;
1480 break;
1481 case nir_texop_lod:
1482 args->opcode = ac_image_get_lod;
1483 break;
1484 case nir_texop_fragment_fetch:
1485 case nir_texop_fragment_mask_fetch:
1486 args->opcode = ac_image_load;
1487 args->level_zero = false;
1488 break;
1489 default:
1490 break;
1491 }
1492
1493 if (instr->op == nir_texop_tg4 && ctx->ac.chip_class <= GFX8) {
1494 nir_deref_instr *texture_deref_instr = get_tex_texture_deref(instr);
1495 nir_variable *var = nir_deref_instr_get_variable(texture_deref_instr);
1496 const struct glsl_type *type = glsl_without_array(var->type);
1497 enum glsl_base_type stype = glsl_get_sampler_result_type(type);
1498 if (stype == GLSL_TYPE_UINT || stype == GLSL_TYPE_INT) {
1499 return lower_gather4_integer(&ctx->ac, var, args, instr);
1500 }
1501 }
1502
1503 /* Fixup for GFX9 which allocates 1D textures as 2D. */
1504 if (instr->op == nir_texop_lod && ctx->ac.chip_class == GFX9) {
1505 if ((args->dim == ac_image_2darray ||
1506 args->dim == ac_image_2d) && !args->coords[1]) {
1507 args->coords[1] = ctx->ac.i32_0;
1508 }
1509 }
1510
1511 args->attributes = AC_FUNC_ATTR_READNONE;
1512 bool cs_derivs = ctx->stage == MESA_SHADER_COMPUTE &&
1513 ctx->info->cs.derivative_group != DERIVATIVE_GROUP_NONE;
1514 if (ctx->stage == MESA_SHADER_FRAGMENT || cs_derivs) {
1515 /* Prevent texture instructions with implicit derivatives from being
1516 * sinked into branches. */
1517 switch (instr->op) {
1518 case nir_texop_tex:
1519 case nir_texop_txb:
1520 case nir_texop_lod:
1521 args->attributes |= AC_FUNC_ATTR_CONVERGENT;
1522 break;
1523 default:
1524 break;
1525 }
1526 }
1527
1528 return ac_build_image_opcode(&ctx->ac, args);
1529 }
1530
1531 static LLVMValueRef visit_vulkan_resource_reindex(struct ac_nir_context *ctx,
1532 nir_intrinsic_instr *instr)
1533 {
1534 LLVMValueRef ptr = get_src(ctx, instr->src[0]);
1535 LLVMValueRef index = get_src(ctx, instr->src[1]);
1536
1537 LLVMValueRef result = LLVMBuildGEP(ctx->ac.builder, ptr, &index, 1, "");
1538 LLVMSetMetadata(result, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
1539 return result;
1540 }
1541
1542 static LLVMValueRef visit_load_push_constant(struct ac_nir_context *ctx,
1543 nir_intrinsic_instr *instr)
1544 {
1545 LLVMValueRef ptr, addr;
1546 LLVMValueRef src0 = get_src(ctx, instr->src[0]);
1547 unsigned index = nir_intrinsic_base(instr);
1548
1549 addr = LLVMConstInt(ctx->ac.i32, index, 0);
1550 addr = LLVMBuildAdd(ctx->ac.builder, addr, src0, "");
1551
1552 /* Load constant values from user SGPRS when possible, otherwise
1553 * fallback to the default path that loads directly from memory.
1554 */
1555 if (LLVMIsConstant(src0) &&
1556 instr->dest.ssa.bit_size == 32) {
1557 unsigned count = instr->dest.ssa.num_components;
1558 unsigned offset = index;
1559
1560 offset += LLVMConstIntGetZExtValue(src0);
1561 offset /= 4;
1562
1563 offset -= ctx->args->base_inline_push_consts;
1564
1565 unsigned num_inline_push_consts = ctx->args->num_inline_push_consts;
1566 if (offset + count <= num_inline_push_consts) {
1567 LLVMValueRef push_constants[num_inline_push_consts];
1568 for (unsigned i = 0; i < num_inline_push_consts; i++)
1569 push_constants[i] = ac_get_arg(&ctx->ac,
1570 ctx->args->inline_push_consts[i]);
1571 return ac_build_gather_values(&ctx->ac,
1572 push_constants + offset,
1573 count);
1574 }
1575 }
1576
1577 ptr = LLVMBuildGEP(ctx->ac.builder,
1578 ac_get_arg(&ctx->ac, ctx->args->push_constants), &addr, 1, "");
1579
1580 if (instr->dest.ssa.bit_size == 8) {
1581 unsigned load_dwords = instr->dest.ssa.num_components > 1 ? 2 : 1;
1582 LLVMTypeRef vec_type = LLVMVectorType(ctx->ac.i8, 4 * load_dwords);
1583 ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
1584 LLVMValueRef res = LLVMBuildLoad(ctx->ac.builder, ptr, "");
1585
1586 LLVMValueRef params[3];
1587 if (load_dwords > 1) {
1588 LLVMValueRef res_vec = LLVMBuildBitCast(ctx->ac.builder, res, ctx->ac.v2i32, "");
1589 params[0] = LLVMBuildExtractElement(ctx->ac.builder, res_vec, LLVMConstInt(ctx->ac.i32, 1, false), "");
1590 params[1] = LLVMBuildExtractElement(ctx->ac.builder, res_vec, LLVMConstInt(ctx->ac.i32, 0, false), "");
1591 } else {
1592 res = LLVMBuildBitCast(ctx->ac.builder, res, ctx->ac.i32, "");
1593 params[0] = ctx->ac.i32_0;
1594 params[1] = res;
1595 }
1596 params[2] = addr;
1597 res = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.alignbyte", ctx->ac.i32, params, 3, 0);
1598
1599 res = LLVMBuildTrunc(ctx->ac.builder, res, LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.num_components * 8), "");
1600 if (instr->dest.ssa.num_components > 1)
1601 res = LLVMBuildBitCast(ctx->ac.builder, res, LLVMVectorType(ctx->ac.i8, instr->dest.ssa.num_components), "");
1602 return res;
1603 } else if (instr->dest.ssa.bit_size == 16) {
1604 unsigned load_dwords = instr->dest.ssa.num_components / 2 + 1;
1605 LLVMTypeRef vec_type = LLVMVectorType(ctx->ac.i16, 2 * load_dwords);
1606 ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
1607 LLVMValueRef res = LLVMBuildLoad(ctx->ac.builder, ptr, "");
1608 res = LLVMBuildBitCast(ctx->ac.builder, res, vec_type, "");
1609 LLVMValueRef cond = LLVMBuildLShr(ctx->ac.builder, addr, ctx->ac.i32_1, "");
1610 cond = LLVMBuildTrunc(ctx->ac.builder, cond, ctx->ac.i1, "");
1611 LLVMValueRef mask[] = { LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
1612 LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
1613 LLVMConstInt(ctx->ac.i32, 4, false)};
1614 LLVMValueRef swizzle_aligned = LLVMConstVector(&mask[0], instr->dest.ssa.num_components);
1615 LLVMValueRef swizzle_unaligned = LLVMConstVector(&mask[1], instr->dest.ssa.num_components);
1616 LLVMValueRef shuffle_aligned = LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_aligned, "");
1617 LLVMValueRef shuffle_unaligned = LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_unaligned, "");
1618 res = LLVMBuildSelect(ctx->ac.builder, cond, shuffle_unaligned, shuffle_aligned, "");
1619 return LLVMBuildBitCast(ctx->ac.builder, res, get_def_type(ctx, &instr->dest.ssa), "");
1620 }
1621
1622 ptr = ac_cast_ptr(&ctx->ac, ptr, get_def_type(ctx, &instr->dest.ssa));
1623
1624 return LLVMBuildLoad(ctx->ac.builder, ptr, "");
1625 }
1626
1627 static LLVMValueRef visit_get_buffer_size(struct ac_nir_context *ctx,
1628 const nir_intrinsic_instr *instr)
1629 {
1630 LLVMValueRef index = get_src(ctx, instr->src[0]);
1631
1632 return get_buffer_size(ctx, ctx->abi->load_ssbo(ctx->abi, index, false), false);
1633 }
1634
1635 static uint32_t widen_mask(uint32_t mask, unsigned multiplier)
1636 {
1637 uint32_t new_mask = 0;
1638 for(unsigned i = 0; i < 32 && (1u << i) <= mask; ++i)
1639 if (mask & (1u << i))
1640 new_mask |= ((1u << multiplier) - 1u) << (i * multiplier);
1641 return new_mask;
1642 }
1643
1644 static LLVMValueRef extract_vector_range(struct ac_llvm_context *ctx, LLVMValueRef src,
1645 unsigned start, unsigned count)
1646 {
1647 LLVMValueRef mask[] = {
1648 ctx->i32_0, ctx->i32_1,
1649 LLVMConstInt(ctx->i32, 2, false), LLVMConstInt(ctx->i32, 3, false) };
1650
1651 unsigned src_elements = ac_get_llvm_num_components(src);
1652
1653 if (count == src_elements) {
1654 assert(start == 0);
1655 return src;
1656 } else if (count == 1) {
1657 assert(start < src_elements);
1658 return LLVMBuildExtractElement(ctx->builder, src, mask[start], "");
1659 } else {
1660 assert(start + count <= src_elements);
1661 assert(count <= 4);
1662 LLVMValueRef swizzle = LLVMConstVector(&mask[start], count);
1663 return LLVMBuildShuffleVector(ctx->builder, src, src, swizzle, "");
1664 }
1665 }
1666
1667 static unsigned get_cache_policy(struct ac_nir_context *ctx,
1668 enum gl_access_qualifier access,
1669 bool may_store_unaligned,
1670 bool writeonly_memory)
1671 {
1672 unsigned cache_policy = 0;
1673
1674 /* GFX6 has a TC L1 bug causing corruption of 8bit/16bit stores. All
1675 * store opcodes not aligned to a dword are affected. The only way to
1676 * get unaligned stores is through shader images.
1677 */
1678 if (((may_store_unaligned && ctx->ac.chip_class == GFX6) ||
1679 /* If this is write-only, don't keep data in L1 to prevent
1680 * evicting L1 cache lines that may be needed by other
1681 * instructions.
1682 */
1683 writeonly_memory ||
1684 access & (ACCESS_COHERENT | ACCESS_VOLATILE))) {
1685 cache_policy |= ac_glc;
1686 }
1687
1688 if (access & ACCESS_STREAM_CACHE_POLICY)
1689 cache_policy |= ac_slc | ac_glc;
1690
1691 return cache_policy;
1692 }
1693
1694 static LLVMValueRef enter_waterfall_ssbo(struct ac_nir_context *ctx,
1695 struct waterfall_context *wctx,
1696 const nir_intrinsic_instr *instr,
1697 nir_src src)
1698 {
1699 return enter_waterfall(ctx, wctx, get_src(ctx, src),
1700 nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
1701 }
1702
1703 static void visit_store_ssbo(struct ac_nir_context *ctx,
1704 nir_intrinsic_instr *instr)
1705 {
1706 if (ctx->ac.postponed_kill) {
1707 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
1708 ctx->ac.postponed_kill, "");
1709 ac_build_ifcc(&ctx->ac, cond, 7000);
1710 }
1711
1712 LLVMValueRef src_data = get_src(ctx, instr->src[0]);
1713 int elem_size_bytes = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src_data)) / 8;
1714 unsigned writemask = nir_intrinsic_write_mask(instr);
1715 enum gl_access_qualifier access = nir_intrinsic_access(instr);
1716 bool writeonly_memory = access & ACCESS_NON_READABLE;
1717 unsigned cache_policy = get_cache_policy(ctx, access, false, writeonly_memory);
1718
1719 struct waterfall_context wctx;
1720 LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[1]);
1721
1722 LLVMValueRef rsrc = ctx->abi->load_ssbo(ctx->abi, rsrc_base, true);
1723 LLVMValueRef base_data = src_data;
1724 base_data = ac_trim_vector(&ctx->ac, base_data, instr->num_components);
1725 LLVMValueRef base_offset = get_src(ctx, instr->src[2]);
1726
1727 while (writemask) {
1728 int start, count;
1729 LLVMValueRef data, offset;
1730 LLVMTypeRef data_type;
1731
1732 u_bit_scan_consecutive_range(&writemask, &start, &count);
1733
1734 /* Due to an LLVM limitation with LLVM < 9, split 3-element
1735 * writes into a 2-element and a 1-element write. */
1736 if (count == 3 &&
1737 (elem_size_bytes != 4 || !ac_has_vec3_support(ctx->ac.chip_class, false))) {
1738 writemask |= 1 << (start + 2);
1739 count = 2;
1740 }
1741 int num_bytes = count * elem_size_bytes; /* count in bytes */
1742
1743 /* we can only store 4 DWords at the same time.
1744 * can only happen for 64 Bit vectors. */
1745 if (num_bytes > 16) {
1746 writemask |= ((1u << (count - 2)) - 1u) << (start + 2);
1747 count = 2;
1748 num_bytes = 16;
1749 }
1750
1751 /* check alignment of 16 Bit stores */
1752 if (elem_size_bytes == 2 && num_bytes > 2 && (start % 2) == 1) {
1753 writemask |= ((1u << (count - 1)) - 1u) << (start + 1);
1754 count = 1;
1755 num_bytes = 2;
1756 }
1757
1758 /* Due to alignment issues, split stores of 8-bit/16-bit
1759 * vectors.
1760 */
1761 if (ctx->ac.chip_class == GFX6 && count > 1 && elem_size_bytes < 4) {
1762 writemask |= ((1u << (count - 1)) - 1u) << (start + 1);
1763 count = 1;
1764 num_bytes = elem_size_bytes;
1765 }
1766
1767 data = extract_vector_range(&ctx->ac, base_data, start, count);
1768
1769 offset = LLVMBuildAdd(ctx->ac.builder, base_offset,
1770 LLVMConstInt(ctx->ac.i32, start * elem_size_bytes, false), "");
1771
1772 if (num_bytes == 1) {
1773 ac_build_tbuffer_store_byte(&ctx->ac, rsrc, data,
1774 offset, ctx->ac.i32_0,
1775 cache_policy);
1776 } else if (num_bytes == 2) {
1777 ac_build_tbuffer_store_short(&ctx->ac, rsrc, data,
1778 offset, ctx->ac.i32_0,
1779 cache_policy);
1780 } else {
1781 int num_channels = num_bytes / 4;
1782
1783 switch (num_bytes) {
1784 case 16: /* v4f32 */
1785 data_type = ctx->ac.v4f32;
1786 break;
1787 case 12: /* v3f32 */
1788 data_type = ctx->ac.v3f32;
1789 break;
1790 case 8: /* v2f32 */
1791 data_type = ctx->ac.v2f32;
1792 break;
1793 case 4: /* f32 */
1794 data_type = ctx->ac.f32;
1795 break;
1796 default:
1797 unreachable("Malformed vector store.");
1798 }
1799 data = LLVMBuildBitCast(ctx->ac.builder, data, data_type, "");
1800
1801 ac_build_buffer_store_dword(&ctx->ac, rsrc, data,
1802 num_channels, offset,
1803 ctx->ac.i32_0, 0,
1804 cache_policy);
1805 }
1806 }
1807
1808 exit_waterfall(ctx, &wctx, NULL);
1809
1810 if (ctx->ac.postponed_kill)
1811 ac_build_endif(&ctx->ac, 7000);
1812 }
1813
1814 static LLVMValueRef emit_ssbo_comp_swap_64(struct ac_nir_context *ctx,
1815 LLVMValueRef descriptor,
1816 LLVMValueRef offset,
1817 LLVMValueRef compare,
1818 LLVMValueRef exchange)
1819 {
1820 LLVMBasicBlockRef start_block = NULL, then_block = NULL;
1821 if (ctx->abi->robust_buffer_access) {
1822 LLVMValueRef size = ac_llvm_extract_elem(&ctx->ac, descriptor, 2);
1823
1824 LLVMValueRef cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, offset, size, "");
1825 start_block = LLVMGetInsertBlock(ctx->ac.builder);
1826
1827 ac_build_ifcc(&ctx->ac, cond, -1);
1828
1829 then_block = LLVMGetInsertBlock(ctx->ac.builder);
1830 }
1831
1832 LLVMValueRef ptr_parts[2] = {
1833 ac_llvm_extract_elem(&ctx->ac, descriptor, 0),
1834 LLVMBuildAnd(ctx->ac.builder,
1835 ac_llvm_extract_elem(&ctx->ac, descriptor, 1),
1836 LLVMConstInt(ctx->ac.i32, 65535, 0), "")
1837 };
1838
1839 ptr_parts[1] = LLVMBuildTrunc(ctx->ac.builder, ptr_parts[1], ctx->ac.i16, "");
1840 ptr_parts[1] = LLVMBuildSExt(ctx->ac.builder, ptr_parts[1], ctx->ac.i32, "");
1841
1842 offset = LLVMBuildZExt(ctx->ac.builder, offset, ctx->ac.i64, "");
1843
1844 LLVMValueRef ptr = ac_build_gather_values(&ctx->ac, ptr_parts, 2);
1845 ptr = LLVMBuildBitCast(ctx->ac.builder, ptr, ctx->ac.i64, "");
1846 ptr = LLVMBuildAdd(ctx->ac.builder, ptr, offset, "");
1847 ptr = LLVMBuildIntToPtr(ctx->ac.builder, ptr, LLVMPointerType(ctx->ac.i64, AC_ADDR_SPACE_GLOBAL), "");
1848
1849 LLVMValueRef result = ac_build_atomic_cmp_xchg(&ctx->ac, ptr, compare, exchange, "singlethread-one-as");
1850 result = LLVMBuildExtractValue(ctx->ac.builder, result, 0, "");
1851
1852 if (ctx->abi->robust_buffer_access) {
1853 ac_build_endif(&ctx->ac, -1);
1854
1855 LLVMBasicBlockRef incoming_blocks[2] = {
1856 start_block,
1857 then_block,
1858 };
1859
1860 LLVMValueRef incoming_values[2] = {
1861 LLVMConstInt(ctx->ac.i64, 0, 0),
1862 result,
1863 };
1864 LLVMValueRef ret = LLVMBuildPhi(ctx->ac.builder, ctx->ac.i64, "");
1865 LLVMAddIncoming(ret, incoming_values, incoming_blocks, 2);
1866 return ret;
1867 } else {
1868 return result;
1869 }
1870 }
1871
1872 static LLVMValueRef visit_atomic_ssbo(struct ac_nir_context *ctx,
1873 nir_intrinsic_instr *instr)
1874 {
1875 if (ctx->ac.postponed_kill) {
1876 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
1877 ctx->ac.postponed_kill, "");
1878 ac_build_ifcc(&ctx->ac, cond, 7001);
1879 }
1880
1881 LLVMTypeRef return_type = LLVMTypeOf(get_src(ctx, instr->src[2]));
1882 const char *op;
1883 char name[64], type[8];
1884 LLVMValueRef params[6], descriptor;
1885 LLVMValueRef result;
1886 int arg_count = 0;
1887
1888 struct waterfall_context wctx;
1889 LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[0]);
1890
1891 switch (instr->intrinsic) {
1892 case nir_intrinsic_ssbo_atomic_add:
1893 op = "add";
1894 break;
1895 case nir_intrinsic_ssbo_atomic_imin:
1896 op = "smin";
1897 break;
1898 case nir_intrinsic_ssbo_atomic_umin:
1899 op = "umin";
1900 break;
1901 case nir_intrinsic_ssbo_atomic_imax:
1902 op = "smax";
1903 break;
1904 case nir_intrinsic_ssbo_atomic_umax:
1905 op = "umax";
1906 break;
1907 case nir_intrinsic_ssbo_atomic_and:
1908 op = "and";
1909 break;
1910 case nir_intrinsic_ssbo_atomic_or:
1911 op = "or";
1912 break;
1913 case nir_intrinsic_ssbo_atomic_xor:
1914 op = "xor";
1915 break;
1916 case nir_intrinsic_ssbo_atomic_exchange:
1917 op = "swap";
1918 break;
1919 case nir_intrinsic_ssbo_atomic_comp_swap:
1920 op = "cmpswap";
1921 break;
1922 default:
1923 abort();
1924 }
1925
1926 descriptor = ctx->abi->load_ssbo(ctx->abi,
1927 rsrc_base,
1928 true);
1929
1930 if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap &&
1931 return_type == ctx->ac.i64) {
1932 result = emit_ssbo_comp_swap_64(ctx, descriptor,
1933 get_src(ctx, instr->src[1]),
1934 get_src(ctx, instr->src[2]),
1935 get_src(ctx, instr->src[3]));
1936 } else {
1937 if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap) {
1938 params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[3]), 0);
1939 }
1940 params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
1941 params[arg_count++] = descriptor;
1942
1943 if (LLVM_VERSION_MAJOR >= 9) {
1944 /* XXX: The new raw/struct atomic intrinsics are buggy with
1945 * LLVM 8, see r358579.
1946 */
1947 params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
1948 params[arg_count++] = ctx->ac.i32_0; /* soffset */
1949 params[arg_count++] = ctx->ac.i32_0; /* slc */
1950
1951 ac_build_type_name_for_intr(return_type, type, sizeof(type));
1952 snprintf(name, sizeof(name),
1953 "llvm.amdgcn.raw.buffer.atomic.%s.%s", op, type);
1954 } else {
1955 params[arg_count++] = ctx->ac.i32_0; /* vindex */
1956 params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
1957 params[arg_count++] = ctx->ac.i1false; /* slc */
1958
1959 assert(return_type == ctx->ac.i32);
1960 snprintf(name, sizeof(name),
1961 "llvm.amdgcn.buffer.atomic.%s", op);
1962 }
1963
1964 result = ac_build_intrinsic(&ctx->ac, name, return_type, params,
1965 arg_count, 0);
1966 }
1967
1968 result = exit_waterfall(ctx, &wctx, result);
1969 if (ctx->ac.postponed_kill)
1970 ac_build_endif(&ctx->ac, 7001);
1971 return result;
1972 }
1973
1974 static LLVMValueRef visit_load_buffer(struct ac_nir_context *ctx,
1975 nir_intrinsic_instr *instr)
1976 {
1977 struct waterfall_context wctx;
1978 LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[0]);
1979
1980 int elem_size_bytes = instr->dest.ssa.bit_size / 8;
1981 int num_components = instr->num_components;
1982 enum gl_access_qualifier access = nir_intrinsic_access(instr);
1983 unsigned cache_policy = get_cache_policy(ctx, access, false, false);
1984
1985 LLVMValueRef offset = get_src(ctx, instr->src[1]);
1986 LLVMValueRef rsrc = ctx->abi->load_ssbo(ctx->abi, rsrc_base, false);
1987 LLVMValueRef vindex = ctx->ac.i32_0;
1988
1989 LLVMTypeRef def_type = get_def_type(ctx, &instr->dest.ssa);
1990 LLVMTypeRef def_elem_type = num_components > 1 ? LLVMGetElementType(def_type) : def_type;
1991
1992 LLVMValueRef results[4];
1993 for (int i = 0; i < num_components;) {
1994 int num_elems = num_components - i;
1995 if (elem_size_bytes < 4 && nir_intrinsic_align(instr) % 4 != 0)
1996 num_elems = 1;
1997 if (num_elems * elem_size_bytes > 16)
1998 num_elems = 16 / elem_size_bytes;
1999 int load_bytes = num_elems * elem_size_bytes;
2000
2001 LLVMValueRef immoffset = LLVMConstInt(ctx->ac.i32, i * elem_size_bytes, false);
2002
2003 LLVMValueRef ret;
2004
2005 if (load_bytes == 1) {
2006 ret = ac_build_tbuffer_load_byte(&ctx->ac,
2007 rsrc,
2008 offset,
2009 ctx->ac.i32_0,
2010 immoffset,
2011 cache_policy);
2012 } else if (load_bytes == 2) {
2013 ret = ac_build_tbuffer_load_short(&ctx->ac,
2014 rsrc,
2015 offset,
2016 ctx->ac.i32_0,
2017 immoffset,
2018 cache_policy);
2019 } else {
2020 int num_channels = util_next_power_of_two(load_bytes) / 4;
2021 bool can_speculate = access & ACCESS_CAN_REORDER;
2022
2023 ret = ac_build_buffer_load(&ctx->ac, rsrc, num_channels,
2024 vindex, offset, immoffset, 0,
2025 cache_policy, can_speculate, false);
2026 }
2027
2028 LLVMTypeRef byte_vec = LLVMVectorType(ctx->ac.i8, ac_get_type_size(LLVMTypeOf(ret)));
2029 ret = LLVMBuildBitCast(ctx->ac.builder, ret, byte_vec, "");
2030 ret = ac_trim_vector(&ctx->ac, ret, load_bytes);
2031
2032 LLVMTypeRef ret_type = LLVMVectorType(def_elem_type, num_elems);
2033 ret = LLVMBuildBitCast(ctx->ac.builder, ret, ret_type, "");
2034
2035 for (unsigned j = 0; j < num_elems; j++) {
2036 results[i + j] = LLVMBuildExtractElement(ctx->ac.builder, ret, LLVMConstInt(ctx->ac.i32, j, false), "");
2037 }
2038 i += num_elems;
2039 }
2040
2041 LLVMValueRef ret = ac_build_gather_values(&ctx->ac, results, num_components);
2042 return exit_waterfall(ctx, &wctx, ret);
2043 }
2044
2045 static LLVMValueRef enter_waterfall_ubo(struct ac_nir_context *ctx,
2046 struct waterfall_context *wctx,
2047 const nir_intrinsic_instr *instr)
2048 {
2049 return enter_waterfall(ctx, wctx, get_src(ctx, instr->src[0]),
2050 nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
2051 }
2052
2053 static LLVMValueRef visit_load_ubo_buffer(struct ac_nir_context *ctx,
2054 nir_intrinsic_instr *instr)
2055 {
2056 struct waterfall_context wctx;
2057 LLVMValueRef rsrc_base = enter_waterfall_ubo(ctx, &wctx, instr);
2058
2059 LLVMValueRef ret;
2060 LLVMValueRef rsrc = rsrc_base;
2061 LLVMValueRef offset = get_src(ctx, instr->src[1]);
2062 int num_components = instr->num_components;
2063
2064 if (ctx->abi->load_ubo)
2065 rsrc = ctx->abi->load_ubo(ctx->abi, rsrc);
2066
2067 if (instr->dest.ssa.bit_size == 64)
2068 num_components *= 2;
2069
2070 if (instr->dest.ssa.bit_size == 16 || instr->dest.ssa.bit_size == 8) {
2071 unsigned load_bytes = instr->dest.ssa.bit_size / 8;
2072 LLVMValueRef results[num_components];
2073 for (unsigned i = 0; i < num_components; ++i) {
2074 LLVMValueRef immoffset = LLVMConstInt(ctx->ac.i32,
2075 load_bytes * i, 0);
2076
2077 if (load_bytes == 1) {
2078 results[i] = ac_build_tbuffer_load_byte(&ctx->ac,
2079 rsrc,
2080 offset,
2081 ctx->ac.i32_0,
2082 immoffset,
2083 0);
2084 } else {
2085 assert(load_bytes == 2);
2086 results[i] = ac_build_tbuffer_load_short(&ctx->ac,
2087 rsrc,
2088 offset,
2089 ctx->ac.i32_0,
2090 immoffset,
2091 0);
2092 }
2093 }
2094 ret = ac_build_gather_values(&ctx->ac, results, num_components);
2095 } else {
2096 ret = ac_build_buffer_load(&ctx->ac, rsrc, num_components, NULL, offset,
2097 NULL, 0, 0, true, true);
2098
2099 ret = ac_trim_vector(&ctx->ac, ret, num_components);
2100 }
2101
2102 ret = LLVMBuildBitCast(ctx->ac.builder, ret,
2103 get_def_type(ctx, &instr->dest.ssa), "");
2104
2105 return exit_waterfall(ctx, &wctx, ret);
2106 }
2107
2108 static void
2109 get_deref_offset(struct ac_nir_context *ctx, nir_deref_instr *instr,
2110 bool vs_in, unsigned *vertex_index_out,
2111 LLVMValueRef *vertex_index_ref,
2112 unsigned *const_out, LLVMValueRef *indir_out)
2113 {
2114 nir_variable *var = nir_deref_instr_get_variable(instr);
2115 nir_deref_path path;
2116 unsigned idx_lvl = 1;
2117
2118 nir_deref_path_init(&path, instr, NULL);
2119
2120 if (vertex_index_out != NULL || vertex_index_ref != NULL) {
2121 if (vertex_index_ref) {
2122 *vertex_index_ref = get_src(ctx, path.path[idx_lvl]->arr.index);
2123 if (vertex_index_out)
2124 *vertex_index_out = 0;
2125 } else {
2126 *vertex_index_out = nir_src_as_uint(path.path[idx_lvl]->arr.index);
2127 }
2128 ++idx_lvl;
2129 }
2130
2131 uint32_t const_offset = 0;
2132 LLVMValueRef offset = NULL;
2133
2134 if (var->data.compact) {
2135 assert(instr->deref_type == nir_deref_type_array);
2136 const_offset = nir_src_as_uint(instr->arr.index);
2137 goto out;
2138 }
2139
2140 for (; path.path[idx_lvl]; ++idx_lvl) {
2141 const struct glsl_type *parent_type = path.path[idx_lvl - 1]->type;
2142 if (path.path[idx_lvl]->deref_type == nir_deref_type_struct) {
2143 unsigned index = path.path[idx_lvl]->strct.index;
2144
2145 for (unsigned i = 0; i < index; i++) {
2146 const struct glsl_type *ft = glsl_get_struct_field(parent_type, i);
2147 const_offset += glsl_count_attribute_slots(ft, vs_in);
2148 }
2149 } else if(path.path[idx_lvl]->deref_type == nir_deref_type_array) {
2150 unsigned size = glsl_count_attribute_slots(path.path[idx_lvl]->type, vs_in);
2151 if (nir_src_is_const(path.path[idx_lvl]->arr.index)) {
2152 const_offset += size *
2153 nir_src_as_uint(path.path[idx_lvl]->arr.index);
2154 } else {
2155 LLVMValueRef array_off = LLVMBuildMul(ctx->ac.builder, LLVMConstInt(ctx->ac.i32, size, 0),
2156 get_src(ctx, path.path[idx_lvl]->arr.index), "");
2157 if (offset)
2158 offset = LLVMBuildAdd(ctx->ac.builder, offset, array_off, "");
2159 else
2160 offset = array_off;
2161 }
2162 } else
2163 unreachable("Uhandled deref type in get_deref_instr_offset");
2164 }
2165
2166 out:
2167 nir_deref_path_finish(&path);
2168
2169 if (const_offset && offset)
2170 offset = LLVMBuildAdd(ctx->ac.builder, offset,
2171 LLVMConstInt(ctx->ac.i32, const_offset, 0),
2172 "");
2173
2174 *const_out = const_offset;
2175 *indir_out = offset;
2176 }
2177
2178 static LLVMValueRef load_tess_varyings(struct ac_nir_context *ctx,
2179 nir_intrinsic_instr *instr,
2180 bool load_inputs)
2181 {
2182 LLVMValueRef result;
2183 LLVMValueRef vertex_index = NULL;
2184 LLVMValueRef indir_index = NULL;
2185 unsigned const_index = 0;
2186
2187 nir_variable *var = nir_deref_instr_get_variable(nir_instr_as_deref(instr->src[0].ssa->parent_instr));
2188
2189 unsigned location = var->data.location;
2190 unsigned driver_location = var->data.driver_location;
2191 const bool is_patch = var->data.patch ||
2192 var->data.location == VARYING_SLOT_TESS_LEVEL_INNER ||
2193 var->data.location == VARYING_SLOT_TESS_LEVEL_OUTER;
2194 const bool is_compact = var->data.compact;
2195
2196 get_deref_offset(ctx, nir_instr_as_deref(instr->src[0].ssa->parent_instr),
2197 false, NULL, is_patch ? NULL : &vertex_index,
2198 &const_index, &indir_index);
2199
2200 LLVMTypeRef dest_type = get_def_type(ctx, &instr->dest.ssa);
2201
2202 LLVMTypeRef src_component_type;
2203 if (LLVMGetTypeKind(dest_type) == LLVMVectorTypeKind)
2204 src_component_type = LLVMGetElementType(dest_type);
2205 else
2206 src_component_type = dest_type;
2207
2208 result = ctx->abi->load_tess_varyings(ctx->abi, src_component_type,
2209 vertex_index, indir_index,
2210 const_index, location, driver_location,
2211 var->data.location_frac,
2212 instr->num_components,
2213 is_patch, is_compact, load_inputs);
2214 if (instr->dest.ssa.bit_size == 16) {
2215 result = ac_to_integer(&ctx->ac, result);
2216 result = LLVMBuildTrunc(ctx->ac.builder, result, dest_type, "");
2217 }
2218 return LLVMBuildBitCast(ctx->ac.builder, result, dest_type, "");
2219 }
2220
2221 static unsigned
2222 type_scalar_size_bytes(const struct glsl_type *type)
2223 {
2224 assert(glsl_type_is_vector_or_scalar(type) ||
2225 glsl_type_is_matrix(type));
2226 return glsl_type_is_boolean(type) ? 4 : glsl_get_bit_size(type) / 8;
2227 }
2228
2229 static LLVMValueRef visit_load_var(struct ac_nir_context *ctx,
2230 nir_intrinsic_instr *instr)
2231 {
2232 nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2233 nir_variable *var = nir_deref_instr_get_variable(deref);
2234
2235 LLVMValueRef values[8];
2236 int idx = 0;
2237 int ve = instr->dest.ssa.num_components;
2238 unsigned comp = 0;
2239 LLVMValueRef indir_index;
2240 LLVMValueRef ret;
2241 unsigned const_index;
2242 unsigned stride = 4;
2243 int mode = deref->mode;
2244
2245 if (var) {
2246 bool vs_in = ctx->stage == MESA_SHADER_VERTEX &&
2247 var->data.mode == nir_var_shader_in;
2248 idx = var->data.driver_location;
2249 comp = var->data.location_frac;
2250 mode = var->data.mode;
2251
2252 get_deref_offset(ctx, deref, vs_in, NULL, NULL,
2253 &const_index, &indir_index);
2254
2255 if (var->data.compact) {
2256 stride = 1;
2257 const_index += comp;
2258 comp = 0;
2259 }
2260 }
2261
2262 if (instr->dest.ssa.bit_size == 64 &&
2263 (deref->mode == nir_var_shader_in ||
2264 deref->mode == nir_var_shader_out ||
2265 deref->mode == nir_var_function_temp))
2266 ve *= 2;
2267
2268 switch (mode) {
2269 case nir_var_shader_in:
2270 if (ctx->stage == MESA_SHADER_TESS_CTRL ||
2271 ctx->stage == MESA_SHADER_TESS_EVAL) {
2272 return load_tess_varyings(ctx, instr, true);
2273 }
2274
2275 if (ctx->stage == MESA_SHADER_GEOMETRY) {
2276 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
2277 LLVMValueRef indir_index;
2278 unsigned const_index, vertex_index;
2279 get_deref_offset(ctx, deref, false, &vertex_index, NULL,
2280 &const_index, &indir_index);
2281 assert(indir_index == NULL);
2282
2283 return ctx->abi->load_inputs(ctx->abi, var->data.location,
2284 var->data.driver_location,
2285 var->data.location_frac,
2286 instr->num_components, vertex_index, const_index, type);
2287 }
2288
2289 for (unsigned chan = comp; chan < ve + comp; chan++) {
2290 if (indir_index) {
2291 unsigned count = glsl_count_attribute_slots(
2292 var->type,
2293 ctx->stage == MESA_SHADER_VERTEX);
2294 count -= chan / 4;
2295 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2296 &ctx->ac, ctx->abi->inputs + idx + chan, count,
2297 stride, false, true);
2298
2299 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2300 tmp_vec,
2301 indir_index, "");
2302 } else
2303 values[chan] = ctx->abi->inputs[idx + chan + const_index * stride];
2304 }
2305 break;
2306 case nir_var_function_temp:
2307 for (unsigned chan = 0; chan < ve; chan++) {
2308 if (indir_index) {
2309 unsigned count = glsl_count_attribute_slots(
2310 var->type, false);
2311 count -= chan / 4;
2312 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2313 &ctx->ac, ctx->locals + idx + chan, count,
2314 stride, true, true);
2315
2316 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2317 tmp_vec,
2318 indir_index, "");
2319 } else {
2320 values[chan] = LLVMBuildLoad(ctx->ac.builder, ctx->locals[idx + chan + const_index * stride], "");
2321 }
2322 }
2323 break;
2324 case nir_var_shader_out:
2325 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
2326 return load_tess_varyings(ctx, instr, false);
2327 }
2328
2329 if (ctx->stage == MESA_SHADER_FRAGMENT &&
2330 var->data.fb_fetch_output &&
2331 ctx->abi->emit_fbfetch)
2332 return ctx->abi->emit_fbfetch(ctx->abi);
2333
2334 for (unsigned chan = comp; chan < ve + comp; chan++) {
2335 if (indir_index) {
2336 unsigned count = glsl_count_attribute_slots(
2337 var->type, false);
2338 count -= chan / 4;
2339 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2340 &ctx->ac, ctx->abi->outputs + idx + chan, count,
2341 stride, true, true);
2342
2343 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2344 tmp_vec,
2345 indir_index, "");
2346 } else {
2347 values[chan] = LLVMBuildLoad(ctx->ac.builder,
2348 ctx->abi->outputs[idx + chan + const_index * stride],
2349 "");
2350 }
2351 }
2352 break;
2353 case nir_var_mem_global: {
2354 LLVMValueRef address = get_src(ctx, instr->src[0]);
2355 LLVMTypeRef result_type = get_def_type(ctx, &instr->dest.ssa);
2356 unsigned explicit_stride = glsl_get_explicit_stride(deref->type);
2357 unsigned natural_stride = type_scalar_size_bytes(deref->type);
2358 unsigned stride = explicit_stride ? explicit_stride : natural_stride;
2359 int elem_size_bytes = ac_get_elem_bits(&ctx->ac, result_type) / 8;
2360 bool split_loads = ctx->ac.chip_class == GFX6 && elem_size_bytes < 4;
2361
2362 if (stride != natural_stride || split_loads) {
2363 if (LLVMGetTypeKind(result_type) == LLVMVectorTypeKind)
2364 result_type = LLVMGetElementType(result_type);
2365
2366 LLVMTypeRef ptr_type = LLVMPointerType(result_type,
2367 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2368 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2369
2370 for (unsigned i = 0; i < instr->dest.ssa.num_components; ++i) {
2371 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, i * stride / natural_stride, 0);
2372 values[i] = LLVMBuildLoad(ctx->ac.builder,
2373 ac_build_gep_ptr(&ctx->ac, address, offset), "");
2374 }
2375 return ac_build_gather_values(&ctx->ac, values, instr->dest.ssa.num_components);
2376 } else {
2377 LLVMTypeRef ptr_type = LLVMPointerType(result_type,
2378 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2379 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2380 LLVMValueRef val = LLVMBuildLoad(ctx->ac.builder, address, "");
2381 return val;
2382 }
2383 }
2384 default:
2385 unreachable("unhandle variable mode");
2386 }
2387 ret = ac_build_varying_gather_values(&ctx->ac, values, ve, comp);
2388 return LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->dest.ssa), "");
2389 }
2390
2391 static void
2392 visit_store_var(struct ac_nir_context *ctx,
2393 nir_intrinsic_instr *instr)
2394 {
2395 if (ctx->ac.postponed_kill) {
2396 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
2397 ctx->ac.postponed_kill, "");
2398 ac_build_ifcc(&ctx->ac, cond, 7002);
2399 }
2400
2401 nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2402 nir_variable *var = nir_deref_instr_get_variable(deref);
2403
2404 LLVMValueRef temp_ptr, value;
2405 int idx = 0;
2406 unsigned comp = 0;
2407 LLVMValueRef src = ac_to_float(&ctx->ac, get_src(ctx, instr->src[1]));
2408 int writemask = instr->const_index[0];
2409 LLVMValueRef indir_index;
2410 unsigned const_index;
2411
2412 if (var) {
2413 get_deref_offset(ctx, deref, false,
2414 NULL, NULL, &const_index, &indir_index);
2415 idx = var->data.driver_location;
2416 comp = var->data.location_frac;
2417
2418 if (var->data.compact) {
2419 const_index += comp;
2420 comp = 0;
2421 }
2422 }
2423
2424 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src)) == 64 &&
2425 (deref->mode == nir_var_shader_out ||
2426 deref->mode == nir_var_function_temp)) {
2427
2428 src = LLVMBuildBitCast(ctx->ac.builder, src,
2429 LLVMVectorType(ctx->ac.f32, ac_get_llvm_num_components(src) * 2),
2430 "");
2431
2432 writemask = widen_mask(writemask, 2);
2433 }
2434
2435 writemask = writemask << comp;
2436
2437 switch (deref->mode) {
2438 case nir_var_shader_out:
2439
2440 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
2441 LLVMValueRef vertex_index = NULL;
2442 LLVMValueRef indir_index = NULL;
2443 unsigned const_index = 0;
2444 const bool is_patch = var->data.patch ||
2445 var->data.location == VARYING_SLOT_TESS_LEVEL_INNER ||
2446 var->data.location == VARYING_SLOT_TESS_LEVEL_OUTER;
2447
2448 get_deref_offset(ctx, deref, false, NULL,
2449 is_patch ? NULL : &vertex_index,
2450 &const_index, &indir_index);
2451
2452 ctx->abi->store_tcs_outputs(ctx->abi, var,
2453 vertex_index, indir_index,
2454 const_index, src, writemask);
2455 break;
2456 }
2457
2458 for (unsigned chan = 0; chan < 8; chan++) {
2459 int stride = 4;
2460 if (!(writemask & (1 << chan)))
2461 continue;
2462
2463 value = ac_llvm_extract_elem(&ctx->ac, src, chan - comp);
2464
2465 if (var->data.compact)
2466 stride = 1;
2467 if (indir_index) {
2468 unsigned count = glsl_count_attribute_slots(
2469 var->type, false);
2470 count -= chan / 4;
2471 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2472 &ctx->ac, ctx->abi->outputs + idx + chan, count,
2473 stride, true, true);
2474
2475 tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
2476 value, indir_index, "");
2477 build_store_values_extended(&ctx->ac, ctx->abi->outputs + idx + chan,
2478 count, stride, tmp_vec);
2479
2480 } else {
2481 temp_ptr = ctx->abi->outputs[idx + chan + const_index * stride];
2482
2483 LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
2484 }
2485 }
2486 break;
2487 case nir_var_function_temp:
2488 for (unsigned chan = 0; chan < 8; chan++) {
2489 if (!(writemask & (1 << chan)))
2490 continue;
2491
2492 value = ac_llvm_extract_elem(&ctx->ac, src, chan);
2493 if (indir_index) {
2494 unsigned count = glsl_count_attribute_slots(
2495 var->type, false);
2496 count -= chan / 4;
2497 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2498 &ctx->ac, ctx->locals + idx + chan, count,
2499 4, true, true);
2500
2501 tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
2502 value, indir_index, "");
2503 build_store_values_extended(&ctx->ac, ctx->locals + idx + chan,
2504 count, 4, tmp_vec);
2505 } else {
2506 temp_ptr = ctx->locals[idx + chan + const_index * 4];
2507
2508 LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
2509 }
2510 }
2511 break;
2512
2513 case nir_var_mem_global: {
2514 int writemask = instr->const_index[0];
2515 LLVMValueRef address = get_src(ctx, instr->src[0]);
2516 LLVMValueRef val = get_src(ctx, instr->src[1]);
2517
2518 unsigned explicit_stride = glsl_get_explicit_stride(deref->type);
2519 unsigned natural_stride = type_scalar_size_bytes(deref->type);
2520 unsigned stride = explicit_stride ? explicit_stride : natural_stride;
2521 int elem_size_bytes = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(val)) / 8;
2522 bool split_stores = ctx->ac.chip_class == GFX6 && elem_size_bytes < 4;
2523
2524 LLVMTypeRef ptr_type = LLVMPointerType(LLVMTypeOf(val),
2525 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2526 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2527
2528 if (writemask == (1u << ac_get_llvm_num_components(val)) - 1 &&
2529 stride == natural_stride && !split_stores) {
2530 LLVMTypeRef ptr_type = LLVMPointerType(LLVMTypeOf(val),
2531 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2532 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2533
2534 val = LLVMBuildBitCast(ctx->ac.builder, val,
2535 LLVMGetElementType(LLVMTypeOf(address)), "");
2536 LLVMBuildStore(ctx->ac.builder, val, address);
2537 } else {
2538 LLVMTypeRef val_type = LLVMTypeOf(val);
2539 if (LLVMGetTypeKind(LLVMTypeOf(val)) == LLVMVectorTypeKind)
2540 val_type = LLVMGetElementType(val_type);
2541
2542 LLVMTypeRef ptr_type = LLVMPointerType(val_type,
2543 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2544 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2545 for (unsigned chan = 0; chan < 4; chan++) {
2546 if (!(writemask & (1 << chan)))
2547 continue;
2548
2549 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, chan * stride / natural_stride, 0);
2550
2551 LLVMValueRef ptr = ac_build_gep_ptr(&ctx->ac, address, offset);
2552 LLVMValueRef src = ac_llvm_extract_elem(&ctx->ac, val,
2553 chan);
2554 src = LLVMBuildBitCast(ctx->ac.builder, src,
2555 LLVMGetElementType(LLVMTypeOf(ptr)), "");
2556 LLVMBuildStore(ctx->ac.builder, src, ptr);
2557 }
2558 }
2559 break;
2560 }
2561 default:
2562 abort();
2563 break;
2564 }
2565
2566 if (ctx->ac.postponed_kill)
2567 ac_build_endif(&ctx->ac, 7002);
2568 }
2569
2570 static int image_type_to_components_count(enum glsl_sampler_dim dim, bool array)
2571 {
2572 switch (dim) {
2573 case GLSL_SAMPLER_DIM_BUF:
2574 return 1;
2575 case GLSL_SAMPLER_DIM_1D:
2576 return array ? 2 : 1;
2577 case GLSL_SAMPLER_DIM_2D:
2578 return array ? 3 : 2;
2579 case GLSL_SAMPLER_DIM_MS:
2580 return array ? 4 : 3;
2581 case GLSL_SAMPLER_DIM_3D:
2582 case GLSL_SAMPLER_DIM_CUBE:
2583 return 3;
2584 case GLSL_SAMPLER_DIM_RECT:
2585 case GLSL_SAMPLER_DIM_SUBPASS:
2586 return 2;
2587 case GLSL_SAMPLER_DIM_SUBPASS_MS:
2588 return 3;
2589 default:
2590 break;
2591 }
2592 return 0;
2593 }
2594
2595 static LLVMValueRef adjust_sample_index_using_fmask(struct ac_llvm_context *ctx,
2596 LLVMValueRef coord_x, LLVMValueRef coord_y,
2597 LLVMValueRef coord_z,
2598 LLVMValueRef sample_index,
2599 LLVMValueRef fmask_desc_ptr)
2600 {
2601 unsigned sample_chan = coord_z ? 3 : 2;
2602 LLVMValueRef addr[4] = {coord_x, coord_y, coord_z};
2603 addr[sample_chan] = sample_index;
2604
2605 ac_apply_fmask_to_sample(ctx, fmask_desc_ptr, addr, coord_z != NULL);
2606 return addr[sample_chan];
2607 }
2608
2609 static nir_deref_instr *get_image_deref(const nir_intrinsic_instr *instr)
2610 {
2611 assert(instr->src[0].is_ssa);
2612 return nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2613 }
2614
2615 static LLVMValueRef get_image_descriptor(struct ac_nir_context *ctx,
2616 const nir_intrinsic_instr *instr,
2617 LLVMValueRef dynamic_index,
2618 enum ac_descriptor_type desc_type,
2619 bool write)
2620 {
2621 nir_deref_instr *deref_instr =
2622 instr->src[0].ssa->parent_instr->type == nir_instr_type_deref ?
2623 nir_instr_as_deref(instr->src[0].ssa->parent_instr) : NULL;
2624
2625 return get_sampler_desc(ctx, deref_instr, desc_type, &instr->instr, dynamic_index, true, write);
2626 }
2627
2628 static void get_image_coords(struct ac_nir_context *ctx,
2629 const nir_intrinsic_instr *instr,
2630 LLVMValueRef dynamic_desc_index,
2631 struct ac_image_args *args,
2632 enum glsl_sampler_dim dim,
2633 bool is_array)
2634 {
2635 LLVMValueRef src0 = get_src(ctx, instr->src[1]);
2636 LLVMValueRef masks[] = {
2637 LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
2638 LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
2639 };
2640 LLVMValueRef sample_index = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
2641
2642 int count;
2643 ASSERTED bool add_frag_pos = (dim == GLSL_SAMPLER_DIM_SUBPASS ||
2644 dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2645 bool is_ms = (dim == GLSL_SAMPLER_DIM_MS ||
2646 dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2647 bool gfx9_1d = ctx->ac.chip_class == GFX9 && dim == GLSL_SAMPLER_DIM_1D;
2648 assert(!add_frag_pos && "Input attachments should be lowered by this point.");
2649 count = image_type_to_components_count(dim, is_array);
2650
2651 if (is_ms && (instr->intrinsic == nir_intrinsic_image_deref_load ||
2652 instr->intrinsic == nir_intrinsic_bindless_image_load)) {
2653 LLVMValueRef fmask_load_address[3];
2654
2655 fmask_load_address[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
2656 fmask_load_address[1] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[1], "");
2657 if (is_array)
2658 fmask_load_address[2] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[2], "");
2659 else
2660 fmask_load_address[2] = NULL;
2661
2662 sample_index = adjust_sample_index_using_fmask(&ctx->ac,
2663 fmask_load_address[0],
2664 fmask_load_address[1],
2665 fmask_load_address[2],
2666 sample_index,
2667 get_sampler_desc(ctx, nir_instr_as_deref(instr->src[0].ssa->parent_instr),
2668 AC_DESC_FMASK, &instr->instr, dynamic_desc_index, true, false));
2669 }
2670 if (count == 1 && !gfx9_1d) {
2671 if (instr->src[1].ssa->num_components)
2672 args->coords[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
2673 else
2674 args->coords[0] = src0;
2675 } else {
2676 int chan;
2677 if (is_ms)
2678 count--;
2679 for (chan = 0; chan < count; ++chan) {
2680 args->coords[chan] = ac_llvm_extract_elem(&ctx->ac, src0, chan);
2681 }
2682
2683 if (gfx9_1d) {
2684 if (is_array) {
2685 args->coords[2] = args->coords[1];
2686 args->coords[1] = ctx->ac.i32_0;
2687 } else
2688 args->coords[1] = ctx->ac.i32_0;
2689 count++;
2690 }
2691 if (ctx->ac.chip_class == GFX9 &&
2692 dim == GLSL_SAMPLER_DIM_2D &&
2693 !is_array) {
2694 /* The hw can't bind a slice of a 3D image as a 2D
2695 * image, because it ignores BASE_ARRAY if the target
2696 * is 3D. The workaround is to read BASE_ARRAY and set
2697 * it as the 3rd address operand for all 2D images.
2698 */
2699 LLVMValueRef first_layer, const5, mask;
2700
2701 const5 = LLVMConstInt(ctx->ac.i32, 5, 0);
2702 mask = LLVMConstInt(ctx->ac.i32, S_008F24_BASE_ARRAY(~0), 0);
2703 first_layer = LLVMBuildExtractElement(ctx->ac.builder, args->resource, const5, "");
2704 first_layer = LLVMBuildAnd(ctx->ac.builder, first_layer, mask, "");
2705
2706 args->coords[count] = first_layer;
2707 count++;
2708 }
2709
2710
2711 if (is_ms) {
2712 args->coords[count] = sample_index;
2713 count++;
2714 }
2715 }
2716 }
2717
2718 static LLVMValueRef get_image_buffer_descriptor(struct ac_nir_context *ctx,
2719 const nir_intrinsic_instr *instr,
2720 LLVMValueRef dynamic_index,
2721 bool write, bool atomic)
2722 {
2723 LLVMValueRef rsrc = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_BUFFER, write);
2724 if (ctx->ac.chip_class == GFX9 && LLVM_VERSION_MAJOR < 9 && atomic) {
2725 LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 2, 0), "");
2726 LLVMValueRef stride = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 1, 0), "");
2727 stride = LLVMBuildLShr(ctx->ac.builder, stride, LLVMConstInt(ctx->ac.i32, 16, 0), "");
2728
2729 LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->ac.builder,
2730 LLVMBuildICmp(ctx->ac.builder, LLVMIntUGT, elem_count, stride, ""),
2731 elem_count, stride, "");
2732
2733 rsrc = LLVMBuildInsertElement(ctx->ac.builder, rsrc, new_elem_count,
2734 LLVMConstInt(ctx->ac.i32, 2, 0), "");
2735 }
2736 return rsrc;
2737 }
2738
2739 static LLVMValueRef enter_waterfall_image(struct ac_nir_context *ctx,
2740 struct waterfall_context *wctx,
2741 const nir_intrinsic_instr *instr)
2742 {
2743 nir_deref_instr *deref_instr = NULL;
2744
2745 if (instr->src[0].ssa->parent_instr->type == nir_instr_type_deref)
2746 deref_instr = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2747
2748 LLVMValueRef value = get_sampler_desc_index(ctx, deref_instr, &instr->instr, true);
2749 return enter_waterfall(ctx, wctx, value, nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
2750 }
2751
2752 static LLVMValueRef visit_image_load(struct ac_nir_context *ctx,
2753 const nir_intrinsic_instr *instr,
2754 bool bindless)
2755 {
2756 LLVMValueRef res;
2757
2758 enum glsl_sampler_dim dim;
2759 enum gl_access_qualifier access;
2760 bool is_array;
2761 if (bindless) {
2762 dim = nir_intrinsic_image_dim(instr);
2763 access = nir_intrinsic_access(instr);
2764 is_array = nir_intrinsic_image_array(instr);
2765 } else {
2766 const nir_deref_instr *image_deref = get_image_deref(instr);
2767 const struct glsl_type *type = image_deref->type;
2768 const nir_variable *var = nir_deref_instr_get_variable(image_deref);
2769 dim = glsl_get_sampler_dim(type);
2770 access = var->data.access;
2771 is_array = glsl_sampler_type_is_array(type);
2772 }
2773
2774 struct waterfall_context wctx;
2775 LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
2776
2777 struct ac_image_args args = {};
2778
2779 args.cache_policy = get_cache_policy(ctx, access, false, false);
2780
2781 if (dim == GLSL_SAMPLER_DIM_BUF) {
2782 unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
2783 unsigned num_channels = util_last_bit(mask);
2784 LLVMValueRef rsrc, vindex;
2785
2786 rsrc = get_image_buffer_descriptor(ctx, instr, dynamic_index, false, false);
2787 vindex = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
2788 ctx->ac.i32_0, "");
2789
2790 assert(instr->dest.is_ssa);
2791 bool can_speculate = access & ACCESS_CAN_REORDER;
2792 res = ac_build_buffer_load_format(&ctx->ac, rsrc, vindex,
2793 ctx->ac.i32_0, num_channels,
2794 args.cache_policy,
2795 can_speculate,
2796 instr->dest.ssa.bit_size == 16);
2797 res = ac_build_expand_to_vec4(&ctx->ac, res, num_channels);
2798
2799 res = ac_trim_vector(&ctx->ac, res, instr->dest.ssa.num_components);
2800 res = ac_to_integer(&ctx->ac, res);
2801 } else {
2802 bool level_zero = nir_src_is_const(instr->src[3]) && nir_src_as_uint(instr->src[3]) == 0;
2803
2804 args.opcode = level_zero ? ac_image_load : ac_image_load_mip;
2805 args.resource = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_IMAGE, false);
2806 get_image_coords(ctx, instr, dynamic_index, &args, dim, is_array);
2807 args.dim = ac_get_image_dim(ctx->ac.chip_class, dim, is_array);
2808 if (!level_zero)
2809 args.lod = get_src(ctx, instr->src[3]);
2810 args.dmask = 15;
2811 args.attributes = AC_FUNC_ATTR_READONLY;
2812
2813 assert(instr->dest.is_ssa);
2814 args.d16 = instr->dest.ssa.bit_size == 16;
2815
2816 res = ac_build_image_opcode(&ctx->ac, &args);
2817 }
2818 return exit_waterfall(ctx, &wctx, res);
2819 }
2820
2821 static void visit_image_store(struct ac_nir_context *ctx,
2822 const nir_intrinsic_instr *instr,
2823 bool bindless)
2824 {
2825 if (ctx->ac.postponed_kill) {
2826 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
2827 ctx->ac.postponed_kill, "");
2828 ac_build_ifcc(&ctx->ac, cond, 7003);
2829 }
2830
2831 enum glsl_sampler_dim dim;
2832 enum gl_access_qualifier access;
2833 bool is_array;
2834
2835 if (bindless) {
2836 dim = nir_intrinsic_image_dim(instr);
2837 access = nir_intrinsic_access(instr);
2838 is_array = nir_intrinsic_image_array(instr);
2839 } else {
2840 const nir_deref_instr *image_deref = get_image_deref(instr);
2841 const struct glsl_type *type = image_deref->type;
2842 const nir_variable *var = nir_deref_instr_get_variable(image_deref);
2843 dim = glsl_get_sampler_dim(type);
2844 access = var->data.access;
2845 is_array = glsl_sampler_type_is_array(type);
2846 }
2847
2848 struct waterfall_context wctx;
2849 LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
2850
2851 bool writeonly_memory = access & ACCESS_NON_READABLE;
2852 struct ac_image_args args = {};
2853
2854 args.cache_policy = get_cache_policy(ctx, access, true, writeonly_memory);
2855
2856 if (dim == GLSL_SAMPLER_DIM_BUF) {
2857 LLVMValueRef rsrc = get_image_buffer_descriptor(ctx, instr, dynamic_index, true, false);
2858 LLVMValueRef src = ac_to_float(&ctx->ac, get_src(ctx, instr->src[3]));
2859 unsigned src_channels = ac_get_llvm_num_components(src);
2860 LLVMValueRef vindex;
2861
2862 if (src_channels == 3)
2863 src = ac_build_expand_to_vec4(&ctx->ac, src, 3);
2864
2865 vindex = LLVMBuildExtractElement(ctx->ac.builder,
2866 get_src(ctx, instr->src[1]),
2867 ctx->ac.i32_0, "");
2868
2869 ac_build_buffer_store_format(&ctx->ac, rsrc, src, vindex,
2870 ctx->ac.i32_0, args.cache_policy);
2871 } else {
2872 bool level_zero = nir_src_is_const(instr->src[4]) && nir_src_as_uint(instr->src[4]) == 0;
2873
2874 args.opcode = level_zero ? ac_image_store : ac_image_store_mip;
2875 args.data[0] = ac_to_float(&ctx->ac, get_src(ctx, instr->src[3]));
2876 args.resource = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_IMAGE, true);
2877 get_image_coords(ctx, instr, dynamic_index, &args, dim, is_array);
2878 args.dim = ac_get_image_dim(ctx->ac.chip_class, dim, is_array);
2879 if (!level_zero)
2880 args.lod = get_src(ctx, instr->src[4]);
2881 args.dmask = 15;
2882 args.d16 = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(args.data[0])) == 16;
2883
2884 ac_build_image_opcode(&ctx->ac, &args);
2885 }
2886
2887 exit_waterfall(ctx, &wctx, NULL);
2888 if (ctx->ac.postponed_kill)
2889 ac_build_endif(&ctx->ac, 7003);
2890 }
2891
2892 static LLVMValueRef visit_image_atomic(struct ac_nir_context *ctx,
2893 const nir_intrinsic_instr *instr,
2894 bool bindless)
2895 {
2896 if (ctx->ac.postponed_kill) {
2897 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
2898 ctx->ac.postponed_kill, "");
2899 ac_build_ifcc(&ctx->ac, cond, 7004);
2900 }
2901
2902 LLVMValueRef params[7];
2903 int param_count = 0;
2904
2905 bool cmpswap = instr->intrinsic == nir_intrinsic_image_deref_atomic_comp_swap ||
2906 instr->intrinsic == nir_intrinsic_bindless_image_atomic_comp_swap;
2907 const char *atomic_name;
2908 char intrinsic_name[64];
2909 enum ac_atomic_op atomic_subop;
2910 ASSERTED int length;
2911
2912 enum glsl_sampler_dim dim;
2913 bool is_array;
2914 if (bindless) {
2915 if (instr->intrinsic == nir_intrinsic_bindless_image_atomic_imin ||
2916 instr->intrinsic == nir_intrinsic_bindless_image_atomic_umin ||
2917 instr->intrinsic == nir_intrinsic_bindless_image_atomic_imax ||
2918 instr->intrinsic == nir_intrinsic_bindless_image_atomic_umax) {
2919 ASSERTED const GLenum format = nir_intrinsic_format(instr);
2920 assert(format == GL_R32UI || format == GL_R32I);
2921 }
2922 dim = nir_intrinsic_image_dim(instr);
2923 is_array = nir_intrinsic_image_array(instr);
2924 } else {
2925 const struct glsl_type *type = get_image_deref(instr)->type;
2926 dim = glsl_get_sampler_dim(type);
2927 is_array = glsl_sampler_type_is_array(type);
2928 }
2929
2930 struct waterfall_context wctx;
2931 LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
2932
2933 switch (instr->intrinsic) {
2934 case nir_intrinsic_bindless_image_atomic_add:
2935 case nir_intrinsic_image_deref_atomic_add:
2936 atomic_name = "add";
2937 atomic_subop = ac_atomic_add;
2938 break;
2939 case nir_intrinsic_bindless_image_atomic_imin:
2940 case nir_intrinsic_image_deref_atomic_imin:
2941 atomic_name = "smin";
2942 atomic_subop = ac_atomic_smin;
2943 break;
2944 case nir_intrinsic_bindless_image_atomic_umin:
2945 case nir_intrinsic_image_deref_atomic_umin:
2946 atomic_name = "umin";
2947 atomic_subop = ac_atomic_umin;
2948 break;
2949 case nir_intrinsic_bindless_image_atomic_imax:
2950 case nir_intrinsic_image_deref_atomic_imax:
2951 atomic_name = "smax";
2952 atomic_subop = ac_atomic_smax;
2953 break;
2954 case nir_intrinsic_bindless_image_atomic_umax:
2955 case nir_intrinsic_image_deref_atomic_umax:
2956 atomic_name = "umax";
2957 atomic_subop = ac_atomic_umax;
2958 break;
2959 case nir_intrinsic_bindless_image_atomic_and:
2960 case nir_intrinsic_image_deref_atomic_and:
2961 atomic_name = "and";
2962 atomic_subop = ac_atomic_and;
2963 break;
2964 case nir_intrinsic_bindless_image_atomic_or:
2965 case nir_intrinsic_image_deref_atomic_or:
2966 atomic_name = "or";
2967 atomic_subop = ac_atomic_or;
2968 break;
2969 case nir_intrinsic_bindless_image_atomic_xor:
2970 case nir_intrinsic_image_deref_atomic_xor:
2971 atomic_name = "xor";
2972 atomic_subop = ac_atomic_xor;
2973 break;
2974 case nir_intrinsic_bindless_image_atomic_exchange:
2975 case nir_intrinsic_image_deref_atomic_exchange:
2976 atomic_name = "swap";
2977 atomic_subop = ac_atomic_swap;
2978 break;
2979 case nir_intrinsic_bindless_image_atomic_comp_swap:
2980 case nir_intrinsic_image_deref_atomic_comp_swap:
2981 atomic_name = "cmpswap";
2982 atomic_subop = 0; /* not used */
2983 break;
2984 case nir_intrinsic_bindless_image_atomic_inc_wrap:
2985 case nir_intrinsic_image_deref_atomic_inc_wrap: {
2986 atomic_name = "inc";
2987 atomic_subop = ac_atomic_inc_wrap;
2988 /* ATOMIC_INC instruction does:
2989 * value = (value + 1) % (data + 1)
2990 * but we want:
2991 * value = (value + 1) % data
2992 * So replace 'data' by 'data - 1'.
2993 */
2994 ctx->ssa_defs[instr->src[3].ssa->index] =
2995 LLVMBuildSub(ctx->ac.builder,
2996 ctx->ssa_defs[instr->src[3].ssa->index],
2997 ctx->ac.i32_1, "");
2998 break;
2999 }
3000 case nir_intrinsic_bindless_image_atomic_dec_wrap:
3001 case nir_intrinsic_image_deref_atomic_dec_wrap:
3002 atomic_name = "dec";
3003 atomic_subop = ac_atomic_dec_wrap;
3004 break;
3005 default:
3006 abort();
3007 }
3008
3009 if (cmpswap)
3010 params[param_count++] = get_src(ctx, instr->src[4]);
3011 params[param_count++] = get_src(ctx, instr->src[3]);
3012
3013 LLVMValueRef result;
3014 if (dim == GLSL_SAMPLER_DIM_BUF) {
3015 params[param_count++] = get_image_buffer_descriptor(ctx, instr, dynamic_index, true, true);
3016 params[param_count++] = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
3017 ctx->ac.i32_0, ""); /* vindex */
3018 params[param_count++] = ctx->ac.i32_0; /* voffset */
3019 if (LLVM_VERSION_MAJOR >= 9) {
3020 /* XXX: The new raw/struct atomic intrinsics are buggy
3021 * with LLVM 8, see r358579.
3022 */
3023 params[param_count++] = ctx->ac.i32_0; /* soffset */
3024 params[param_count++] = ctx->ac.i32_0; /* slc */
3025
3026 length = snprintf(intrinsic_name, sizeof(intrinsic_name),
3027 "llvm.amdgcn.struct.buffer.atomic.%s.i32", atomic_name);
3028 } else {
3029 params[param_count++] = ctx->ac.i1false; /* slc */
3030
3031 length = snprintf(intrinsic_name, sizeof(intrinsic_name),
3032 "llvm.amdgcn.buffer.atomic.%s", atomic_name);
3033 }
3034
3035 assert(length < sizeof(intrinsic_name));
3036 result = ac_build_intrinsic(&ctx->ac, intrinsic_name, ctx->ac.i32,
3037 params, param_count, 0);
3038 } else {
3039 struct ac_image_args args = {};
3040 args.opcode = cmpswap ? ac_image_atomic_cmpswap : ac_image_atomic;
3041 args.atomic = atomic_subop;
3042 args.data[0] = params[0];
3043 if (cmpswap)
3044 args.data[1] = params[1];
3045 args.resource = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_IMAGE, true);
3046 get_image_coords(ctx, instr, dynamic_index, &args, dim, is_array);
3047 args.dim = ac_get_image_dim(ctx->ac.chip_class, dim, is_array);
3048
3049 result = ac_build_image_opcode(&ctx->ac, &args);
3050 }
3051
3052 result = exit_waterfall(ctx, &wctx, result);
3053 if (ctx->ac.postponed_kill)
3054 ac_build_endif(&ctx->ac, 7004);
3055 return result;
3056 }
3057
3058 static LLVMValueRef visit_image_samples(struct ac_nir_context *ctx,
3059 nir_intrinsic_instr *instr)
3060 {
3061 struct waterfall_context wctx;
3062 LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
3063 LLVMValueRef rsrc = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_IMAGE, false);
3064
3065 LLVMValueRef ret = ac_build_image_get_sample_count(&ctx->ac, rsrc);
3066
3067 return exit_waterfall(ctx, &wctx, ret);
3068 }
3069
3070 static LLVMValueRef visit_image_size(struct ac_nir_context *ctx,
3071 const nir_intrinsic_instr *instr,
3072 bool bindless)
3073 {
3074 LLVMValueRef res;
3075
3076 enum glsl_sampler_dim dim;
3077 bool is_array;
3078 if (bindless) {
3079 dim = nir_intrinsic_image_dim(instr);
3080 is_array = nir_intrinsic_image_array(instr);
3081 } else {
3082 const struct glsl_type *type = get_image_deref(instr)->type;
3083 dim = glsl_get_sampler_dim(type);
3084 is_array = glsl_sampler_type_is_array(type);
3085 }
3086
3087 struct waterfall_context wctx;
3088 LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
3089
3090 if (dim == GLSL_SAMPLER_DIM_BUF) {
3091 res = get_buffer_size(ctx, get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_BUFFER, false), true);
3092 } else {
3093
3094 struct ac_image_args args = { 0 };
3095
3096 args.dim = ac_get_image_dim(ctx->ac.chip_class, dim, is_array);
3097 args.dmask = 0xf;
3098 args.resource = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_IMAGE, false);
3099 args.opcode = ac_image_get_resinfo;
3100 args.lod = ctx->ac.i32_0;
3101 args.attributes = AC_FUNC_ATTR_READNONE;
3102
3103 res = ac_build_image_opcode(&ctx->ac, &args);
3104
3105 LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
3106
3107 if (dim == GLSL_SAMPLER_DIM_CUBE && is_array) {
3108 LLVMValueRef six = LLVMConstInt(ctx->ac.i32, 6, false);
3109 LLVMValueRef z = LLVMBuildExtractElement(ctx->ac.builder, res, two, "");
3110 z = LLVMBuildSDiv(ctx->ac.builder, z, six, "");
3111 res = LLVMBuildInsertElement(ctx->ac.builder, res, z, two, "");
3112 }
3113
3114 if (ctx->ac.chip_class == GFX9 && dim == GLSL_SAMPLER_DIM_1D && is_array) {
3115 LLVMValueRef layers = LLVMBuildExtractElement(ctx->ac.builder, res, two, "");
3116 res = LLVMBuildInsertElement(ctx->ac.builder, res, layers,
3117 ctx->ac.i32_1, "");
3118 }
3119 }
3120 return exit_waterfall(ctx, &wctx, res);
3121 }
3122
3123 static void emit_membar(struct ac_llvm_context *ac,
3124 const nir_intrinsic_instr *instr)
3125 {
3126 unsigned wait_flags = 0;
3127
3128 switch (instr->intrinsic) {
3129 case nir_intrinsic_memory_barrier:
3130 case nir_intrinsic_group_memory_barrier:
3131 wait_flags = AC_WAIT_LGKM | AC_WAIT_VLOAD | AC_WAIT_VSTORE;
3132 break;
3133 case nir_intrinsic_memory_barrier_buffer:
3134 case nir_intrinsic_memory_barrier_image:
3135 wait_flags = AC_WAIT_VLOAD | AC_WAIT_VSTORE;
3136 break;
3137 case nir_intrinsic_memory_barrier_shared:
3138 wait_flags = AC_WAIT_LGKM;
3139 break;
3140 default:
3141 break;
3142 }
3143
3144 ac_build_waitcnt(ac, wait_flags);
3145 }
3146
3147 void ac_emit_barrier(struct ac_llvm_context *ac, gl_shader_stage stage)
3148 {
3149 /* GFX6 only (thanks to a hw bug workaround):
3150 * The real barrier instruction isn’t needed, because an entire patch
3151 * always fits into a single wave.
3152 */
3153 if (ac->chip_class == GFX6 && stage == MESA_SHADER_TESS_CTRL) {
3154 ac_build_waitcnt(ac, AC_WAIT_LGKM | AC_WAIT_VLOAD | AC_WAIT_VSTORE);
3155 return;
3156 }
3157 ac_build_s_barrier(ac);
3158 }
3159
3160 static void emit_discard(struct ac_nir_context *ctx,
3161 const nir_intrinsic_instr *instr)
3162 {
3163 LLVMValueRef cond;
3164
3165 if (instr->intrinsic == nir_intrinsic_discard_if) {
3166 cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
3167 get_src(ctx, instr->src[0]),
3168 ctx->ac.i32_0, "");
3169 } else {
3170 assert(instr->intrinsic == nir_intrinsic_discard);
3171 cond = ctx->ac.i1false;
3172 }
3173
3174 ac_build_kill_if_false(&ctx->ac, cond);
3175 }
3176
3177 static void emit_demote(struct ac_nir_context *ctx,
3178 const nir_intrinsic_instr *instr)
3179 {
3180 LLVMValueRef cond;
3181
3182 if (instr->intrinsic == nir_intrinsic_demote_if) {
3183 cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
3184 get_src(ctx, instr->src[0]),
3185 ctx->ac.i32_0, "");
3186 } else {
3187 assert(instr->intrinsic == nir_intrinsic_demote);
3188 cond = ctx->ac.i1false;
3189 }
3190
3191 /* Kill immediately while maintaining WQM. */
3192 ac_build_kill_if_false(&ctx->ac, ac_build_wqm_vote(&ctx->ac, cond));
3193
3194 LLVMValueRef mask = LLVMBuildLoad(ctx->ac.builder, ctx->ac.postponed_kill, "");
3195 mask = LLVMBuildAnd(ctx->ac.builder, mask, cond, "");
3196 LLVMBuildStore(ctx->ac.builder, mask, ctx->ac.postponed_kill);
3197 return;
3198 }
3199
3200 static LLVMValueRef
3201 visit_load_local_invocation_index(struct ac_nir_context *ctx)
3202 {
3203 LLVMValueRef result;
3204 LLVMValueRef thread_id = ac_get_thread_id(&ctx->ac);
3205 result = LLVMBuildAnd(ctx->ac.builder,
3206 ac_get_arg(&ctx->ac, ctx->args->tg_size),
3207 LLVMConstInt(ctx->ac.i32, 0xfc0, false), "");
3208
3209 if (ctx->ac.wave_size == 32)
3210 result = LLVMBuildLShr(ctx->ac.builder, result,
3211 LLVMConstInt(ctx->ac.i32, 1, false), "");
3212
3213 return LLVMBuildAdd(ctx->ac.builder, result, thread_id, "");
3214 }
3215
3216 static LLVMValueRef
3217 visit_load_subgroup_id(struct ac_nir_context *ctx)
3218 {
3219 if (ctx->stage == MESA_SHADER_COMPUTE) {
3220 LLVMValueRef result;
3221 result = LLVMBuildAnd(ctx->ac.builder,
3222 ac_get_arg(&ctx->ac, ctx->args->tg_size),
3223 LLVMConstInt(ctx->ac.i32, 0xfc0, false), "");
3224 return LLVMBuildLShr(ctx->ac.builder, result, LLVMConstInt(ctx->ac.i32, 6, false), "");
3225 } else {
3226 return LLVMConstInt(ctx->ac.i32, 0, false);
3227 }
3228 }
3229
3230 static LLVMValueRef
3231 visit_load_num_subgroups(struct ac_nir_context *ctx)
3232 {
3233 if (ctx->stage == MESA_SHADER_COMPUTE) {
3234 return LLVMBuildAnd(ctx->ac.builder,
3235 ac_get_arg(&ctx->ac, ctx->args->tg_size),
3236 LLVMConstInt(ctx->ac.i32, 0x3f, false), "");
3237 } else {
3238 return LLVMConstInt(ctx->ac.i32, 1, false);
3239 }
3240 }
3241
3242 static LLVMValueRef
3243 visit_first_invocation(struct ac_nir_context *ctx)
3244 {
3245 LLVMValueRef active_set = ac_build_ballot(&ctx->ac, ctx->ac.i32_1);
3246 const char *intr = ctx->ac.wave_size == 32 ? "llvm.cttz.i32" : "llvm.cttz.i64";
3247
3248 /* The second argument is whether cttz(0) should be defined, but we do not care. */
3249 LLVMValueRef args[] = {active_set, ctx->ac.i1false};
3250 LLVMValueRef result = ac_build_intrinsic(&ctx->ac, intr,
3251 ctx->ac.iN_wavemask, args, 2,
3252 AC_FUNC_ATTR_NOUNWIND |
3253 AC_FUNC_ATTR_READNONE);
3254
3255 return LLVMBuildTrunc(ctx->ac.builder, result, ctx->ac.i32, "");
3256 }
3257
3258 static LLVMValueRef
3259 visit_load_shared(struct ac_nir_context *ctx,
3260 const nir_intrinsic_instr *instr)
3261 {
3262 LLVMValueRef values[4], derived_ptr, index, ret;
3263
3264 LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[0],
3265 instr->dest.ssa.bit_size);
3266
3267 for (int chan = 0; chan < instr->num_components; chan++) {
3268 index = LLVMConstInt(ctx->ac.i32, chan, 0);
3269 derived_ptr = LLVMBuildGEP(ctx->ac.builder, ptr, &index, 1, "");
3270 values[chan] = LLVMBuildLoad(ctx->ac.builder, derived_ptr, "");
3271 }
3272
3273 ret = ac_build_gather_values(&ctx->ac, values, instr->num_components);
3274 return LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->dest.ssa), "");
3275 }
3276
3277 static void
3278 visit_store_shared(struct ac_nir_context *ctx,
3279 const nir_intrinsic_instr *instr)
3280 {
3281 LLVMValueRef derived_ptr, data,index;
3282 LLVMBuilderRef builder = ctx->ac.builder;
3283
3284 LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[1],
3285 instr->src[0].ssa->bit_size);
3286 LLVMValueRef src = get_src(ctx, instr->src[0]);
3287
3288 int writemask = nir_intrinsic_write_mask(instr);
3289 for (int chan = 0; chan < 4; chan++) {
3290 if (!(writemask & (1 << chan))) {
3291 continue;
3292 }
3293 data = ac_llvm_extract_elem(&ctx->ac, src, chan);
3294 index = LLVMConstInt(ctx->ac.i32, chan, 0);
3295 derived_ptr = LLVMBuildGEP(builder, ptr, &index, 1, "");
3296 LLVMBuildStore(builder, data, derived_ptr);
3297 }
3298 }
3299
3300 static LLVMValueRef visit_var_atomic(struct ac_nir_context *ctx,
3301 const nir_intrinsic_instr *instr,
3302 LLVMValueRef ptr, int src_idx)
3303 {
3304 if (ctx->ac.postponed_kill) {
3305 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
3306 ctx->ac.postponed_kill, "");
3307 ac_build_ifcc(&ctx->ac, cond, 7005);
3308 }
3309
3310 LLVMValueRef result;
3311 LLVMValueRef src = get_src(ctx, instr->src[src_idx]);
3312
3313 const char *sync_scope = LLVM_VERSION_MAJOR >= 9 ? "workgroup-one-as" : "workgroup";
3314
3315 if (instr->src[0].ssa->parent_instr->type == nir_instr_type_deref) {
3316 nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
3317 if (deref->mode == nir_var_mem_global) {
3318 /* use "singlethread" sync scope to implement relaxed ordering */
3319 sync_scope = LLVM_VERSION_MAJOR >= 9 ? "singlethread-one-as" : "singlethread";
3320
3321 LLVMTypeRef ptr_type = LLVMPointerType(LLVMTypeOf(src), LLVMGetPointerAddressSpace(LLVMTypeOf(ptr)));
3322 ptr = LLVMBuildBitCast(ctx->ac.builder, ptr, ptr_type , "");
3323 }
3324 }
3325
3326 if (instr->intrinsic == nir_intrinsic_shared_atomic_comp_swap ||
3327 instr->intrinsic == nir_intrinsic_deref_atomic_comp_swap) {
3328 LLVMValueRef src1 = get_src(ctx, instr->src[src_idx + 1]);
3329 result = ac_build_atomic_cmp_xchg(&ctx->ac, ptr, src, src1, sync_scope);
3330 result = LLVMBuildExtractValue(ctx->ac.builder, result, 0, "");
3331 } else {
3332 LLVMAtomicRMWBinOp op;
3333 switch (instr->intrinsic) {
3334 case nir_intrinsic_shared_atomic_add:
3335 case nir_intrinsic_deref_atomic_add:
3336 op = LLVMAtomicRMWBinOpAdd;
3337 break;
3338 case nir_intrinsic_shared_atomic_umin:
3339 case nir_intrinsic_deref_atomic_umin:
3340 op = LLVMAtomicRMWBinOpUMin;
3341 break;
3342 case nir_intrinsic_shared_atomic_umax:
3343 case nir_intrinsic_deref_atomic_umax:
3344 op = LLVMAtomicRMWBinOpUMax;
3345 break;
3346 case nir_intrinsic_shared_atomic_imin:
3347 case nir_intrinsic_deref_atomic_imin:
3348 op = LLVMAtomicRMWBinOpMin;
3349 break;
3350 case nir_intrinsic_shared_atomic_imax:
3351 case nir_intrinsic_deref_atomic_imax:
3352 op = LLVMAtomicRMWBinOpMax;
3353 break;
3354 case nir_intrinsic_shared_atomic_and:
3355 case nir_intrinsic_deref_atomic_and:
3356 op = LLVMAtomicRMWBinOpAnd;
3357 break;
3358 case nir_intrinsic_shared_atomic_or:
3359 case nir_intrinsic_deref_atomic_or:
3360 op = LLVMAtomicRMWBinOpOr;
3361 break;
3362 case nir_intrinsic_shared_atomic_xor:
3363 case nir_intrinsic_deref_atomic_xor:
3364 op = LLVMAtomicRMWBinOpXor;
3365 break;
3366 case nir_intrinsic_shared_atomic_exchange:
3367 case nir_intrinsic_deref_atomic_exchange:
3368 op = LLVMAtomicRMWBinOpXchg;
3369 break;
3370 default:
3371 return NULL;
3372 }
3373
3374 result = ac_build_atomic_rmw(&ctx->ac, op, ptr, ac_to_integer(&ctx->ac, src), sync_scope);
3375 }
3376
3377 if (ctx->ac.postponed_kill)
3378 ac_build_endif(&ctx->ac, 7005);
3379 return result;
3380 }
3381
3382 static LLVMValueRef load_sample_pos(struct ac_nir_context *ctx)
3383 {
3384 LLVMValueRef values[2];
3385 LLVMValueRef pos[2];
3386
3387 pos[0] = ac_to_float(&ctx->ac,
3388 ac_get_arg(&ctx->ac, ctx->args->frag_pos[0]));
3389 pos[1] = ac_to_float(&ctx->ac,
3390 ac_get_arg(&ctx->ac, ctx->args->frag_pos[1]));
3391
3392 values[0] = ac_build_fract(&ctx->ac, pos[0], 32);
3393 values[1] = ac_build_fract(&ctx->ac, pos[1], 32);
3394 return ac_build_gather_values(&ctx->ac, values, 2);
3395 }
3396
3397 static LLVMValueRef lookup_interp_param(struct ac_nir_context *ctx,
3398 enum glsl_interp_mode interp, unsigned location)
3399 {
3400 switch (interp) {
3401 case INTERP_MODE_FLAT:
3402 default:
3403 return NULL;
3404 case INTERP_MODE_SMOOTH:
3405 case INTERP_MODE_NONE:
3406 if (location == INTERP_CENTER)
3407 return ac_get_arg(&ctx->ac, ctx->args->persp_center);
3408 else if (location == INTERP_CENTROID)
3409 return ctx->abi->persp_centroid;
3410 else if (location == INTERP_SAMPLE)
3411 return ac_get_arg(&ctx->ac, ctx->args->persp_sample);
3412 break;
3413 case INTERP_MODE_NOPERSPECTIVE:
3414 if (location == INTERP_CENTER)
3415 return ac_get_arg(&ctx->ac, ctx->args->linear_center);
3416 else if (location == INTERP_CENTROID)
3417 return ctx->abi->linear_centroid;
3418 else if (location == INTERP_SAMPLE)
3419 return ac_get_arg(&ctx->ac, ctx->args->linear_sample);
3420 break;
3421 }
3422 return NULL;
3423 }
3424
3425 static LLVMValueRef barycentric_center(struct ac_nir_context *ctx,
3426 unsigned mode)
3427 {
3428 LLVMValueRef interp_param = lookup_interp_param(ctx, mode, INTERP_CENTER);
3429 return LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2i32, "");
3430 }
3431
3432 static LLVMValueRef barycentric_offset(struct ac_nir_context *ctx,
3433 unsigned mode,
3434 LLVMValueRef offset)
3435 {
3436 LLVMValueRef interp_param = lookup_interp_param(ctx, mode, INTERP_CENTER);
3437 LLVMValueRef src_c0 = ac_to_float(&ctx->ac, LLVMBuildExtractElement(ctx->ac.builder, offset, ctx->ac.i32_0, ""));
3438 LLVMValueRef src_c1 = ac_to_float(&ctx->ac, LLVMBuildExtractElement(ctx->ac.builder, offset, ctx->ac.i32_1, ""));
3439
3440 LLVMValueRef ij_out[2];
3441 LLVMValueRef ddxy_out = ac_build_ddxy_interp(&ctx->ac, interp_param);
3442
3443 /*
3444 * take the I then J parameters, and the DDX/Y for it, and
3445 * calculate the IJ inputs for the interpolator.
3446 * temp1 = ddx * offset/sample.x + I;
3447 * interp_param.I = ddy * offset/sample.y + temp1;
3448 * temp1 = ddx * offset/sample.x + J;
3449 * interp_param.J = ddy * offset/sample.y + temp1;
3450 */
3451 for (unsigned i = 0; i < 2; i++) {
3452 LLVMValueRef ix_ll = LLVMConstInt(ctx->ac.i32, i, false);
3453 LLVMValueRef iy_ll = LLVMConstInt(ctx->ac.i32, i + 2, false);
3454 LLVMValueRef ddx_el = LLVMBuildExtractElement(ctx->ac.builder,
3455 ddxy_out, ix_ll, "");
3456 LLVMValueRef ddy_el = LLVMBuildExtractElement(ctx->ac.builder,
3457 ddxy_out, iy_ll, "");
3458 LLVMValueRef interp_el = LLVMBuildExtractElement(ctx->ac.builder,
3459 interp_param, ix_ll, "");
3460 LLVMValueRef temp1, temp2;
3461
3462 interp_el = LLVMBuildBitCast(ctx->ac.builder, interp_el,
3463 ctx->ac.f32, "");
3464
3465 temp1 = ac_build_fmad(&ctx->ac, ddx_el, src_c0, interp_el);
3466 temp2 = ac_build_fmad(&ctx->ac, ddy_el, src_c1, temp1);
3467
3468 ij_out[i] = LLVMBuildBitCast(ctx->ac.builder,
3469 temp2, ctx->ac.i32, "");
3470 }
3471 interp_param = ac_build_gather_values(&ctx->ac, ij_out, 2);
3472 return LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2i32, "");
3473 }
3474
3475 static LLVMValueRef barycentric_centroid(struct ac_nir_context *ctx,
3476 unsigned mode)
3477 {
3478 LLVMValueRef interp_param = lookup_interp_param(ctx, mode, INTERP_CENTROID);
3479 return LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2i32, "");
3480 }
3481
3482 static LLVMValueRef barycentric_at_sample(struct ac_nir_context *ctx,
3483 unsigned mode,
3484 LLVMValueRef sample_id)
3485 {
3486 if (ctx->abi->interp_at_sample_force_center)
3487 return barycentric_center(ctx, mode);
3488
3489 LLVMValueRef halfval = LLVMConstReal(ctx->ac.f32, 0.5f);
3490
3491 /* fetch sample ID */
3492 LLVMValueRef sample_pos = ctx->abi->load_sample_position(ctx->abi, sample_id);
3493
3494 LLVMValueRef src_c0 = LLVMBuildExtractElement(ctx->ac.builder, sample_pos, ctx->ac.i32_0, "");
3495 src_c0 = LLVMBuildFSub(ctx->ac.builder, src_c0, halfval, "");
3496 LLVMValueRef src_c1 = LLVMBuildExtractElement(ctx->ac.builder, sample_pos, ctx->ac.i32_1, "");
3497 src_c1 = LLVMBuildFSub(ctx->ac.builder, src_c1, halfval, "");
3498 LLVMValueRef coords[] = { src_c0, src_c1 };
3499 LLVMValueRef offset = ac_build_gather_values(&ctx->ac, coords, 2);
3500
3501 return barycentric_offset(ctx, mode, offset);
3502 }
3503
3504
3505 static LLVMValueRef barycentric_sample(struct ac_nir_context *ctx,
3506 unsigned mode)
3507 {
3508 LLVMValueRef interp_param = lookup_interp_param(ctx, mode, INTERP_SAMPLE);
3509 return LLVMBuildBitCast(ctx->ac.builder, interp_param, ctx->ac.v2i32, "");
3510 }
3511
3512 static LLVMValueRef barycentric_model(struct ac_nir_context *ctx)
3513 {
3514 return LLVMBuildBitCast(ctx->ac.builder,
3515 ac_get_arg(&ctx->ac, ctx->args->pull_model),
3516 ctx->ac.v3i32, "");
3517 }
3518
3519 static LLVMValueRef load_interpolated_input(struct ac_nir_context *ctx,
3520 LLVMValueRef interp_param,
3521 unsigned index, unsigned comp_start,
3522 unsigned num_components,
3523 unsigned bitsize)
3524 {
3525 LLVMValueRef attr_number = LLVMConstInt(ctx->ac.i32, index, false);
3526 LLVMValueRef interp_param_f;
3527
3528 interp_param_f = LLVMBuildBitCast(ctx->ac.builder,
3529 interp_param, ctx->ac.v2f32, "");
3530 LLVMValueRef i = LLVMBuildExtractElement(
3531 ctx->ac.builder, interp_param_f, ctx->ac.i32_0, "");
3532 LLVMValueRef j = LLVMBuildExtractElement(
3533 ctx->ac.builder, interp_param_f, ctx->ac.i32_1, "");
3534
3535 /* Workaround for issue 2647: kill threads with infinite interpolation coeffs */
3536 if (ctx->verified_interp &&
3537 !_mesa_hash_table_search(ctx->verified_interp, interp_param)) {
3538 LLVMValueRef args[2];
3539 args[0] = i;
3540 args[1] = LLVMConstInt(ctx->ac.i32, S_NAN | Q_NAN | N_INFINITY | P_INFINITY, false);
3541 LLVMValueRef cond = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.class.f32", ctx->ac.i1,
3542 args, 2, AC_FUNC_ATTR_READNONE);
3543 ac_build_kill_if_false(&ctx->ac, LLVMBuildNot(ctx->ac.builder, cond, ""));
3544 _mesa_hash_table_insert(ctx->verified_interp, interp_param, interp_param);
3545 }
3546
3547 LLVMValueRef values[4];
3548 assert(bitsize == 16 || bitsize == 32);
3549 for (unsigned comp = 0; comp < num_components; comp++) {
3550 LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, comp_start + comp, false);
3551 if (bitsize == 16) {
3552 values[comp] = ac_build_fs_interp_f16(&ctx->ac, llvm_chan, attr_number,
3553 ac_get_arg(&ctx->ac, ctx->args->prim_mask), i, j);
3554 } else {
3555 values[comp] = ac_build_fs_interp(&ctx->ac, llvm_chan, attr_number,
3556 ac_get_arg(&ctx->ac, ctx->args->prim_mask), i, j);
3557 }
3558 }
3559
3560 return ac_to_integer(&ctx->ac, ac_build_gather_values(&ctx->ac, values, num_components));
3561 }
3562
3563 static LLVMValueRef load_input(struct ac_nir_context *ctx,
3564 nir_intrinsic_instr *instr)
3565 {
3566 unsigned offset_idx = instr->intrinsic == nir_intrinsic_load_input ? 0 : 1;
3567
3568 /* We only lower inputs for fragment shaders ATM */
3569 ASSERTED nir_const_value *offset = nir_src_as_const_value(instr->src[offset_idx]);
3570 assert(offset);
3571 assert(offset[0].i32 == 0);
3572
3573 unsigned component = nir_intrinsic_component(instr);
3574 unsigned index = nir_intrinsic_base(instr);
3575 unsigned vertex_id = 2; /* P0 */
3576
3577 if (instr->intrinsic == nir_intrinsic_load_input_vertex) {
3578 nir_const_value *src0 = nir_src_as_const_value(instr->src[0]);
3579
3580 switch (src0[0].i32) {
3581 case 0:
3582 vertex_id = 2;
3583 break;
3584 case 1:
3585 vertex_id = 0;
3586 break;
3587 case 2:
3588 vertex_id = 1;
3589 break;
3590 default:
3591 unreachable("Invalid vertex index");
3592 }
3593 }
3594
3595 LLVMValueRef attr_number = LLVMConstInt(ctx->ac.i32, index, false);
3596 LLVMValueRef values[8];
3597
3598 /* Each component of a 64-bit value takes up two GL-level channels. */
3599 unsigned num_components = instr->dest.ssa.num_components;
3600 unsigned bit_size = instr->dest.ssa.bit_size;
3601 unsigned channels =
3602 bit_size == 64 ? num_components * 2 : num_components;
3603
3604 for (unsigned chan = 0; chan < channels; chan++) {
3605 if (component + chan > 4)
3606 attr_number = LLVMConstInt(ctx->ac.i32, index + 1, false);
3607 LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, (component + chan) % 4, false);
3608 values[chan] = ac_build_fs_interp_mov(&ctx->ac,
3609 LLVMConstInt(ctx->ac.i32, vertex_id, false),
3610 llvm_chan,
3611 attr_number,
3612 ac_get_arg(&ctx->ac, ctx->args->prim_mask));
3613 values[chan] = LLVMBuildBitCast(ctx->ac.builder, values[chan], ctx->ac.i32, "");
3614 values[chan] = LLVMBuildTruncOrBitCast(ctx->ac.builder, values[chan],
3615 bit_size == 16 ? ctx->ac.i16 : ctx->ac.i32, "");
3616 }
3617
3618 LLVMValueRef result = ac_build_gather_values(&ctx->ac, values, channels);
3619 if (bit_size == 64) {
3620 LLVMTypeRef type = num_components == 1 ? ctx->ac.i64 :
3621 LLVMVectorType(ctx->ac.i64, num_components);
3622 result = LLVMBuildBitCast(ctx->ac.builder, result, type, "");
3623 }
3624 return result;
3625 }
3626
3627 static void visit_intrinsic(struct ac_nir_context *ctx,
3628 nir_intrinsic_instr *instr)
3629 {
3630 LLVMValueRef result = NULL;
3631
3632 switch (instr->intrinsic) {
3633 case nir_intrinsic_ballot:
3634 result = ac_build_ballot(&ctx->ac, get_src(ctx, instr->src[0]));
3635 if (ctx->ac.ballot_mask_bits > ctx->ac.wave_size)
3636 result = LLVMBuildZExt(ctx->ac.builder, result, ctx->ac.iN_ballotmask, "");
3637 break;
3638 case nir_intrinsic_read_invocation:
3639 result = ac_build_readlane(&ctx->ac, get_src(ctx, instr->src[0]),
3640 get_src(ctx, instr->src[1]));
3641 break;
3642 case nir_intrinsic_read_first_invocation:
3643 result = ac_build_readlane(&ctx->ac, get_src(ctx, instr->src[0]), NULL);
3644 break;
3645 case nir_intrinsic_load_subgroup_invocation:
3646 result = ac_get_thread_id(&ctx->ac);
3647 break;
3648 case nir_intrinsic_load_work_group_id: {
3649 LLVMValueRef values[3];
3650
3651 for (int i = 0; i < 3; i++) {
3652 values[i] = ctx->args->workgroup_ids[i].used ?
3653 ac_get_arg(&ctx->ac, ctx->args->workgroup_ids[i]) : ctx->ac.i32_0;
3654 }
3655
3656 result = ac_build_gather_values(&ctx->ac, values, 3);
3657 break;
3658 }
3659 case nir_intrinsic_load_base_vertex:
3660 case nir_intrinsic_load_first_vertex:
3661 result = ctx->abi->load_base_vertex(ctx->abi);
3662 break;
3663 case nir_intrinsic_load_local_group_size:
3664 result = ctx->abi->load_local_group_size(ctx->abi);
3665 break;
3666 case nir_intrinsic_load_vertex_id:
3667 result = LLVMBuildAdd(ctx->ac.builder,
3668 ac_get_arg(&ctx->ac, ctx->args->vertex_id),
3669 ac_get_arg(&ctx->ac, ctx->args->base_vertex), "");
3670 break;
3671 case nir_intrinsic_load_vertex_id_zero_base: {
3672 result = ctx->abi->vertex_id;
3673 break;
3674 }
3675 case nir_intrinsic_load_local_invocation_id: {
3676 result = ac_get_arg(&ctx->ac, ctx->args->local_invocation_ids);
3677 break;
3678 }
3679 case nir_intrinsic_load_base_instance:
3680 result = ac_get_arg(&ctx->ac, ctx->args->start_instance);
3681 break;
3682 case nir_intrinsic_load_draw_id:
3683 result = ac_get_arg(&ctx->ac, ctx->args->draw_id);
3684 break;
3685 case nir_intrinsic_load_view_index:
3686 result = ac_get_arg(&ctx->ac, ctx->args->view_index);
3687 break;
3688 case nir_intrinsic_load_invocation_id:
3689 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
3690 result = ac_unpack_param(&ctx->ac,
3691 ac_get_arg(&ctx->ac, ctx->args->tcs_rel_ids),
3692 8, 5);
3693 } else {
3694 if (ctx->ac.chip_class >= GFX10) {
3695 result = LLVMBuildAnd(ctx->ac.builder,
3696 ac_get_arg(&ctx->ac, ctx->args->gs_invocation_id),
3697 LLVMConstInt(ctx->ac.i32, 127, 0), "");
3698 } else {
3699 result = ac_get_arg(&ctx->ac, ctx->args->gs_invocation_id);
3700 }
3701 }
3702 break;
3703 case nir_intrinsic_load_primitive_id:
3704 if (ctx->stage == MESA_SHADER_GEOMETRY) {
3705 result = ac_get_arg(&ctx->ac, ctx->args->gs_prim_id);
3706 } else if (ctx->stage == MESA_SHADER_TESS_CTRL) {
3707 result = ac_get_arg(&ctx->ac, ctx->args->tcs_patch_id);
3708 } else if (ctx->stage == MESA_SHADER_TESS_EVAL) {
3709 result = ac_get_arg(&ctx->ac, ctx->args->tes_patch_id);
3710 } else
3711 fprintf(stderr, "Unknown primitive id intrinsic: %d", ctx->stage);
3712 break;
3713 case nir_intrinsic_load_sample_id:
3714 result = ac_unpack_param(&ctx->ac,
3715 ac_get_arg(&ctx->ac, ctx->args->ancillary),
3716 8, 4);
3717 break;
3718 case nir_intrinsic_load_sample_pos:
3719 result = load_sample_pos(ctx);
3720 break;
3721 case nir_intrinsic_load_sample_mask_in:
3722 result = ctx->abi->load_sample_mask_in(ctx->abi);
3723 break;
3724 case nir_intrinsic_load_frag_coord: {
3725 LLVMValueRef values[4] = {
3726 ac_get_arg(&ctx->ac, ctx->args->frag_pos[0]),
3727 ac_get_arg(&ctx->ac, ctx->args->frag_pos[1]),
3728 ac_get_arg(&ctx->ac, ctx->args->frag_pos[2]),
3729 ac_build_fdiv(&ctx->ac, ctx->ac.f32_1,
3730 ac_get_arg(&ctx->ac, ctx->args->frag_pos[3]))
3731 };
3732 result = ac_to_integer(&ctx->ac,
3733 ac_build_gather_values(&ctx->ac, values, 4));
3734 break;
3735 }
3736 case nir_intrinsic_load_layer_id:
3737 result = ctx->abi->inputs[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER, 0)];
3738 break;
3739 case nir_intrinsic_load_front_face:
3740 result = ac_get_arg(&ctx->ac, ctx->args->front_face);
3741 break;
3742 case nir_intrinsic_load_helper_invocation:
3743 result = ac_build_load_helper_invocation(&ctx->ac);
3744 break;
3745 case nir_intrinsic_is_helper_invocation:
3746 result = ac_build_is_helper_invocation(&ctx->ac);
3747 break;
3748 case nir_intrinsic_load_color0:
3749 result = ctx->abi->color0;
3750 break;
3751 case nir_intrinsic_load_color1:
3752 result = ctx->abi->color1;
3753 break;
3754 case nir_intrinsic_load_user_data_amd:
3755 assert(LLVMTypeOf(ctx->abi->user_data) == ctx->ac.v4i32);
3756 result = ctx->abi->user_data;
3757 break;
3758 case nir_intrinsic_load_instance_id:
3759 result = ctx->abi->instance_id;
3760 break;
3761 case nir_intrinsic_load_num_work_groups:
3762 result = ac_get_arg(&ctx->ac, ctx->args->num_work_groups);
3763 break;
3764 case nir_intrinsic_load_local_invocation_index:
3765 result = visit_load_local_invocation_index(ctx);
3766 break;
3767 case nir_intrinsic_load_subgroup_id:
3768 result = visit_load_subgroup_id(ctx);
3769 break;
3770 case nir_intrinsic_load_num_subgroups:
3771 result = visit_load_num_subgroups(ctx);
3772 break;
3773 case nir_intrinsic_first_invocation:
3774 result = visit_first_invocation(ctx);
3775 break;
3776 case nir_intrinsic_load_push_constant:
3777 result = visit_load_push_constant(ctx, instr);
3778 break;
3779 case nir_intrinsic_vulkan_resource_index: {
3780 LLVMValueRef index = get_src(ctx, instr->src[0]);
3781 unsigned desc_set = nir_intrinsic_desc_set(instr);
3782 unsigned binding = nir_intrinsic_binding(instr);
3783
3784 result = ctx->abi->load_resource(ctx->abi, index, desc_set,
3785 binding);
3786 break;
3787 }
3788 case nir_intrinsic_vulkan_resource_reindex:
3789 result = visit_vulkan_resource_reindex(ctx, instr);
3790 break;
3791 case nir_intrinsic_store_ssbo:
3792 visit_store_ssbo(ctx, instr);
3793 break;
3794 case nir_intrinsic_load_ssbo:
3795 result = visit_load_buffer(ctx, instr);
3796 break;
3797 case nir_intrinsic_ssbo_atomic_add:
3798 case nir_intrinsic_ssbo_atomic_imin:
3799 case nir_intrinsic_ssbo_atomic_umin:
3800 case nir_intrinsic_ssbo_atomic_imax:
3801 case nir_intrinsic_ssbo_atomic_umax:
3802 case nir_intrinsic_ssbo_atomic_and:
3803 case nir_intrinsic_ssbo_atomic_or:
3804 case nir_intrinsic_ssbo_atomic_xor:
3805 case nir_intrinsic_ssbo_atomic_exchange:
3806 case nir_intrinsic_ssbo_atomic_comp_swap:
3807 result = visit_atomic_ssbo(ctx, instr);
3808 break;
3809 case nir_intrinsic_load_ubo:
3810 result = visit_load_ubo_buffer(ctx, instr);
3811 break;
3812 case nir_intrinsic_get_buffer_size:
3813 result = visit_get_buffer_size(ctx, instr);
3814 break;
3815 case nir_intrinsic_load_deref:
3816 result = visit_load_var(ctx, instr);
3817 break;
3818 case nir_intrinsic_store_deref:
3819 visit_store_var(ctx, instr);
3820 break;
3821 case nir_intrinsic_load_shared:
3822 result = visit_load_shared(ctx, instr);
3823 break;
3824 case nir_intrinsic_store_shared:
3825 visit_store_shared(ctx, instr);
3826 break;
3827 case nir_intrinsic_bindless_image_samples:
3828 case nir_intrinsic_image_deref_samples:
3829 result = visit_image_samples(ctx, instr);
3830 break;
3831 case nir_intrinsic_bindless_image_load:
3832 result = visit_image_load(ctx, instr, true);
3833 break;
3834 case nir_intrinsic_image_deref_load:
3835 result = visit_image_load(ctx, instr, false);
3836 break;
3837 case nir_intrinsic_bindless_image_store:
3838 visit_image_store(ctx, instr, true);
3839 break;
3840 case nir_intrinsic_image_deref_store:
3841 visit_image_store(ctx, instr, false);
3842 break;
3843 case nir_intrinsic_bindless_image_atomic_add:
3844 case nir_intrinsic_bindless_image_atomic_imin:
3845 case nir_intrinsic_bindless_image_atomic_umin:
3846 case nir_intrinsic_bindless_image_atomic_imax:
3847 case nir_intrinsic_bindless_image_atomic_umax:
3848 case nir_intrinsic_bindless_image_atomic_and:
3849 case nir_intrinsic_bindless_image_atomic_or:
3850 case nir_intrinsic_bindless_image_atomic_xor:
3851 case nir_intrinsic_bindless_image_atomic_exchange:
3852 case nir_intrinsic_bindless_image_atomic_comp_swap:
3853 case nir_intrinsic_bindless_image_atomic_inc_wrap:
3854 case nir_intrinsic_bindless_image_atomic_dec_wrap:
3855 result = visit_image_atomic(ctx, instr, true);
3856 break;
3857 case nir_intrinsic_image_deref_atomic_add:
3858 case nir_intrinsic_image_deref_atomic_imin:
3859 case nir_intrinsic_image_deref_atomic_umin:
3860 case nir_intrinsic_image_deref_atomic_imax:
3861 case nir_intrinsic_image_deref_atomic_umax:
3862 case nir_intrinsic_image_deref_atomic_and:
3863 case nir_intrinsic_image_deref_atomic_or:
3864 case nir_intrinsic_image_deref_atomic_xor:
3865 case nir_intrinsic_image_deref_atomic_exchange:
3866 case nir_intrinsic_image_deref_atomic_comp_swap:
3867 case nir_intrinsic_image_deref_atomic_inc_wrap:
3868 case nir_intrinsic_image_deref_atomic_dec_wrap:
3869 result = visit_image_atomic(ctx, instr, false);
3870 break;
3871 case nir_intrinsic_bindless_image_size:
3872 result = visit_image_size(ctx, instr, true);
3873 break;
3874 case nir_intrinsic_image_deref_size:
3875 result = visit_image_size(ctx, instr, false);
3876 break;
3877 case nir_intrinsic_shader_clock:
3878 result = ac_build_shader_clock(&ctx->ac,
3879 nir_intrinsic_memory_scope(instr));
3880 break;
3881 case nir_intrinsic_discard:
3882 case nir_intrinsic_discard_if:
3883 emit_discard(ctx, instr);
3884 break;
3885 case nir_intrinsic_demote:
3886 case nir_intrinsic_demote_if:
3887 emit_demote(ctx, instr);
3888 break;
3889 case nir_intrinsic_memory_barrier:
3890 case nir_intrinsic_group_memory_barrier:
3891 case nir_intrinsic_memory_barrier_buffer:
3892 case nir_intrinsic_memory_barrier_image:
3893 case nir_intrinsic_memory_barrier_shared:
3894 emit_membar(&ctx->ac, instr);
3895 break;
3896 case nir_intrinsic_memory_barrier_tcs_patch:
3897 break;
3898 case nir_intrinsic_control_barrier:
3899 ac_emit_barrier(&ctx->ac, ctx->stage);
3900 break;
3901 case nir_intrinsic_shared_atomic_add:
3902 case nir_intrinsic_shared_atomic_imin:
3903 case nir_intrinsic_shared_atomic_umin:
3904 case nir_intrinsic_shared_atomic_imax:
3905 case nir_intrinsic_shared_atomic_umax:
3906 case nir_intrinsic_shared_atomic_and:
3907 case nir_intrinsic_shared_atomic_or:
3908 case nir_intrinsic_shared_atomic_xor:
3909 case nir_intrinsic_shared_atomic_exchange:
3910 case nir_intrinsic_shared_atomic_comp_swap: {
3911 LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[0],
3912 instr->src[1].ssa->bit_size);
3913 result = visit_var_atomic(ctx, instr, ptr, 1);
3914 break;
3915 }
3916 case nir_intrinsic_deref_atomic_add:
3917 case nir_intrinsic_deref_atomic_imin:
3918 case nir_intrinsic_deref_atomic_umin:
3919 case nir_intrinsic_deref_atomic_imax:
3920 case nir_intrinsic_deref_atomic_umax:
3921 case nir_intrinsic_deref_atomic_and:
3922 case nir_intrinsic_deref_atomic_or:
3923 case nir_intrinsic_deref_atomic_xor:
3924 case nir_intrinsic_deref_atomic_exchange:
3925 case nir_intrinsic_deref_atomic_comp_swap: {
3926 LLVMValueRef ptr = get_src(ctx, instr->src[0]);
3927 result = visit_var_atomic(ctx, instr, ptr, 1);
3928 break;
3929 }
3930 case nir_intrinsic_load_barycentric_pixel:
3931 result = barycentric_center(ctx, nir_intrinsic_interp_mode(instr));
3932 break;
3933 case nir_intrinsic_load_barycentric_centroid:
3934 result = barycentric_centroid(ctx, nir_intrinsic_interp_mode(instr));
3935 break;
3936 case nir_intrinsic_load_barycentric_sample:
3937 result = barycentric_sample(ctx, nir_intrinsic_interp_mode(instr));
3938 break;
3939 case nir_intrinsic_load_barycentric_model:
3940 result = barycentric_model(ctx);
3941 break;
3942 case nir_intrinsic_load_barycentric_at_offset: {
3943 LLVMValueRef offset = ac_to_float(&ctx->ac, get_src(ctx, instr->src[0]));
3944 result = barycentric_offset(ctx, nir_intrinsic_interp_mode(instr), offset);
3945 break;
3946 }
3947 case nir_intrinsic_load_barycentric_at_sample: {
3948 LLVMValueRef sample_id = get_src(ctx, instr->src[0]);
3949 result = barycentric_at_sample(ctx, nir_intrinsic_interp_mode(instr), sample_id);
3950 break;
3951 }
3952 case nir_intrinsic_load_interpolated_input: {
3953 /* We assume any indirect loads have been lowered away */
3954 ASSERTED nir_const_value *offset = nir_src_as_const_value(instr->src[1]);
3955 assert(offset);
3956 assert(offset[0].i32 == 0);
3957
3958 LLVMValueRef interp_param = get_src(ctx, instr->src[0]);
3959 unsigned index = nir_intrinsic_base(instr);
3960 unsigned component = nir_intrinsic_component(instr);
3961 result = load_interpolated_input(ctx, interp_param, index,
3962 component,
3963 instr->dest.ssa.num_components,
3964 instr->dest.ssa.bit_size);
3965 break;
3966 }
3967 case nir_intrinsic_load_input:
3968 case nir_intrinsic_load_input_vertex:
3969 result = load_input(ctx, instr);
3970 break;
3971 case nir_intrinsic_emit_vertex:
3972 ctx->abi->emit_vertex(ctx->abi, nir_intrinsic_stream_id(instr), ctx->abi->outputs);
3973 break;
3974 case nir_intrinsic_emit_vertex_with_counter: {
3975 unsigned stream = nir_intrinsic_stream_id(instr);
3976 LLVMValueRef next_vertex = get_src(ctx, instr->src[0]);
3977 ctx->abi->emit_vertex_with_counter(ctx->abi, stream,
3978 next_vertex,
3979 ctx->abi->outputs);
3980 break;
3981 }
3982 case nir_intrinsic_end_primitive:
3983 case nir_intrinsic_end_primitive_with_counter:
3984 ctx->abi->emit_primitive(ctx->abi, nir_intrinsic_stream_id(instr));
3985 break;
3986 case nir_intrinsic_load_tess_coord:
3987 result = ctx->abi->load_tess_coord(ctx->abi);
3988 break;
3989 case nir_intrinsic_load_tess_level_outer:
3990 result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_OUTER, false);
3991 break;
3992 case nir_intrinsic_load_tess_level_inner:
3993 result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_INNER, false);
3994 break;
3995 case nir_intrinsic_load_tess_level_outer_default:
3996 result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_OUTER, true);
3997 break;
3998 case nir_intrinsic_load_tess_level_inner_default:
3999 result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_INNER, true);
4000 break;
4001 case nir_intrinsic_load_patch_vertices_in:
4002 result = ctx->abi->load_patch_vertices_in(ctx->abi);
4003 break;
4004 case nir_intrinsic_vote_all: {
4005 LLVMValueRef tmp = ac_build_vote_all(&ctx->ac, get_src(ctx, instr->src[0]));
4006 result = LLVMBuildSExt(ctx->ac.builder, tmp, ctx->ac.i32, "");
4007 break;
4008 }
4009 case nir_intrinsic_vote_any: {
4010 LLVMValueRef tmp = ac_build_vote_any(&ctx->ac, get_src(ctx, instr->src[0]));
4011 result = LLVMBuildSExt(ctx->ac.builder, tmp, ctx->ac.i32, "");
4012 break;
4013 }
4014 case nir_intrinsic_shuffle:
4015 if (ctx->ac.chip_class == GFX8 ||
4016 ctx->ac.chip_class == GFX9 ||
4017 (ctx->ac.chip_class == GFX10 && ctx->ac.wave_size == 32)) {
4018 result = ac_build_shuffle(&ctx->ac, get_src(ctx, instr->src[0]),
4019 get_src(ctx, instr->src[1]));
4020 } else {
4021 LLVMValueRef src = get_src(ctx, instr->src[0]);
4022 LLVMValueRef index = get_src(ctx, instr->src[1]);
4023 LLVMTypeRef type = LLVMTypeOf(src);
4024 struct waterfall_context wctx;
4025 LLVMValueRef index_val;
4026
4027 index_val = enter_waterfall(ctx, &wctx, index, true);
4028
4029 src = LLVMBuildZExt(ctx->ac.builder, src,
4030 ctx->ac.i32, "");
4031
4032 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.readlane",
4033 ctx->ac.i32,
4034 (LLVMValueRef []) { src, index_val }, 2,
4035 AC_FUNC_ATTR_READNONE |
4036 AC_FUNC_ATTR_CONVERGENT);
4037
4038 result = LLVMBuildTrunc(ctx->ac.builder, result, type, "");
4039
4040 result = exit_waterfall(ctx, &wctx, result);
4041 }
4042 break;
4043 case nir_intrinsic_reduce:
4044 result = ac_build_reduce(&ctx->ac,
4045 get_src(ctx, instr->src[0]),
4046 instr->const_index[0],
4047 instr->const_index[1]);
4048 break;
4049 case nir_intrinsic_inclusive_scan:
4050 result = ac_build_inclusive_scan(&ctx->ac,
4051 get_src(ctx, instr->src[0]),
4052 instr->const_index[0]);
4053 break;
4054 case nir_intrinsic_exclusive_scan:
4055 result = ac_build_exclusive_scan(&ctx->ac,
4056 get_src(ctx, instr->src[0]),
4057 instr->const_index[0]);
4058 break;
4059 case nir_intrinsic_quad_broadcast: {
4060 unsigned lane = nir_src_as_uint(instr->src[1]);
4061 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]),
4062 lane, lane, lane, lane);
4063 break;
4064 }
4065 case nir_intrinsic_quad_swap_horizontal:
4066 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 1, 0, 3 ,2);
4067 break;
4068 case nir_intrinsic_quad_swap_vertical:
4069 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 2, 3, 0 ,1);
4070 break;
4071 case nir_intrinsic_quad_swap_diagonal:
4072 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 3, 2, 1 ,0);
4073 break;
4074 case nir_intrinsic_quad_swizzle_amd: {
4075 uint32_t mask = nir_intrinsic_swizzle_mask(instr);
4076 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]),
4077 mask & 0x3, (mask >> 2) & 0x3,
4078 (mask >> 4) & 0x3, (mask >> 6) & 0x3);
4079 break;
4080 }
4081 case nir_intrinsic_masked_swizzle_amd: {
4082 uint32_t mask = nir_intrinsic_swizzle_mask(instr);
4083 result = ac_build_ds_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), mask);
4084 break;
4085 }
4086 case nir_intrinsic_write_invocation_amd:
4087 result = ac_build_writelane(&ctx->ac, get_src(ctx, instr->src[0]),
4088 get_src(ctx, instr->src[1]),
4089 get_src(ctx, instr->src[2]));
4090 break;
4091 case nir_intrinsic_mbcnt_amd:
4092 result = ac_build_mbcnt(&ctx->ac, get_src(ctx, instr->src[0]));
4093 break;
4094 case nir_intrinsic_load_scratch: {
4095 LLVMValueRef offset = get_src(ctx, instr->src[0]);
4096 LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->scratch,
4097 offset);
4098 LLVMTypeRef comp_type =
4099 LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
4100 LLVMTypeRef vec_type =
4101 instr->dest.ssa.num_components == 1 ? comp_type :
4102 LLVMVectorType(comp_type, instr->dest.ssa.num_components);
4103 unsigned addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
4104 ptr = LLVMBuildBitCast(ctx->ac.builder, ptr,
4105 LLVMPointerType(vec_type, addr_space), "");
4106 result = LLVMBuildLoad(ctx->ac.builder, ptr, "");
4107 break;
4108 }
4109 case nir_intrinsic_store_scratch: {
4110 LLVMValueRef offset = get_src(ctx, instr->src[1]);
4111 LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->scratch,
4112 offset);
4113 LLVMTypeRef comp_type =
4114 LLVMIntTypeInContext(ctx->ac.context, instr->src[0].ssa->bit_size);
4115 unsigned addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
4116 ptr = LLVMBuildBitCast(ctx->ac.builder, ptr,
4117 LLVMPointerType(comp_type, addr_space), "");
4118 LLVMValueRef src = get_src(ctx, instr->src[0]);
4119 unsigned wrmask = nir_intrinsic_write_mask(instr);
4120 while (wrmask) {
4121 int start, count;
4122 u_bit_scan_consecutive_range(&wrmask, &start, &count);
4123
4124 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, start, false);
4125 LLVMValueRef offset_ptr = LLVMBuildGEP(ctx->ac.builder, ptr, &offset, 1, "");
4126 LLVMTypeRef vec_type =
4127 count == 1 ? comp_type : LLVMVectorType(comp_type, count);
4128 offset_ptr = LLVMBuildBitCast(ctx->ac.builder,
4129 offset_ptr,
4130 LLVMPointerType(vec_type, addr_space),
4131 "");
4132 LLVMValueRef offset_src =
4133 ac_extract_components(&ctx->ac, src, start, count);
4134 LLVMBuildStore(ctx->ac.builder, offset_src, offset_ptr);
4135 }
4136 break;
4137 }
4138 case nir_intrinsic_load_constant: {
4139 unsigned base = nir_intrinsic_base(instr);
4140 unsigned range = nir_intrinsic_range(instr);
4141
4142 LLVMValueRef offset = get_src(ctx, instr->src[0]);
4143 offset = LLVMBuildAdd(ctx->ac.builder, offset,
4144 LLVMConstInt(ctx->ac.i32, base, false), "");
4145
4146 /* Clamp the offset to avoid out-of-bound access because global
4147 * instructions can't handle them.
4148 */
4149 LLVMValueRef size = LLVMConstInt(ctx->ac.i32, base + range, false);
4150 LLVMValueRef cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT,
4151 offset, size, "");
4152 offset = LLVMBuildSelect(ctx->ac.builder, cond, offset, size, "");
4153
4154 LLVMValueRef ptr = ac_build_gep0(&ctx->ac, ctx->constant_data,
4155 offset);
4156 LLVMTypeRef comp_type =
4157 LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
4158 LLVMTypeRef vec_type =
4159 instr->dest.ssa.num_components == 1 ? comp_type :
4160 LLVMVectorType(comp_type, instr->dest.ssa.num_components);
4161 unsigned addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
4162 ptr = LLVMBuildBitCast(ctx->ac.builder, ptr,
4163 LLVMPointerType(vec_type, addr_space), "");
4164 result = LLVMBuildLoad(ctx->ac.builder, ptr, "");
4165 break;
4166 }
4167 default:
4168 fprintf(stderr, "Unknown intrinsic: ");
4169 nir_print_instr(&instr->instr, stderr);
4170 fprintf(stderr, "\n");
4171 break;
4172 }
4173 if (result) {
4174 ctx->ssa_defs[instr->dest.ssa.index] = result;
4175 }
4176 }
4177
4178 static LLVMValueRef get_bindless_index_from_uniform(struct ac_nir_context *ctx,
4179 unsigned base_index,
4180 unsigned constant_index,
4181 LLVMValueRef dynamic_index)
4182 {
4183 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, base_index * 4, 0);
4184 LLVMValueRef index = LLVMBuildAdd(ctx->ac.builder, dynamic_index,
4185 LLVMConstInt(ctx->ac.i32, constant_index, 0), "");
4186
4187 /* Bindless uniforms are 64bit so multiple index by 8 */
4188 index = LLVMBuildMul(ctx->ac.builder, index, LLVMConstInt(ctx->ac.i32, 8, 0), "");
4189 offset = LLVMBuildAdd(ctx->ac.builder, offset, index, "");
4190
4191 LLVMValueRef ubo_index = ctx->abi->load_ubo(ctx->abi, ctx->ac.i32_0);
4192
4193 LLVMValueRef ret = ac_build_buffer_load(&ctx->ac, ubo_index, 1, NULL, offset,
4194 NULL, 0, 0, true, true);
4195
4196 return LLVMBuildBitCast(ctx->ac.builder, ret, ctx->ac.i32, "");
4197 }
4198
4199 struct sampler_desc_address {
4200 unsigned descriptor_set;
4201 unsigned base_index; /* binding in vulkan */
4202 unsigned constant_index;
4203 LLVMValueRef dynamic_index;
4204 bool image;
4205 bool bindless;
4206 };
4207
4208 static struct sampler_desc_address
4209 get_sampler_desc_internal(struct ac_nir_context *ctx,
4210 nir_deref_instr *deref_instr,
4211 const nir_instr *instr,
4212 bool image)
4213 {
4214 LLVMValueRef index = NULL;
4215 unsigned constant_index = 0;
4216 unsigned descriptor_set;
4217 unsigned base_index;
4218 bool bindless = false;
4219
4220 if (!deref_instr) {
4221 descriptor_set = 0;
4222 if (image) {
4223 nir_intrinsic_instr *img_instr = nir_instr_as_intrinsic(instr);
4224 base_index = 0;
4225 bindless = true;
4226 index = get_src(ctx, img_instr->src[0]);
4227 } else {
4228 nir_tex_instr *tex_instr = nir_instr_as_tex(instr);
4229 int sampSrcIdx = nir_tex_instr_src_index(tex_instr,
4230 nir_tex_src_sampler_handle);
4231 if (sampSrcIdx != -1) {
4232 base_index = 0;
4233 bindless = true;
4234 index = get_src(ctx, tex_instr->src[sampSrcIdx].src);
4235 } else {
4236 assert(tex_instr && !image);
4237 base_index = tex_instr->sampler_index;
4238 }
4239 }
4240 } else {
4241 while(deref_instr->deref_type != nir_deref_type_var) {
4242 if (deref_instr->deref_type == nir_deref_type_array) {
4243 unsigned array_size = glsl_get_aoa_size(deref_instr->type);
4244 if (!array_size)
4245 array_size = 1;
4246
4247 if (nir_src_is_const(deref_instr->arr.index)) {
4248 constant_index += array_size * nir_src_as_uint(deref_instr->arr.index);
4249 } else {
4250 LLVMValueRef indirect = get_src(ctx, deref_instr->arr.index);
4251
4252 indirect = LLVMBuildMul(ctx->ac.builder, indirect,
4253 LLVMConstInt(ctx->ac.i32, array_size, false), "");
4254
4255 if (!index)
4256 index = indirect;
4257 else
4258 index = LLVMBuildAdd(ctx->ac.builder, index, indirect, "");
4259 }
4260
4261 deref_instr = nir_src_as_deref(deref_instr->parent);
4262 } else if (deref_instr->deref_type == nir_deref_type_struct) {
4263 unsigned sidx = deref_instr->strct.index;
4264 deref_instr = nir_src_as_deref(deref_instr->parent);
4265 constant_index += glsl_get_struct_location_offset(deref_instr->type, sidx);
4266 } else {
4267 unreachable("Unsupported deref type");
4268 }
4269 }
4270 descriptor_set = deref_instr->var->data.descriptor_set;
4271
4272 if (deref_instr->var->data.bindless) {
4273 /* For now just assert on unhandled variable types */
4274 assert(deref_instr->var->data.mode == nir_var_uniform);
4275
4276 base_index = deref_instr->var->data.driver_location;
4277 bindless = true;
4278
4279 index = index ? index : ctx->ac.i32_0;
4280 index = get_bindless_index_from_uniform(ctx, base_index,
4281 constant_index, index);
4282 } else
4283 base_index = deref_instr->var->data.binding;
4284 }
4285 return (struct sampler_desc_address) {
4286 .descriptor_set = descriptor_set,
4287 .base_index = base_index,
4288 .constant_index = constant_index,
4289 .dynamic_index = index,
4290 .image = image,
4291 .bindless = bindless,
4292 };
4293 }
4294
4295 /* Extract any possibly divergent index into a separate value that can be fed
4296 * into get_sampler_desc with the same arguments. */
4297 static LLVMValueRef get_sampler_desc_index(struct ac_nir_context *ctx,
4298 nir_deref_instr *deref_instr,
4299 const nir_instr *instr,
4300 bool image)
4301 {
4302 struct sampler_desc_address addr = get_sampler_desc_internal(ctx, deref_instr, instr, image);
4303 return addr.dynamic_index;
4304 }
4305
4306 static LLVMValueRef get_sampler_desc(struct ac_nir_context *ctx,
4307 nir_deref_instr *deref_instr,
4308 enum ac_descriptor_type desc_type,
4309 const nir_instr *instr,
4310 LLVMValueRef index,
4311 bool image, bool write)
4312 {
4313 struct sampler_desc_address addr = get_sampler_desc_internal(ctx, deref_instr, instr, image);
4314 return ctx->abi->load_sampler_desc(ctx->abi,
4315 addr.descriptor_set,
4316 addr.base_index,
4317 addr.constant_index, index,
4318 desc_type, addr.image, write, addr.bindless);
4319 }
4320
4321 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
4322 *
4323 * GFX6-GFX7:
4324 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
4325 * filtering manually. The driver sets img7 to a mask clearing
4326 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
4327 * s_and_b32 samp0, samp0, img7
4328 *
4329 * GFX8:
4330 * The ANISO_OVERRIDE sampler field enables this fix in TA.
4331 */
4332 static LLVMValueRef sici_fix_sampler_aniso(struct ac_nir_context *ctx,
4333 LLVMValueRef res, LLVMValueRef samp)
4334 {
4335 LLVMBuilderRef builder = ctx->ac.builder;
4336 LLVMValueRef img7, samp0;
4337
4338 if (ctx->ac.chip_class >= GFX8)
4339 return samp;
4340
4341 img7 = LLVMBuildExtractElement(builder, res,
4342 LLVMConstInt(ctx->ac.i32, 7, 0), "");
4343 samp0 = LLVMBuildExtractElement(builder, samp,
4344 LLVMConstInt(ctx->ac.i32, 0, 0), "");
4345 samp0 = LLVMBuildAnd(builder, samp0, img7, "");
4346 return LLVMBuildInsertElement(builder, samp, samp0,
4347 LLVMConstInt(ctx->ac.i32, 0, 0), "");
4348 }
4349
4350 static void tex_fetch_ptrs(struct ac_nir_context *ctx,
4351 nir_tex_instr *instr,
4352 struct waterfall_context *wctx,
4353 LLVMValueRef *res_ptr, LLVMValueRef *samp_ptr,
4354 LLVMValueRef *fmask_ptr)
4355 {
4356 nir_deref_instr *texture_deref_instr = NULL;
4357 nir_deref_instr *sampler_deref_instr = NULL;
4358 int plane = -1;
4359
4360 for (unsigned i = 0; i < instr->num_srcs; i++) {
4361 switch (instr->src[i].src_type) {
4362 case nir_tex_src_texture_deref:
4363 texture_deref_instr = nir_src_as_deref(instr->src[i].src);
4364 break;
4365 case nir_tex_src_sampler_deref:
4366 sampler_deref_instr = nir_src_as_deref(instr->src[i].src);
4367 break;
4368 case nir_tex_src_plane:
4369 plane = nir_src_as_int(instr->src[i].src);
4370 break;
4371 default:
4372 break;
4373 }
4374 }
4375
4376 LLVMValueRef texture_dynamic_index = get_sampler_desc_index(ctx, texture_deref_instr,
4377 &instr->instr, false);
4378 if (!sampler_deref_instr)
4379 sampler_deref_instr = texture_deref_instr;
4380
4381 LLVMValueRef sampler_dynamic_index = get_sampler_desc_index(ctx, sampler_deref_instr,
4382 &instr->instr, false);
4383 if (instr->texture_non_uniform)
4384 texture_dynamic_index = enter_waterfall(ctx, wctx + 0, texture_dynamic_index, true);
4385
4386 if (instr->sampler_non_uniform)
4387 sampler_dynamic_index = enter_waterfall(ctx, wctx + 1, sampler_dynamic_index, true);
4388
4389 enum ac_descriptor_type main_descriptor = instr->sampler_dim == GLSL_SAMPLER_DIM_BUF ? AC_DESC_BUFFER : AC_DESC_IMAGE;
4390
4391 if (plane >= 0) {
4392 assert(instr->op != nir_texop_txf_ms &&
4393 instr->op != nir_texop_samples_identical);
4394 assert(instr->sampler_dim != GLSL_SAMPLER_DIM_BUF);
4395
4396 main_descriptor = AC_DESC_PLANE_0 + plane;
4397 }
4398
4399 if (instr->op == nir_texop_fragment_mask_fetch) {
4400 /* The fragment mask is fetched from the compressed
4401 * multisampled surface.
4402 */
4403 main_descriptor = AC_DESC_FMASK;
4404 }
4405
4406 *res_ptr = get_sampler_desc(ctx, texture_deref_instr, main_descriptor, &instr->instr,
4407 texture_dynamic_index, false, false);
4408
4409 if (samp_ptr) {
4410 *samp_ptr = get_sampler_desc(ctx, sampler_deref_instr, AC_DESC_SAMPLER, &instr->instr,
4411 sampler_dynamic_index, false, false);
4412 if (instr->sampler_dim < GLSL_SAMPLER_DIM_RECT)
4413 *samp_ptr = sici_fix_sampler_aniso(ctx, *res_ptr, *samp_ptr);
4414 }
4415 if (fmask_ptr && (instr->op == nir_texop_txf_ms ||
4416 instr->op == nir_texop_samples_identical))
4417 *fmask_ptr = get_sampler_desc(ctx, texture_deref_instr, AC_DESC_FMASK,
4418 &instr->instr, texture_dynamic_index, false, false);
4419 }
4420
4421 static LLVMValueRef apply_round_slice(struct ac_llvm_context *ctx,
4422 LLVMValueRef coord)
4423 {
4424 coord = ac_to_float(ctx, coord);
4425 coord = ac_build_round(ctx, coord);
4426 coord = ac_to_integer(ctx, coord);
4427 return coord;
4428 }
4429
4430 static void visit_tex(struct ac_nir_context *ctx, nir_tex_instr *instr)
4431 {
4432 LLVMValueRef result = NULL;
4433 struct ac_image_args args = { 0 };
4434 LLVMValueRef fmask_ptr = NULL, sample_index = NULL;
4435 LLVMValueRef ddx = NULL, ddy = NULL;
4436 unsigned offset_src = 0;
4437 struct waterfall_context wctx[2] = {{{0}}};
4438
4439 tex_fetch_ptrs(ctx, instr, wctx, &args.resource, &args.sampler, &fmask_ptr);
4440
4441 for (unsigned i = 0; i < instr->num_srcs; i++) {
4442 switch (instr->src[i].src_type) {
4443 case nir_tex_src_coord: {
4444 LLVMValueRef coord = get_src(ctx, instr->src[i].src);
4445 for (unsigned chan = 0; chan < instr->coord_components; ++chan)
4446 args.coords[chan] = ac_llvm_extract_elem(&ctx->ac, coord, chan);
4447 break;
4448 }
4449 case nir_tex_src_projector:
4450 break;
4451 case nir_tex_src_comparator:
4452 if (instr->is_shadow) {
4453 args.compare = get_src(ctx, instr->src[i].src);
4454 args.compare = ac_to_float(&ctx->ac, args.compare);
4455 }
4456 break;
4457 case nir_tex_src_offset:
4458 args.offset = get_src(ctx, instr->src[i].src);
4459 offset_src = i;
4460 break;
4461 case nir_tex_src_bias:
4462 args.bias = get_src(ctx, instr->src[i].src);
4463 break;
4464 case nir_tex_src_lod: {
4465 if (nir_src_is_const(instr->src[i].src) && nir_src_as_uint(instr->src[i].src) == 0)
4466 args.level_zero = true;
4467 else
4468 args.lod = get_src(ctx, instr->src[i].src);
4469 break;
4470 }
4471 case nir_tex_src_ms_index:
4472 sample_index = get_src(ctx, instr->src[i].src);
4473 break;
4474 case nir_tex_src_ms_mcs:
4475 break;
4476 case nir_tex_src_ddx:
4477 ddx = get_src(ctx, instr->src[i].src);
4478 break;
4479 case nir_tex_src_ddy:
4480 ddy = get_src(ctx, instr->src[i].src);
4481 break;
4482 case nir_tex_src_min_lod:
4483 args.min_lod = get_src(ctx, instr->src[i].src);
4484 break;
4485 case nir_tex_src_texture_offset:
4486 case nir_tex_src_sampler_offset:
4487 case nir_tex_src_plane:
4488 default:
4489 break;
4490 }
4491 }
4492
4493 if (instr->op == nir_texop_txs && instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
4494 result = get_buffer_size(ctx, args.resource, true);
4495 goto write_result;
4496 }
4497
4498 if (instr->op == nir_texop_texture_samples) {
4499 LLVMValueRef res, samples, is_msaa;
4500 LLVMValueRef default_sample;
4501
4502 res = LLVMBuildBitCast(ctx->ac.builder, args.resource, ctx->ac.v8i32, "");
4503 samples = LLVMBuildExtractElement(ctx->ac.builder, res,
4504 LLVMConstInt(ctx->ac.i32, 3, false), "");
4505 is_msaa = LLVMBuildLShr(ctx->ac.builder, samples,
4506 LLVMConstInt(ctx->ac.i32, 28, false), "");
4507 is_msaa = LLVMBuildAnd(ctx->ac.builder, is_msaa,
4508 LLVMConstInt(ctx->ac.i32, 0xe, false), "");
4509 is_msaa = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, is_msaa,
4510 LLVMConstInt(ctx->ac.i32, 0xe, false), "");
4511
4512 samples = LLVMBuildLShr(ctx->ac.builder, samples,
4513 LLVMConstInt(ctx->ac.i32, 16, false), "");
4514 samples = LLVMBuildAnd(ctx->ac.builder, samples,
4515 LLVMConstInt(ctx->ac.i32, 0xf, false), "");
4516 samples = LLVMBuildShl(ctx->ac.builder, ctx->ac.i32_1,
4517 samples, "");
4518
4519 if (ctx->abi->robust_buffer_access) {
4520 LLVMValueRef dword1, is_null_descriptor;
4521
4522 /* Extract the second dword of the descriptor, if it's
4523 * all zero, then it's a null descriptor.
4524 */
4525 dword1 = LLVMBuildExtractElement(ctx->ac.builder, res,
4526 LLVMConstInt(ctx->ac.i32, 1, false), "");
4527 is_null_descriptor =
4528 LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, dword1,
4529 LLVMConstInt(ctx->ac.i32, 0, false), "");
4530 default_sample =
4531 LLVMBuildSelect(ctx->ac.builder, is_null_descriptor,
4532 ctx->ac.i32_0, ctx->ac.i32_1, "");
4533 } else {
4534 default_sample = ctx->ac.i32_1;
4535 }
4536
4537 samples = LLVMBuildSelect(ctx->ac.builder, is_msaa, samples,
4538 default_sample, "");
4539 result = samples;
4540 goto write_result;
4541 }
4542
4543 if (args.offset && instr->op != nir_texop_txf && instr->op != nir_texop_txf_ms) {
4544 LLVMValueRef offset[3], pack;
4545 for (unsigned chan = 0; chan < 3; ++chan)
4546 offset[chan] = ctx->ac.i32_0;
4547
4548 unsigned num_components = ac_get_llvm_num_components(args.offset);
4549 for (unsigned chan = 0; chan < num_components; chan++) {
4550 offset[chan] = ac_llvm_extract_elem(&ctx->ac, args.offset, chan);
4551 offset[chan] = LLVMBuildAnd(ctx->ac.builder, offset[chan],
4552 LLVMConstInt(ctx->ac.i32, 0x3f, false), "");
4553 if (chan)
4554 offset[chan] = LLVMBuildShl(ctx->ac.builder, offset[chan],
4555 LLVMConstInt(ctx->ac.i32, chan * 8, false), "");
4556 }
4557 pack = LLVMBuildOr(ctx->ac.builder, offset[0], offset[1], "");
4558 pack = LLVMBuildOr(ctx->ac.builder, pack, offset[2], "");
4559 args.offset = pack;
4560 }
4561
4562 /* Section 8.23.1 (Depth Texture Comparison Mode) of the
4563 * OpenGL 4.5 spec says:
4564 *
4565 * "If the texture’s internal format indicates a fixed-point
4566 * depth texture, then D_t and D_ref are clamped to the
4567 * range [0, 1]; otherwise no clamping is performed."
4568 *
4569 * TC-compatible HTILE promotes Z16 and Z24 to Z32_FLOAT,
4570 * so the depth comparison value isn't clamped for Z16 and
4571 * Z24 anymore. Do it manually here for GFX8-9; GFX10 has
4572 * an explicitly clamped 32-bit float format.
4573 */
4574 if (args.compare &&
4575 ctx->ac.chip_class >= GFX8 &&
4576 ctx->ac.chip_class <= GFX9 &&
4577 ctx->abi->clamp_shadow_reference) {
4578 LLVMValueRef upgraded, clamped;
4579
4580 upgraded = LLVMBuildExtractElement(ctx->ac.builder, args.sampler,
4581 LLVMConstInt(ctx->ac.i32, 3, false), "");
4582 upgraded = LLVMBuildLShr(ctx->ac.builder, upgraded,
4583 LLVMConstInt(ctx->ac.i32, 29, false), "");
4584 upgraded = LLVMBuildTrunc(ctx->ac.builder, upgraded, ctx->ac.i1, "");
4585 clamped = ac_build_clamp(&ctx->ac, args.compare);
4586 args.compare = LLVMBuildSelect(ctx->ac.builder, upgraded, clamped,
4587 args.compare, "");
4588 }
4589
4590 /* pack derivatives */
4591 if (ddx || ddy) {
4592 int num_src_deriv_channels, num_dest_deriv_channels;
4593 switch (instr->sampler_dim) {
4594 case GLSL_SAMPLER_DIM_3D:
4595 case GLSL_SAMPLER_DIM_CUBE:
4596 num_src_deriv_channels = 3;
4597 num_dest_deriv_channels = 3;
4598 break;
4599 case GLSL_SAMPLER_DIM_2D:
4600 default:
4601 num_src_deriv_channels = 2;
4602 num_dest_deriv_channels = 2;
4603 break;
4604 case GLSL_SAMPLER_DIM_1D:
4605 num_src_deriv_channels = 1;
4606 if (ctx->ac.chip_class == GFX9) {
4607 num_dest_deriv_channels = 2;
4608 } else {
4609 num_dest_deriv_channels = 1;
4610 }
4611 break;
4612 }
4613
4614 for (unsigned i = 0; i < num_src_deriv_channels; i++) {
4615 args.derivs[i] = ac_to_float(&ctx->ac,
4616 ac_llvm_extract_elem(&ctx->ac, ddx, i));
4617 args.derivs[num_dest_deriv_channels + i] = ac_to_float(&ctx->ac,
4618 ac_llvm_extract_elem(&ctx->ac, ddy, i));
4619 }
4620 for (unsigned i = num_src_deriv_channels; i < num_dest_deriv_channels; i++) {
4621 args.derivs[i] = ctx->ac.f32_0;
4622 args.derivs[num_dest_deriv_channels + i] = ctx->ac.f32_0;
4623 }
4624 }
4625
4626 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE && args.coords[0]) {
4627 for (unsigned chan = 0; chan < instr->coord_components; chan++)
4628 args.coords[chan] = ac_to_float(&ctx->ac, args.coords[chan]);
4629 if (instr->coord_components == 3)
4630 args.coords[3] = LLVMGetUndef(ctx->ac.f32);
4631 ac_prepare_cube_coords(&ctx->ac,
4632 instr->op == nir_texop_txd, instr->is_array,
4633 instr->op == nir_texop_lod, args.coords, args.derivs);
4634 }
4635
4636 /* Texture coordinates fixups */
4637 if (instr->coord_components > 1 &&
4638 instr->sampler_dim == GLSL_SAMPLER_DIM_1D &&
4639 instr->is_array &&
4640 instr->op != nir_texop_txf) {
4641 args.coords[1] = apply_round_slice(&ctx->ac, args.coords[1]);
4642 }
4643
4644 if (instr->coord_components > 2 &&
4645 (instr->sampler_dim == GLSL_SAMPLER_DIM_2D ||
4646 instr->sampler_dim == GLSL_SAMPLER_DIM_MS ||
4647 instr->sampler_dim == GLSL_SAMPLER_DIM_SUBPASS ||
4648 instr->sampler_dim == GLSL_SAMPLER_DIM_SUBPASS_MS) &&
4649 instr->is_array &&
4650 instr->op != nir_texop_txf &&
4651 instr->op != nir_texop_txf_ms &&
4652 instr->op != nir_texop_fragment_fetch &&
4653 instr->op != nir_texop_fragment_mask_fetch) {
4654 args.coords[2] = apply_round_slice(&ctx->ac, args.coords[2]);
4655 }
4656
4657 if (ctx->ac.chip_class == GFX9 &&
4658 instr->sampler_dim == GLSL_SAMPLER_DIM_1D &&
4659 instr->op != nir_texop_lod) {
4660 LLVMValueRef filler;
4661 if (instr->op == nir_texop_txf)
4662 filler = ctx->ac.i32_0;
4663 else
4664 filler = LLVMConstReal(ctx->ac.f32, 0.5);
4665
4666 if (instr->is_array)
4667 args.coords[2] = args.coords[1];
4668 args.coords[1] = filler;
4669 }
4670
4671 /* Pack sample index */
4672 if (sample_index && (instr->op == nir_texop_txf_ms ||
4673 instr->op == nir_texop_fragment_fetch))
4674 args.coords[instr->coord_components] = sample_index;
4675
4676 if (instr->op == nir_texop_samples_identical) {
4677 struct ac_image_args txf_args = { 0 };
4678 memcpy(txf_args.coords, args.coords, sizeof(txf_args.coords));
4679
4680 txf_args.dmask = 0xf;
4681 txf_args.resource = fmask_ptr;
4682 txf_args.dim = instr->is_array ? ac_image_2darray : ac_image_2d;
4683 result = build_tex_intrinsic(ctx, instr, &txf_args);
4684
4685 result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
4686 result = emit_int_cmp(&ctx->ac, LLVMIntEQ, result, ctx->ac.i32_0);
4687 goto write_result;
4688 }
4689
4690 if ((instr->sampler_dim == GLSL_SAMPLER_DIM_SUBPASS_MS ||
4691 instr->sampler_dim == GLSL_SAMPLER_DIM_MS) &&
4692 instr->op != nir_texop_txs &&
4693 instr->op != nir_texop_fragment_fetch &&
4694 instr->op != nir_texop_fragment_mask_fetch) {
4695 unsigned sample_chan = instr->is_array ? 3 : 2;
4696 args.coords[sample_chan] = adjust_sample_index_using_fmask(
4697 &ctx->ac, args.coords[0], args.coords[1],
4698 instr->is_array ? args.coords[2] : NULL,
4699 args.coords[sample_chan], fmask_ptr);
4700 }
4701
4702 if (args.offset && (instr->op == nir_texop_txf || instr->op == nir_texop_txf_ms)) {
4703 int num_offsets = instr->src[offset_src].src.ssa->num_components;
4704 num_offsets = MIN2(num_offsets, instr->coord_components);
4705 for (unsigned i = 0; i < num_offsets; ++i) {
4706 args.coords[i] = LLVMBuildAdd(
4707 ctx->ac.builder, args.coords[i],
4708 LLVMConstInt(ctx->ac.i32, nir_src_comp_as_uint(instr->src[offset_src].src, i), false), "");
4709 }
4710 args.offset = NULL;
4711 }
4712
4713 /* DMASK was repurposed for GATHER4. 4 components are always
4714 * returned and DMASK works like a swizzle - it selects
4715 * the component to fetch. The only valid DMASK values are
4716 * 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
4717 * (red,red,red,red) etc.) The ISA document doesn't mention
4718 * this.
4719 */
4720 args.dmask = 0xf;
4721 if (instr->op == nir_texop_tg4) {
4722 if (instr->is_shadow)
4723 args.dmask = 1;
4724 else
4725 args.dmask = 1 << instr->component;
4726 }
4727
4728 if (instr->sampler_dim != GLSL_SAMPLER_DIM_BUF) {
4729 args.dim = ac_get_sampler_dim(ctx->ac.chip_class, instr->sampler_dim, instr->is_array);
4730 args.unorm = instr->sampler_dim == GLSL_SAMPLER_DIM_RECT;
4731 }
4732
4733 /* Adjust the number of coordinates because we only need (x,y) for 2D
4734 * multisampled images and (x,y,layer) for 2D multisampled layered
4735 * images or for multisampled input attachments.
4736 */
4737 if (instr->op == nir_texop_fragment_mask_fetch) {
4738 if (args.dim == ac_image_2dmsaa) {
4739 args.dim = ac_image_2d;
4740 } else {
4741 assert(args.dim == ac_image_2darraymsaa);
4742 args.dim = ac_image_2darray;
4743 }
4744 }
4745
4746 assert(instr->dest.is_ssa);
4747 args.d16 = instr->dest.ssa.bit_size == 16;
4748
4749 result = build_tex_intrinsic(ctx, instr, &args);
4750
4751 if (instr->op == nir_texop_query_levels)
4752 result = LLVMBuildExtractElement(ctx->ac.builder, result, LLVMConstInt(ctx->ac.i32, 3, false), "");
4753 else if (instr->is_shadow && instr->is_new_style_shadow &&
4754 instr->op != nir_texop_txs && instr->op != nir_texop_lod &&
4755 instr->op != nir_texop_tg4)
4756 result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
4757 else if (instr->op == nir_texop_txs &&
4758 instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE &&
4759 instr->is_array) {
4760 LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
4761 LLVMValueRef six = LLVMConstInt(ctx->ac.i32, 6, false);
4762 LLVMValueRef z = LLVMBuildExtractElement(ctx->ac.builder, result, two, "");
4763 z = LLVMBuildSDiv(ctx->ac.builder, z, six, "");
4764 result = LLVMBuildInsertElement(ctx->ac.builder, result, z, two, "");
4765 } else if (ctx->ac.chip_class == GFX9 &&
4766 instr->op == nir_texop_txs &&
4767 instr->sampler_dim == GLSL_SAMPLER_DIM_1D &&
4768 instr->is_array) {
4769 LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
4770 LLVMValueRef layers = LLVMBuildExtractElement(ctx->ac.builder, result, two, "");
4771 result = LLVMBuildInsertElement(ctx->ac.builder, result, layers,
4772 ctx->ac.i32_1, "");
4773 } else if (instr->dest.ssa.num_components != 4)
4774 result = ac_trim_vector(&ctx->ac, result, instr->dest.ssa.num_components);
4775
4776 write_result:
4777 if (result) {
4778 assert(instr->dest.is_ssa);
4779 result = ac_to_integer(&ctx->ac, result);
4780
4781 for (int i = ARRAY_SIZE(wctx); --i >= 0;) {
4782 result = exit_waterfall(ctx, wctx + i, result);
4783 }
4784
4785 ctx->ssa_defs[instr->dest.ssa.index] = result;
4786 }
4787 }
4788
4789 static void visit_phi(struct ac_nir_context *ctx, nir_phi_instr *instr)
4790 {
4791 LLVMTypeRef type = get_def_type(ctx, &instr->dest.ssa);
4792 LLVMValueRef result = LLVMBuildPhi(ctx->ac.builder, type, "");
4793
4794 ctx->ssa_defs[instr->dest.ssa.index] = result;
4795 _mesa_hash_table_insert(ctx->phis, instr, result);
4796 }
4797
4798 static void visit_post_phi(struct ac_nir_context *ctx,
4799 nir_phi_instr *instr,
4800 LLVMValueRef llvm_phi)
4801 {
4802 nir_foreach_phi_src(src, instr) {
4803 LLVMBasicBlockRef block = get_block(ctx, src->pred);
4804 LLVMValueRef llvm_src = get_src(ctx, src->src);
4805
4806 LLVMAddIncoming(llvm_phi, &llvm_src, &block, 1);
4807 }
4808 }
4809
4810 static void phi_post_pass(struct ac_nir_context *ctx)
4811 {
4812 hash_table_foreach(ctx->phis, entry) {
4813 visit_post_phi(ctx, (nir_phi_instr*)entry->key,
4814 (LLVMValueRef)entry->data);
4815 }
4816 }
4817
4818
4819 static bool is_def_used_in_an_export(const nir_ssa_def* def) {
4820 nir_foreach_use(use_src, def) {
4821 if (use_src->parent_instr->type == nir_instr_type_intrinsic) {
4822 nir_intrinsic_instr *instr = nir_instr_as_intrinsic(use_src->parent_instr);
4823 if (instr->intrinsic == nir_intrinsic_store_deref)
4824 return true;
4825 } else if (use_src->parent_instr->type == nir_instr_type_alu) {
4826 nir_alu_instr *instr = nir_instr_as_alu(use_src->parent_instr);
4827 if (instr->op == nir_op_vec4 &&
4828 is_def_used_in_an_export(&instr->dest.dest.ssa)) {
4829 return true;
4830 }
4831 }
4832 }
4833 return false;
4834 }
4835
4836 static void visit_ssa_undef(struct ac_nir_context *ctx,
4837 const nir_ssa_undef_instr *instr)
4838 {
4839 unsigned num_components = instr->def.num_components;
4840 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
4841
4842 if (!ctx->abi->convert_undef_to_zero || is_def_used_in_an_export(&instr->def)) {
4843 LLVMValueRef undef;
4844
4845 if (num_components == 1)
4846 undef = LLVMGetUndef(type);
4847 else {
4848 undef = LLVMGetUndef(LLVMVectorType(type, num_components));
4849 }
4850 ctx->ssa_defs[instr->def.index] = undef;
4851 } else {
4852 LLVMValueRef zero = LLVMConstInt(type, 0, false);
4853 if (num_components > 1) {
4854 zero = ac_build_gather_values_extended(
4855 &ctx->ac, &zero, 4, 0, false, false);
4856 }
4857 ctx->ssa_defs[instr->def.index] = zero;
4858 }
4859 }
4860
4861 static void visit_jump(struct ac_llvm_context *ctx,
4862 const nir_jump_instr *instr)
4863 {
4864 switch (instr->type) {
4865 case nir_jump_break:
4866 ac_build_break(ctx);
4867 break;
4868 case nir_jump_continue:
4869 ac_build_continue(ctx);
4870 break;
4871 default:
4872 fprintf(stderr, "Unknown NIR jump instr: ");
4873 nir_print_instr(&instr->instr, stderr);
4874 fprintf(stderr, "\n");
4875 abort();
4876 }
4877 }
4878
4879 static LLVMTypeRef
4880 glsl_base_to_llvm_type(struct ac_llvm_context *ac,
4881 enum glsl_base_type type)
4882 {
4883 switch (type) {
4884 case GLSL_TYPE_INT:
4885 case GLSL_TYPE_UINT:
4886 case GLSL_TYPE_BOOL:
4887 case GLSL_TYPE_SUBROUTINE:
4888 return ac->i32;
4889 case GLSL_TYPE_INT8:
4890 case GLSL_TYPE_UINT8:
4891 return ac->i8;
4892 case GLSL_TYPE_INT16:
4893 case GLSL_TYPE_UINT16:
4894 return ac->i16;
4895 case GLSL_TYPE_FLOAT:
4896 return ac->f32;
4897 case GLSL_TYPE_FLOAT16:
4898 return ac->f16;
4899 case GLSL_TYPE_INT64:
4900 case GLSL_TYPE_UINT64:
4901 return ac->i64;
4902 case GLSL_TYPE_DOUBLE:
4903 return ac->f64;
4904 default:
4905 unreachable("unknown GLSL type");
4906 }
4907 }
4908
4909 static LLVMTypeRef
4910 glsl_to_llvm_type(struct ac_llvm_context *ac,
4911 const struct glsl_type *type)
4912 {
4913 if (glsl_type_is_scalar(type)) {
4914 return glsl_base_to_llvm_type(ac, glsl_get_base_type(type));
4915 }
4916
4917 if (glsl_type_is_vector(type)) {
4918 return LLVMVectorType(
4919 glsl_base_to_llvm_type(ac, glsl_get_base_type(type)),
4920 glsl_get_vector_elements(type));
4921 }
4922
4923 if (glsl_type_is_matrix(type)) {
4924 return LLVMArrayType(
4925 glsl_to_llvm_type(ac, glsl_get_column_type(type)),
4926 glsl_get_matrix_columns(type));
4927 }
4928
4929 if (glsl_type_is_array(type)) {
4930 return LLVMArrayType(
4931 glsl_to_llvm_type(ac, glsl_get_array_element(type)),
4932 glsl_get_length(type));
4933 }
4934
4935 assert(glsl_type_is_struct_or_ifc(type));
4936
4937 LLVMTypeRef member_types[glsl_get_length(type)];
4938
4939 for (unsigned i = 0; i < glsl_get_length(type); i++) {
4940 member_types[i] =
4941 glsl_to_llvm_type(ac,
4942 glsl_get_struct_field(type, i));
4943 }
4944
4945 return LLVMStructTypeInContext(ac->context, member_types,
4946 glsl_get_length(type), false);
4947 }
4948
4949 static void visit_deref(struct ac_nir_context *ctx,
4950 nir_deref_instr *instr)
4951 {
4952 if (instr->mode != nir_var_mem_shared &&
4953 instr->mode != nir_var_mem_global)
4954 return;
4955
4956 LLVMValueRef result = NULL;
4957 switch(instr->deref_type) {
4958 case nir_deref_type_var: {
4959 struct hash_entry *entry = _mesa_hash_table_search(ctx->vars, instr->var);
4960 result = entry->data;
4961 break;
4962 }
4963 case nir_deref_type_struct:
4964 if (instr->mode == nir_var_mem_global) {
4965 nir_deref_instr *parent = nir_deref_instr_parent(instr);
4966 uint64_t offset = glsl_get_struct_field_offset(parent->type,
4967 instr->strct.index);
4968 result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent),
4969 LLVMConstInt(ctx->ac.i32, offset, 0));
4970 } else {
4971 result = ac_build_gep0(&ctx->ac, get_src(ctx, instr->parent),
4972 LLVMConstInt(ctx->ac.i32, instr->strct.index, 0));
4973 }
4974 break;
4975 case nir_deref_type_array:
4976 if (instr->mode == nir_var_mem_global) {
4977 nir_deref_instr *parent = nir_deref_instr_parent(instr);
4978 unsigned stride = glsl_get_explicit_stride(parent->type);
4979
4980 if ((glsl_type_is_matrix(parent->type) &&
4981 glsl_matrix_type_is_row_major(parent->type)) ||
4982 (glsl_type_is_vector(parent->type) && stride == 0))
4983 stride = type_scalar_size_bytes(parent->type);
4984
4985 assert(stride > 0);
4986 LLVMValueRef index = get_src(ctx, instr->arr.index);
4987 if (LLVMTypeOf(index) != ctx->ac.i64)
4988 index = LLVMBuildZExt(ctx->ac.builder, index, ctx->ac.i64, "");
4989
4990 LLVMValueRef offset = LLVMBuildMul(ctx->ac.builder, index, LLVMConstInt(ctx->ac.i64, stride, 0), "");
4991
4992 result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent), offset);
4993 } else {
4994 result = ac_build_gep0(&ctx->ac, get_src(ctx, instr->parent),
4995 get_src(ctx, instr->arr.index));
4996 }
4997 break;
4998 case nir_deref_type_ptr_as_array:
4999 if (instr->mode == nir_var_mem_global) {
5000 unsigned stride = nir_deref_instr_ptr_as_array_stride(instr);
5001
5002 LLVMValueRef index = get_src(ctx, instr->arr.index);
5003 if (LLVMTypeOf(index) != ctx->ac.i64)
5004 index = LLVMBuildZExt(ctx->ac.builder, index, ctx->ac.i64, "");
5005
5006 LLVMValueRef offset = LLVMBuildMul(ctx->ac.builder, index, LLVMConstInt(ctx->ac.i64, stride, 0), "");
5007
5008 result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent), offset);
5009 } else {
5010 result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent),
5011 get_src(ctx, instr->arr.index));
5012 }
5013 break;
5014 case nir_deref_type_cast: {
5015 result = get_src(ctx, instr->parent);
5016
5017 /* We can't use the structs from LLVM because the shader
5018 * specifies its own offsets. */
5019 LLVMTypeRef pointee_type = ctx->ac.i8;
5020 if (instr->mode == nir_var_mem_shared)
5021 pointee_type = glsl_to_llvm_type(&ctx->ac, instr->type);
5022
5023 unsigned address_space;
5024
5025 switch(instr->mode) {
5026 case nir_var_mem_shared:
5027 address_space = AC_ADDR_SPACE_LDS;
5028 break;
5029 case nir_var_mem_global:
5030 address_space = AC_ADDR_SPACE_GLOBAL;
5031 break;
5032 default:
5033 unreachable("Unhandled address space");
5034 }
5035
5036 LLVMTypeRef type = LLVMPointerType(pointee_type, address_space);
5037
5038 if (LLVMTypeOf(result) != type) {
5039 if (LLVMGetTypeKind(LLVMTypeOf(result)) == LLVMVectorTypeKind) {
5040 result = LLVMBuildBitCast(ctx->ac.builder, result,
5041 type, "");
5042 } else {
5043 result = LLVMBuildIntToPtr(ctx->ac.builder, result,
5044 type, "");
5045 }
5046 }
5047 break;
5048 }
5049 default:
5050 unreachable("Unhandled deref_instr deref type");
5051 }
5052
5053 ctx->ssa_defs[instr->dest.ssa.index] = result;
5054 }
5055
5056 static void visit_cf_list(struct ac_nir_context *ctx,
5057 struct exec_list *list);
5058
5059 static void visit_block(struct ac_nir_context *ctx, nir_block *block)
5060 {
5061 nir_foreach_instr(instr, block)
5062 {
5063 switch (instr->type) {
5064 case nir_instr_type_alu:
5065 visit_alu(ctx, nir_instr_as_alu(instr));
5066 break;
5067 case nir_instr_type_load_const:
5068 visit_load_const(ctx, nir_instr_as_load_const(instr));
5069 break;
5070 case nir_instr_type_intrinsic:
5071 visit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
5072 break;
5073 case nir_instr_type_tex:
5074 visit_tex(ctx, nir_instr_as_tex(instr));
5075 break;
5076 case nir_instr_type_phi:
5077 visit_phi(ctx, nir_instr_as_phi(instr));
5078 break;
5079 case nir_instr_type_ssa_undef:
5080 visit_ssa_undef(ctx, nir_instr_as_ssa_undef(instr));
5081 break;
5082 case nir_instr_type_jump:
5083 visit_jump(&ctx->ac, nir_instr_as_jump(instr));
5084 break;
5085 case nir_instr_type_deref:
5086 visit_deref(ctx, nir_instr_as_deref(instr));
5087 break;
5088 default:
5089 fprintf(stderr, "Unknown NIR instr type: ");
5090 nir_print_instr(instr, stderr);
5091 fprintf(stderr, "\n");
5092 abort();
5093 }
5094 }
5095
5096 _mesa_hash_table_insert(ctx->defs, block,
5097 LLVMGetInsertBlock(ctx->ac.builder));
5098 }
5099
5100 static void visit_if(struct ac_nir_context *ctx, nir_if *if_stmt)
5101 {
5102 LLVMValueRef value = get_src(ctx, if_stmt->condition);
5103
5104 nir_block *then_block =
5105 (nir_block *) exec_list_get_head(&if_stmt->then_list);
5106
5107 ac_build_uif(&ctx->ac, value, then_block->index);
5108
5109 visit_cf_list(ctx, &if_stmt->then_list);
5110
5111 if (!exec_list_is_empty(&if_stmt->else_list)) {
5112 nir_block *else_block =
5113 (nir_block *) exec_list_get_head(&if_stmt->else_list);
5114
5115 ac_build_else(&ctx->ac, else_block->index);
5116 visit_cf_list(ctx, &if_stmt->else_list);
5117 }
5118
5119 ac_build_endif(&ctx->ac, then_block->index);
5120 }
5121
5122 static void visit_loop(struct ac_nir_context *ctx, nir_loop *loop)
5123 {
5124 nir_block *first_loop_block =
5125 (nir_block *) exec_list_get_head(&loop->body);
5126
5127 ac_build_bgnloop(&ctx->ac, first_loop_block->index);
5128
5129 visit_cf_list(ctx, &loop->body);
5130
5131 ac_build_endloop(&ctx->ac, first_loop_block->index);
5132 }
5133
5134 static void visit_cf_list(struct ac_nir_context *ctx,
5135 struct exec_list *list)
5136 {
5137 foreach_list_typed(nir_cf_node, node, node, list)
5138 {
5139 switch (node->type) {
5140 case nir_cf_node_block:
5141 visit_block(ctx, nir_cf_node_as_block(node));
5142 break;
5143
5144 case nir_cf_node_if:
5145 visit_if(ctx, nir_cf_node_as_if(node));
5146 break;
5147
5148 case nir_cf_node_loop:
5149 visit_loop(ctx, nir_cf_node_as_loop(node));
5150 break;
5151
5152 default:
5153 assert(0);
5154 }
5155 }
5156 }
5157
5158 void
5159 ac_handle_shader_output_decl(struct ac_llvm_context *ctx,
5160 struct ac_shader_abi *abi,
5161 struct nir_shader *nir,
5162 struct nir_variable *variable,
5163 gl_shader_stage stage)
5164 {
5165 unsigned output_loc = variable->data.driver_location / 4;
5166 unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
5167
5168 /* tess ctrl has it's own load/store paths for outputs */
5169 if (stage == MESA_SHADER_TESS_CTRL)
5170 return;
5171
5172 if (stage == MESA_SHADER_VERTEX ||
5173 stage == MESA_SHADER_TESS_EVAL ||
5174 stage == MESA_SHADER_GEOMETRY) {
5175 int idx = variable->data.location + variable->data.index;
5176 if (idx == VARYING_SLOT_CLIP_DIST0) {
5177 int length = nir->info.clip_distance_array_size +
5178 nir->info.cull_distance_array_size;
5179
5180 if (length > 4)
5181 attrib_count = 2;
5182 else
5183 attrib_count = 1;
5184 }
5185 }
5186
5187 bool is_16bit = glsl_type_is_16bit(glsl_without_array(variable->type));
5188 LLVMTypeRef type = is_16bit ? ctx->f16 : ctx->f32;
5189 for (unsigned i = 0; i < attrib_count; ++i) {
5190 for (unsigned chan = 0; chan < 4; chan++) {
5191 abi->outputs[ac_llvm_reg_index_soa(output_loc + i, chan)] =
5192 ac_build_alloca_undef(ctx, type, "");
5193 }
5194 }
5195 }
5196
5197 static void
5198 setup_locals(struct ac_nir_context *ctx,
5199 struct nir_function *func)
5200 {
5201 int i, j;
5202 ctx->num_locals = 0;
5203 nir_foreach_variable(variable, &func->impl->locals) {
5204 unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
5205 variable->data.driver_location = ctx->num_locals * 4;
5206 variable->data.location_frac = 0;
5207 ctx->num_locals += attrib_count;
5208 }
5209 ctx->locals = malloc(4 * ctx->num_locals * sizeof(LLVMValueRef));
5210 if (!ctx->locals)
5211 return;
5212
5213 for (i = 0; i < ctx->num_locals; i++) {
5214 for (j = 0; j < 4; j++) {
5215 ctx->locals[i * 4 + j] =
5216 ac_build_alloca_undef(&ctx->ac, ctx->ac.f32, "temp");
5217 }
5218 }
5219 }
5220
5221 static void
5222 setup_scratch(struct ac_nir_context *ctx,
5223 struct nir_shader *shader)
5224 {
5225 if (shader->scratch_size == 0)
5226 return;
5227
5228 ctx->scratch = ac_build_alloca_undef(&ctx->ac,
5229 LLVMArrayType(ctx->ac.i8, shader->scratch_size),
5230 "scratch");
5231 }
5232
5233 static void
5234 setup_constant_data(struct ac_nir_context *ctx,
5235 struct nir_shader *shader)
5236 {
5237 if (!shader->constant_data)
5238 return;
5239
5240 LLVMValueRef data =
5241 LLVMConstStringInContext(ctx->ac.context,
5242 shader->constant_data,
5243 shader->constant_data_size,
5244 true);
5245 LLVMTypeRef type = LLVMArrayType(ctx->ac.i8, shader->constant_data_size);
5246
5247 /* We want to put the constant data in the CONST address space so that
5248 * we can use scalar loads. However, LLVM versions before 10 put these
5249 * variables in the same section as the code, which is unacceptable
5250 * for RadeonSI as it needs to relocate all the data sections after
5251 * the code sections. See https://reviews.llvm.org/D65813.
5252 */
5253 unsigned address_space =
5254 LLVM_VERSION_MAJOR < 10 ? AC_ADDR_SPACE_GLOBAL : AC_ADDR_SPACE_CONST;
5255
5256 LLVMValueRef global =
5257 LLVMAddGlobalInAddressSpace(ctx->ac.module, type,
5258 "const_data",
5259 address_space);
5260
5261 LLVMSetInitializer(global, data);
5262 LLVMSetGlobalConstant(global, true);
5263 LLVMSetVisibility(global, LLVMHiddenVisibility);
5264 ctx->constant_data = global;
5265 }
5266
5267 static void
5268 setup_shared(struct ac_nir_context *ctx,
5269 struct nir_shader *nir)
5270 {
5271 if (ctx->ac.lds)
5272 return;
5273
5274 LLVMTypeRef type = LLVMArrayType(ctx->ac.i8,
5275 nir->info.cs.shared_size);
5276
5277 LLVMValueRef lds =
5278 LLVMAddGlobalInAddressSpace(ctx->ac.module, type,
5279 "compute_lds",
5280 AC_ADDR_SPACE_LDS);
5281 LLVMSetAlignment(lds, 64 * 1024);
5282
5283 ctx->ac.lds = LLVMBuildBitCast(ctx->ac.builder, lds,
5284 LLVMPointerType(ctx->ac.i8,
5285 AC_ADDR_SPACE_LDS), "");
5286 }
5287
5288 void ac_nir_translate(struct ac_llvm_context *ac, struct ac_shader_abi *abi,
5289 const struct ac_shader_args *args, struct nir_shader *nir)
5290 {
5291 struct ac_nir_context ctx = {};
5292 struct nir_function *func;
5293
5294 ctx.ac = *ac;
5295 ctx.abi = abi;
5296 ctx.args = args;
5297
5298 ctx.stage = nir->info.stage;
5299 ctx.info = &nir->info;
5300
5301 ctx.main_function = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx.ac.builder));
5302
5303 nir_foreach_variable(variable, &nir->outputs)
5304 ac_handle_shader_output_decl(&ctx.ac, ctx.abi, nir, variable,
5305 ctx.stage);
5306
5307 ctx.defs = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
5308 _mesa_key_pointer_equal);
5309 ctx.phis = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
5310 _mesa_key_pointer_equal);
5311 ctx.vars = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
5312 _mesa_key_pointer_equal);
5313
5314 if (ctx.abi->kill_ps_if_inf_interp)
5315 ctx.verified_interp = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
5316 _mesa_key_pointer_equal);
5317
5318 func = (struct nir_function *)exec_list_get_head(&nir->functions);
5319
5320 nir_index_ssa_defs(func->impl);
5321 ctx.ssa_defs = calloc(func->impl->ssa_alloc, sizeof(LLVMValueRef));
5322
5323 setup_locals(&ctx, func);
5324 setup_scratch(&ctx, nir);
5325 setup_constant_data(&ctx, nir);
5326
5327 if (gl_shader_stage_is_compute(nir->info.stage))
5328 setup_shared(&ctx, nir);
5329
5330 if (nir->info.stage == MESA_SHADER_FRAGMENT && nir->info.fs.uses_demote) {
5331 ctx.ac.postponed_kill = ac_build_alloca_undef(&ctx.ac, ac->i1, "");
5332 /* true = don't kill. */
5333 LLVMBuildStore(ctx.ac.builder, ctx.ac.i1true, ctx.ac.postponed_kill);
5334 }
5335
5336 visit_cf_list(&ctx, &func->impl->body);
5337 phi_post_pass(&ctx);
5338
5339 if (ctx.ac.postponed_kill)
5340 ac_build_kill_if_false(&ctx.ac, LLVMBuildLoad(ctx.ac.builder,
5341 ctx.ac.postponed_kill, ""));
5342
5343 if (!gl_shader_stage_is_compute(nir->info.stage))
5344 ctx.abi->emit_outputs(ctx.abi, AC_LLVM_MAX_OUTPUTS,
5345 ctx.abi->outputs);
5346
5347 free(ctx.locals);
5348 free(ctx.ssa_defs);
5349 ralloc_free(ctx.defs);
5350 ralloc_free(ctx.phis);
5351 ralloc_free(ctx.vars);
5352 if (ctx.abi->kill_ps_if_inf_interp)
5353 ralloc_free(ctx.verified_interp);
5354 }
5355
5356 bool
5357 ac_lower_indirect_derefs(struct nir_shader *nir, enum chip_class chip_class)
5358 {
5359 bool progress = false;
5360
5361 /* Lower large variables to scratch first so that we won't bloat the
5362 * shader by generating large if ladders for them. We later lower
5363 * scratch to alloca's, assuming LLVM won't generate VGPR indexing.
5364 */
5365 NIR_PASS(progress, nir, nir_lower_vars_to_scratch,
5366 nir_var_function_temp,
5367 256,
5368 glsl_get_natural_size_align_bytes);
5369
5370 /* While it would be nice not to have this flag, we are constrained
5371 * by the reality that LLVM 9.0 has buggy VGPR indexing on GFX9.
5372 */
5373 bool llvm_has_working_vgpr_indexing = chip_class != GFX9;
5374
5375 /* TODO: Indirect indexing of GS inputs is unimplemented.
5376 *
5377 * TCS and TES load inputs directly from LDS or offchip memory, so
5378 * indirect indexing is trivial.
5379 */
5380 nir_variable_mode indirect_mask = 0;
5381 if (nir->info.stage == MESA_SHADER_GEOMETRY ||
5382 (nir->info.stage != MESA_SHADER_TESS_CTRL &&
5383 nir->info.stage != MESA_SHADER_TESS_EVAL &&
5384 !llvm_has_working_vgpr_indexing)) {
5385 indirect_mask |= nir_var_shader_in;
5386 }
5387 if (!llvm_has_working_vgpr_indexing &&
5388 nir->info.stage != MESA_SHADER_TESS_CTRL)
5389 indirect_mask |= nir_var_shader_out;
5390
5391 /* TODO: We shouldn't need to do this, however LLVM isn't currently
5392 * smart enough to handle indirects without causing excess spilling
5393 * causing the gpu to hang.
5394 *
5395 * See the following thread for more details of the problem:
5396 * https://lists.freedesktop.org/archives/mesa-dev/2017-July/162106.html
5397 */
5398 indirect_mask |= nir_var_function_temp;
5399
5400 progress |= nir_lower_indirect_derefs(nir, indirect_mask);
5401 return progress;
5402 }
5403
5404 static unsigned
5405 get_inst_tessfactor_writemask(nir_intrinsic_instr *intrin)
5406 {
5407 if (intrin->intrinsic != nir_intrinsic_store_deref)
5408 return 0;
5409
5410 nir_variable *var =
5411 nir_deref_instr_get_variable(nir_src_as_deref(intrin->src[0]));
5412
5413 if (var->data.mode != nir_var_shader_out)
5414 return 0;
5415
5416 unsigned writemask = 0;
5417 const int location = var->data.location;
5418 unsigned first_component = var->data.location_frac;
5419 unsigned num_comps = intrin->dest.ssa.num_components;
5420
5421 if (location == VARYING_SLOT_TESS_LEVEL_INNER)
5422 writemask = ((1 << (num_comps + 1)) - 1) << first_component;
5423 else if (location == VARYING_SLOT_TESS_LEVEL_OUTER)
5424 writemask = (((1 << (num_comps + 1)) - 1) << first_component) << 4;
5425
5426 return writemask;
5427 }
5428
5429 static void
5430 scan_tess_ctrl(nir_cf_node *cf_node, unsigned *upper_block_tf_writemask,
5431 unsigned *cond_block_tf_writemask,
5432 bool *tessfactors_are_def_in_all_invocs, bool is_nested_cf)
5433 {
5434 switch (cf_node->type) {
5435 case nir_cf_node_block: {
5436 nir_block *block = nir_cf_node_as_block(cf_node);
5437 nir_foreach_instr(instr, block) {
5438 if (instr->type != nir_instr_type_intrinsic)
5439 continue;
5440
5441 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
5442 if (intrin->intrinsic == nir_intrinsic_control_barrier) {
5443
5444 /* If we find a barrier in nested control flow put this in the
5445 * too hard basket. In GLSL this is not possible but it is in
5446 * SPIR-V.
5447 */
5448 if (is_nested_cf) {
5449 *tessfactors_are_def_in_all_invocs = false;
5450 return;
5451 }
5452
5453 /* The following case must be prevented:
5454 * gl_TessLevelInner = ...;
5455 * barrier();
5456 * if (gl_InvocationID == 1)
5457 * gl_TessLevelInner = ...;
5458 *
5459 * If you consider disjoint code segments separated by barriers, each
5460 * such segment that writes tess factor channels should write the same
5461 * channels in all codepaths within that segment.
5462 */
5463 if (upper_block_tf_writemask || cond_block_tf_writemask) {
5464 /* Accumulate the result: */
5465 *tessfactors_are_def_in_all_invocs &=
5466 !(*cond_block_tf_writemask & ~(*upper_block_tf_writemask));
5467
5468 /* Analyze the next code segment from scratch. */
5469 *upper_block_tf_writemask = 0;
5470 *cond_block_tf_writemask = 0;
5471 }
5472 } else
5473 *upper_block_tf_writemask |= get_inst_tessfactor_writemask(intrin);
5474 }
5475
5476 break;
5477 }
5478 case nir_cf_node_if: {
5479 unsigned then_tessfactor_writemask = 0;
5480 unsigned else_tessfactor_writemask = 0;
5481
5482 nir_if *if_stmt = nir_cf_node_as_if(cf_node);
5483 foreach_list_typed(nir_cf_node, nested_node, node, &if_stmt->then_list) {
5484 scan_tess_ctrl(nested_node, &then_tessfactor_writemask,
5485 cond_block_tf_writemask,
5486 tessfactors_are_def_in_all_invocs, true);
5487 }
5488
5489 foreach_list_typed(nir_cf_node, nested_node, node, &if_stmt->else_list) {
5490 scan_tess_ctrl(nested_node, &else_tessfactor_writemask,
5491 cond_block_tf_writemask,
5492 tessfactors_are_def_in_all_invocs, true);
5493 }
5494
5495 if (then_tessfactor_writemask || else_tessfactor_writemask) {
5496 /* If both statements write the same tess factor channels,
5497 * we can say that the upper block writes them too.
5498 */
5499 *upper_block_tf_writemask |= then_tessfactor_writemask &
5500 else_tessfactor_writemask;
5501 *cond_block_tf_writemask |= then_tessfactor_writemask |
5502 else_tessfactor_writemask;
5503 }
5504
5505 break;
5506 }
5507 case nir_cf_node_loop: {
5508 nir_loop *loop = nir_cf_node_as_loop(cf_node);
5509 foreach_list_typed(nir_cf_node, nested_node, node, &loop->body) {
5510 scan_tess_ctrl(nested_node, cond_block_tf_writemask,
5511 cond_block_tf_writemask,
5512 tessfactors_are_def_in_all_invocs, true);
5513 }
5514
5515 break;
5516 }
5517 default:
5518 unreachable("unknown cf node type");
5519 }
5520 }
5521
5522 bool
5523 ac_are_tessfactors_def_in_all_invocs(const struct nir_shader *nir)
5524 {
5525 assert(nir->info.stage == MESA_SHADER_TESS_CTRL);
5526
5527 /* The pass works as follows:
5528 * If all codepaths write tess factors, we can say that all
5529 * invocations define tess factors.
5530 *
5531 * Each tess factor channel is tracked separately.
5532 */
5533 unsigned main_block_tf_writemask = 0; /* if main block writes tess factors */
5534 unsigned cond_block_tf_writemask = 0; /* if cond block writes tess factors */
5535
5536 /* Initial value = true. Here the pass will accumulate results from
5537 * multiple segments surrounded by barriers. If tess factors aren't
5538 * written at all, it's a shader bug and we don't care if this will be
5539 * true.
5540 */
5541 bool tessfactors_are_def_in_all_invocs = true;
5542
5543 nir_foreach_function(function, nir) {
5544 if (function->impl) {
5545 foreach_list_typed(nir_cf_node, node, node, &function->impl->body) {
5546 scan_tess_ctrl(node, &main_block_tf_writemask,
5547 &cond_block_tf_writemask,
5548 &tessfactors_are_def_in_all_invocs,
5549 false);
5550 }
5551 }
5552 }
5553
5554 /* Accumulate the result for the last code segment separated by a
5555 * barrier.
5556 */
5557 if (main_block_tf_writemask || cond_block_tf_writemask) {
5558 tessfactors_are_def_in_all_invocs &=
5559 !(cond_block_tf_writemask & ~main_block_tf_writemask);
5560 }
5561
5562 return tessfactors_are_def_in_all_invocs;
5563 }