ac/nir: support 16-bit data in buffer_load_format opcodes
[mesa.git] / src / amd / llvm / ac_nir_to_llvm.c
1 /*
2 * Copyright © 2016 Bas Nieuwenhuizen
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include <llvm/Config/llvm-config.h>
25
26 #include "ac_nir_to_llvm.h"
27 #include "ac_llvm_build.h"
28 #include "ac_llvm_util.h"
29 #include "ac_binary.h"
30 #include "sid.h"
31 #include "nir/nir.h"
32 #include "nir/nir_deref.h"
33 #include "util/bitscan.h"
34 #include "util/u_math.h"
35 #include "ac_shader_abi.h"
36 #include "ac_shader_util.h"
37
38 struct ac_nir_context {
39 struct ac_llvm_context ac;
40 struct ac_shader_abi *abi;
41 const struct ac_shader_args *args;
42
43 gl_shader_stage stage;
44 shader_info *info;
45
46 LLVMValueRef *ssa_defs;
47
48 LLVMValueRef scratch;
49 LLVMValueRef constant_data;
50
51 struct hash_table *defs;
52 struct hash_table *phis;
53 struct hash_table *vars;
54 struct hash_table *verified_interp;
55
56 LLVMValueRef main_function;
57 LLVMBasicBlockRef continue_block;
58 LLVMBasicBlockRef break_block;
59
60 int num_locals;
61 LLVMValueRef *locals;
62 };
63
64 static LLVMValueRef get_sampler_desc_index(struct ac_nir_context *ctx,
65 nir_deref_instr *deref_instr,
66 const nir_instr *instr,
67 bool image);
68
69 static LLVMValueRef get_sampler_desc(struct ac_nir_context *ctx,
70 nir_deref_instr *deref_instr,
71 enum ac_descriptor_type desc_type,
72 const nir_instr *instr,
73 LLVMValueRef index,
74 bool image, bool write);
75
76 static void
77 build_store_values_extended(struct ac_llvm_context *ac,
78 LLVMValueRef *values,
79 unsigned value_count,
80 unsigned value_stride,
81 LLVMValueRef vec)
82 {
83 LLVMBuilderRef builder = ac->builder;
84 unsigned i;
85
86 for (i = 0; i < value_count; i++) {
87 LLVMValueRef ptr = values[i * value_stride];
88 LLVMValueRef index = LLVMConstInt(ac->i32, i, false);
89 LLVMValueRef value = LLVMBuildExtractElement(builder, vec, index, "");
90 LLVMBuildStore(builder, value, ptr);
91 }
92 }
93
94 static LLVMTypeRef get_def_type(struct ac_nir_context *ctx,
95 const nir_ssa_def *def)
96 {
97 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, def->bit_size);
98 if (def->num_components > 1) {
99 type = LLVMVectorType(type, def->num_components);
100 }
101 return type;
102 }
103
104 static LLVMValueRef get_src(struct ac_nir_context *nir, nir_src src)
105 {
106 assert(src.is_ssa);
107 return nir->ssa_defs[src.ssa->index];
108 }
109
110 static LLVMValueRef
111 get_memory_ptr(struct ac_nir_context *ctx, nir_src src, unsigned bit_size)
112 {
113 LLVMValueRef ptr = get_src(ctx, src);
114 ptr = LLVMBuildGEP(ctx->ac.builder, ctx->ac.lds, &ptr, 1, "");
115 int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
116
117 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, bit_size);
118
119 return LLVMBuildBitCast(ctx->ac.builder, ptr,
120 LLVMPointerType(type, addr_space), "");
121 }
122
123 static LLVMBasicBlockRef get_block(struct ac_nir_context *nir,
124 const struct nir_block *b)
125 {
126 struct hash_entry *entry = _mesa_hash_table_search(nir->defs, b);
127 return (LLVMBasicBlockRef)entry->data;
128 }
129
130 static LLVMValueRef get_alu_src(struct ac_nir_context *ctx,
131 nir_alu_src src,
132 unsigned num_components)
133 {
134 LLVMValueRef value = get_src(ctx, src.src);
135 bool need_swizzle = false;
136
137 assert(value);
138 unsigned src_components = ac_get_llvm_num_components(value);
139 for (unsigned i = 0; i < num_components; ++i) {
140 assert(src.swizzle[i] < src_components);
141 if (src.swizzle[i] != i)
142 need_swizzle = true;
143 }
144
145 if (need_swizzle || num_components != src_components) {
146 LLVMValueRef masks[] = {
147 LLVMConstInt(ctx->ac.i32, src.swizzle[0], false),
148 LLVMConstInt(ctx->ac.i32, src.swizzle[1], false),
149 LLVMConstInt(ctx->ac.i32, src.swizzle[2], false),
150 LLVMConstInt(ctx->ac.i32, src.swizzle[3], false)};
151
152 if (src_components > 1 && num_components == 1) {
153 value = LLVMBuildExtractElement(ctx->ac.builder, value,
154 masks[0], "");
155 } else if (src_components == 1 && num_components > 1) {
156 LLVMValueRef values[] = {value, value, value, value};
157 value = ac_build_gather_values(&ctx->ac, values, num_components);
158 } else {
159 LLVMValueRef swizzle = LLVMConstVector(masks, num_components);
160 value = LLVMBuildShuffleVector(ctx->ac.builder, value, value,
161 swizzle, "");
162 }
163 }
164 assert(!src.negate);
165 assert(!src.abs);
166 return value;
167 }
168
169 static LLVMValueRef emit_int_cmp(struct ac_llvm_context *ctx,
170 LLVMIntPredicate pred, LLVMValueRef src0,
171 LLVMValueRef src1)
172 {
173 LLVMValueRef result = LLVMBuildICmp(ctx->builder, pred, src0, src1, "");
174 return LLVMBuildSelect(ctx->builder, result,
175 LLVMConstInt(ctx->i32, 0xFFFFFFFF, false),
176 ctx->i32_0, "");
177 }
178
179 static LLVMValueRef emit_float_cmp(struct ac_llvm_context *ctx,
180 LLVMRealPredicate pred, LLVMValueRef src0,
181 LLVMValueRef src1)
182 {
183 LLVMValueRef result;
184 src0 = ac_to_float(ctx, src0);
185 src1 = ac_to_float(ctx, src1);
186 result = LLVMBuildFCmp(ctx->builder, pred, src0, src1, "");
187 return LLVMBuildSelect(ctx->builder, result,
188 LLVMConstInt(ctx->i32, 0xFFFFFFFF, false),
189 ctx->i32_0, "");
190 }
191
192 static LLVMValueRef emit_intrin_1f_param(struct ac_llvm_context *ctx,
193 const char *intrin,
194 LLVMTypeRef result_type,
195 LLVMValueRef src0)
196 {
197 char name[64], type[64];
198 LLVMValueRef params[] = {
199 ac_to_float(ctx, src0),
200 };
201
202 ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
203 ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
204 assert(length < sizeof(name));
205 return ac_build_intrinsic(ctx, name, result_type, params, 1, AC_FUNC_ATTR_READNONE);
206 }
207
208 static LLVMValueRef emit_intrin_2f_param(struct ac_llvm_context *ctx,
209 const char *intrin,
210 LLVMTypeRef result_type,
211 LLVMValueRef src0, LLVMValueRef src1)
212 {
213 char name[64], type[64];
214 LLVMValueRef params[] = {
215 ac_to_float(ctx, src0),
216 ac_to_float(ctx, src1),
217 };
218
219 ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
220 ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
221 assert(length < sizeof(name));
222 return ac_build_intrinsic(ctx, name, result_type, params, 2, AC_FUNC_ATTR_READNONE);
223 }
224
225 static LLVMValueRef emit_intrin_3f_param(struct ac_llvm_context *ctx,
226 const char *intrin,
227 LLVMTypeRef result_type,
228 LLVMValueRef src0, LLVMValueRef src1, LLVMValueRef src2)
229 {
230 char name[64], type[64];
231 LLVMValueRef params[] = {
232 ac_to_float(ctx, src0),
233 ac_to_float(ctx, src1),
234 ac_to_float(ctx, src2),
235 };
236
237 ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
238 ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
239 assert(length < sizeof(name));
240 return ac_build_intrinsic(ctx, name, result_type, params, 3, AC_FUNC_ATTR_READNONE);
241 }
242
243 static LLVMValueRef emit_bcsel(struct ac_llvm_context *ctx,
244 LLVMValueRef src0, LLVMValueRef src1, LLVMValueRef src2)
245 {
246 LLVMTypeRef src1_type = LLVMTypeOf(src1);
247 LLVMTypeRef src2_type = LLVMTypeOf(src2);
248
249 assert(LLVMGetTypeKind(LLVMTypeOf(src0)) != LLVMVectorTypeKind);
250
251 if (LLVMGetTypeKind(src1_type) == LLVMPointerTypeKind &&
252 LLVMGetTypeKind(src2_type) != LLVMPointerTypeKind) {
253 src2 = LLVMBuildIntToPtr(ctx->builder, src2, src1_type, "");
254 } else if (LLVMGetTypeKind(src2_type) == LLVMPointerTypeKind &&
255 LLVMGetTypeKind(src1_type) != LLVMPointerTypeKind) {
256 src1 = LLVMBuildIntToPtr(ctx->builder, src1, src2_type, "");
257 }
258
259 LLVMValueRef v = LLVMBuildICmp(ctx->builder, LLVMIntNE, src0,
260 ctx->i32_0, "");
261 return LLVMBuildSelect(ctx->builder, v,
262 ac_to_integer_or_pointer(ctx, src1),
263 ac_to_integer_or_pointer(ctx, src2), "");
264 }
265
266 static LLVMValueRef emit_iabs(struct ac_llvm_context *ctx,
267 LLVMValueRef src0)
268 {
269 return ac_build_imax(ctx, src0, LLVMBuildNeg(ctx->builder, src0, ""));
270 }
271
272 static LLVMValueRef emit_uint_carry(struct ac_llvm_context *ctx,
273 const char *intrin,
274 LLVMValueRef src0, LLVMValueRef src1)
275 {
276 LLVMTypeRef ret_type;
277 LLVMTypeRef types[] = { ctx->i32, ctx->i1 };
278 LLVMValueRef res;
279 LLVMValueRef params[] = { src0, src1 };
280 ret_type = LLVMStructTypeInContext(ctx->context, types,
281 2, true);
282
283 res = ac_build_intrinsic(ctx, intrin, ret_type,
284 params, 2, AC_FUNC_ATTR_READNONE);
285
286 res = LLVMBuildExtractValue(ctx->builder, res, 1, "");
287 res = LLVMBuildZExt(ctx->builder, res, ctx->i32, "");
288 return res;
289 }
290
291 static LLVMValueRef emit_b2f(struct ac_llvm_context *ctx,
292 LLVMValueRef src0,
293 unsigned bitsize)
294 {
295 LLVMValueRef result = LLVMBuildAnd(ctx->builder, src0,
296 LLVMBuildBitCast(ctx->builder, LLVMConstReal(ctx->f32, 1.0), ctx->i32, ""),
297 "");
298 result = LLVMBuildBitCast(ctx->builder, result, ctx->f32, "");
299
300 switch (bitsize) {
301 case 16:
302 return LLVMBuildFPTrunc(ctx->builder, result, ctx->f16, "");
303 case 32:
304 return result;
305 case 64:
306 return LLVMBuildFPExt(ctx->builder, result, ctx->f64, "");
307 default:
308 unreachable("Unsupported bit size.");
309 }
310 }
311
312 static LLVMValueRef emit_f2b(struct ac_llvm_context *ctx,
313 LLVMValueRef src0)
314 {
315 src0 = ac_to_float(ctx, src0);
316 LLVMValueRef zero = LLVMConstNull(LLVMTypeOf(src0));
317 return LLVMBuildSExt(ctx->builder,
318 LLVMBuildFCmp(ctx->builder, LLVMRealUNE, src0, zero, ""),
319 ctx->i32, "");
320 }
321
322 static LLVMValueRef emit_b2i(struct ac_llvm_context *ctx,
323 LLVMValueRef src0,
324 unsigned bitsize)
325 {
326 LLVMValueRef result = LLVMBuildAnd(ctx->builder, src0, ctx->i32_1, "");
327
328 switch (bitsize) {
329 case 8:
330 return LLVMBuildTrunc(ctx->builder, result, ctx->i8, "");
331 case 16:
332 return LLVMBuildTrunc(ctx->builder, result, ctx->i16, "");
333 case 32:
334 return result;
335 case 64:
336 return LLVMBuildZExt(ctx->builder, result, ctx->i64, "");
337 default:
338 unreachable("Unsupported bit size.");
339 }
340 }
341
342 static LLVMValueRef emit_i2b(struct ac_llvm_context *ctx,
343 LLVMValueRef src0)
344 {
345 LLVMValueRef zero = LLVMConstNull(LLVMTypeOf(src0));
346 return LLVMBuildSExt(ctx->builder,
347 LLVMBuildICmp(ctx->builder, LLVMIntNE, src0, zero, ""),
348 ctx->i32, "");
349 }
350
351 static LLVMValueRef emit_f2f16(struct ac_llvm_context *ctx,
352 LLVMValueRef src0)
353 {
354 LLVMValueRef result;
355 LLVMValueRef cond = NULL;
356
357 src0 = ac_to_float(ctx, src0);
358 result = LLVMBuildFPTrunc(ctx->builder, src0, ctx->f16, "");
359
360 if (ctx->chip_class >= GFX8) {
361 LLVMValueRef args[2];
362 /* Check if the result is a denormal - and flush to 0 if so. */
363 args[0] = result;
364 args[1] = LLVMConstInt(ctx->i32, N_SUBNORMAL | P_SUBNORMAL, false);
365 cond = ac_build_intrinsic(ctx, "llvm.amdgcn.class.f16", ctx->i1, args, 2, AC_FUNC_ATTR_READNONE);
366 }
367
368 /* need to convert back up to f32 */
369 result = LLVMBuildFPExt(ctx->builder, result, ctx->f32, "");
370
371 if (ctx->chip_class >= GFX8)
372 result = LLVMBuildSelect(ctx->builder, cond, ctx->f32_0, result, "");
373 else {
374 /* for GFX6-GFX7 */
375 /* 0x38800000 is smallest half float value (2^-14) in 32-bit float,
376 * so compare the result and flush to 0 if it's smaller.
377 */
378 LLVMValueRef temp, cond2;
379 temp = emit_intrin_1f_param(ctx, "llvm.fabs", ctx->f32, result);
380 cond = LLVMBuildFCmp(ctx->builder, LLVMRealOGT,
381 LLVMBuildBitCast(ctx->builder, LLVMConstInt(ctx->i32, 0x38800000, false), ctx->f32, ""),
382 temp, "");
383 cond2 = LLVMBuildFCmp(ctx->builder, LLVMRealONE,
384 temp, ctx->f32_0, "");
385 cond = LLVMBuildAnd(ctx->builder, cond, cond2, "");
386 result = LLVMBuildSelect(ctx->builder, cond, ctx->f32_0, result, "");
387 }
388 return result;
389 }
390
391 static LLVMValueRef emit_umul_high(struct ac_llvm_context *ctx,
392 LLVMValueRef src0, LLVMValueRef src1)
393 {
394 LLVMValueRef dst64, result;
395 src0 = LLVMBuildZExt(ctx->builder, src0, ctx->i64, "");
396 src1 = LLVMBuildZExt(ctx->builder, src1, ctx->i64, "");
397
398 dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
399 dst64 = LLVMBuildLShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
400 result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
401 return result;
402 }
403
404 static LLVMValueRef emit_imul_high(struct ac_llvm_context *ctx,
405 LLVMValueRef src0, LLVMValueRef src1)
406 {
407 LLVMValueRef dst64, result;
408 src0 = LLVMBuildSExt(ctx->builder, src0, ctx->i64, "");
409 src1 = LLVMBuildSExt(ctx->builder, src1, ctx->i64, "");
410
411 dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
412 dst64 = LLVMBuildAShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
413 result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
414 return result;
415 }
416
417 static LLVMValueRef emit_bfm(struct ac_llvm_context *ctx,
418 LLVMValueRef bits, LLVMValueRef offset)
419 {
420 /* mask = ((1 << bits) - 1) << offset */
421 return LLVMBuildShl(ctx->builder,
422 LLVMBuildSub(ctx->builder,
423 LLVMBuildShl(ctx->builder,
424 ctx->i32_1,
425 bits, ""),
426 ctx->i32_1, ""),
427 offset, "");
428 }
429
430 static LLVMValueRef emit_bitfield_select(struct ac_llvm_context *ctx,
431 LLVMValueRef mask, LLVMValueRef insert,
432 LLVMValueRef base)
433 {
434 /* Calculate:
435 * (mask & insert) | (~mask & base) = base ^ (mask & (insert ^ base))
436 * Use the right-hand side, which the LLVM backend can convert to V_BFI.
437 */
438 return LLVMBuildXor(ctx->builder, base,
439 LLVMBuildAnd(ctx->builder, mask,
440 LLVMBuildXor(ctx->builder, insert, base, ""), ""), "");
441 }
442
443 static LLVMValueRef emit_pack_2x16(struct ac_llvm_context *ctx,
444 LLVMValueRef src0,
445 LLVMValueRef (*pack)(struct ac_llvm_context *ctx,
446 LLVMValueRef args[2]))
447 {
448 LLVMValueRef comp[2];
449
450 src0 = ac_to_float(ctx, src0);
451 comp[0] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_0, "");
452 comp[1] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_1, "");
453
454 return LLVMBuildBitCast(ctx->builder, pack(ctx, comp), ctx->i32, "");
455 }
456
457 static LLVMValueRef emit_unpack_half_2x16(struct ac_llvm_context *ctx,
458 LLVMValueRef src0)
459 {
460 LLVMValueRef const16 = LLVMConstInt(ctx->i32, 16, false);
461 LLVMValueRef temps[2], val;
462 int i;
463
464 for (i = 0; i < 2; i++) {
465 val = i == 1 ? LLVMBuildLShr(ctx->builder, src0, const16, "") : src0;
466 val = LLVMBuildTrunc(ctx->builder, val, ctx->i16, "");
467 val = LLVMBuildBitCast(ctx->builder, val, ctx->f16, "");
468 temps[i] = LLVMBuildFPExt(ctx->builder, val, ctx->f32, "");
469 }
470 return ac_build_gather_values(ctx, temps, 2);
471 }
472
473 static LLVMValueRef emit_ddxy(struct ac_nir_context *ctx,
474 nir_op op,
475 LLVMValueRef src0)
476 {
477 unsigned mask;
478 int idx;
479 LLVMValueRef result;
480
481 if (op == nir_op_fddx_fine)
482 mask = AC_TID_MASK_LEFT;
483 else if (op == nir_op_fddy_fine)
484 mask = AC_TID_MASK_TOP;
485 else
486 mask = AC_TID_MASK_TOP_LEFT;
487
488 /* for DDX we want to next X pixel, DDY next Y pixel. */
489 if (op == nir_op_fddx_fine ||
490 op == nir_op_fddx_coarse ||
491 op == nir_op_fddx)
492 idx = 1;
493 else
494 idx = 2;
495
496 result = ac_build_ddxy(&ctx->ac, mask, idx, src0);
497 return result;
498 }
499
500 struct waterfall_context {
501 LLVMBasicBlockRef phi_bb[2];
502 bool use_waterfall;
503 };
504
505 /* To deal with divergent descriptors we can create a loop that handles all
506 * lanes with the same descriptor on a given iteration (henceforth a
507 * waterfall loop).
508 *
509 * These helper create the begin and end of the loop leaving the caller
510 * to implement the body.
511 *
512 * params:
513 * - ctx is the usal nir context
514 * - wctx is a temporary struct containing some loop info. Can be left uninitialized.
515 * - value is the possibly divergent value for which we built the loop
516 * - divergent is whether value is actually divergent. If false we just pass
517 * things through.
518 */
519 static LLVMValueRef enter_waterfall(struct ac_nir_context *ctx,
520 struct waterfall_context *wctx,
521 LLVMValueRef value, bool divergent)
522 {
523 /* If the app claims the value is divergent but it is constant we can
524 * end up with a dynamic index of NULL. */
525 if (!value)
526 divergent = false;
527
528 wctx->use_waterfall = divergent;
529 if (!divergent)
530 return value;
531
532 ac_build_bgnloop(&ctx->ac, 6000);
533
534 LLVMValueRef scalar_value = ac_build_readlane(&ctx->ac, value, NULL);
535
536 LLVMValueRef active = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, value,
537 scalar_value, "uniform_active");
538
539 wctx->phi_bb[0] = LLVMGetInsertBlock(ctx->ac.builder);
540 ac_build_ifcc(&ctx->ac, active, 6001);
541
542 return scalar_value;
543 }
544
545 static LLVMValueRef exit_waterfall(struct ac_nir_context *ctx,
546 struct waterfall_context *wctx,
547 LLVMValueRef value)
548 {
549 LLVMValueRef ret = NULL;
550 LLVMValueRef phi_src[2];
551 LLVMValueRef cc_phi_src[2] = {
552 LLVMConstInt(ctx->ac.i32, 0, false),
553 LLVMConstInt(ctx->ac.i32, 0xffffffff, false),
554 };
555
556 if (!wctx->use_waterfall)
557 return value;
558
559 wctx->phi_bb[1] = LLVMGetInsertBlock(ctx->ac.builder);
560
561 ac_build_endif(&ctx->ac, 6001);
562
563 if (value) {
564 phi_src[0] = LLVMGetUndef(LLVMTypeOf(value));
565 phi_src[1] = value;
566
567 ret = ac_build_phi(&ctx->ac, LLVMTypeOf(value), 2, phi_src, wctx->phi_bb);
568 }
569
570 /*
571 * By using the optimization barrier on the exit decision, we decouple
572 * the operations from the break, and hence avoid LLVM hoisting the
573 * opteration into the break block.
574 */
575 LLVMValueRef cc = ac_build_phi(&ctx->ac, ctx->ac.i32, 2, cc_phi_src, wctx->phi_bb);
576 ac_build_optimization_barrier(&ctx->ac, &cc);
577
578 LLVMValueRef active = LLVMBuildICmp(ctx->ac.builder, LLVMIntNE, cc, ctx->ac.i32_0, "uniform_active2");
579 ac_build_ifcc(&ctx->ac, active, 6002);
580 ac_build_break(&ctx->ac);
581 ac_build_endif(&ctx->ac, 6002);
582
583 ac_build_endloop(&ctx->ac, 6000);
584 return ret;
585 }
586
587 static void visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
588 {
589 LLVMValueRef src[4], result = NULL;
590 unsigned num_components = instr->dest.dest.ssa.num_components;
591 unsigned src_components;
592 LLVMTypeRef def_type = get_def_type(ctx, &instr->dest.dest.ssa);
593 bool saved_inexact = false;
594
595 if (instr->exact)
596 saved_inexact = ac_disable_inexact_math(ctx->ac.builder);
597
598 assert(nir_op_infos[instr->op].num_inputs <= ARRAY_SIZE(src));
599 switch (instr->op) {
600 case nir_op_vec2:
601 case nir_op_vec3:
602 case nir_op_vec4:
603 src_components = 1;
604 break;
605 case nir_op_pack_half_2x16:
606 case nir_op_pack_snorm_2x16:
607 case nir_op_pack_unorm_2x16:
608 src_components = 2;
609 break;
610 case nir_op_unpack_half_2x16:
611 src_components = 1;
612 break;
613 case nir_op_cube_face_coord:
614 case nir_op_cube_face_index:
615 src_components = 3;
616 break;
617 default:
618 src_components = num_components;
619 break;
620 }
621 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
622 src[i] = get_alu_src(ctx, instr->src[i], src_components);
623
624 switch (instr->op) {
625 case nir_op_mov:
626 result = src[0];
627 break;
628 case nir_op_fneg:
629 src[0] = ac_to_float(&ctx->ac, src[0]);
630 result = LLVMBuildFNeg(ctx->ac.builder, src[0], "");
631 if (ctx->ac.float_mode == AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO) {
632 /* fneg will be optimized by backend compiler with sign
633 * bit removed via XOR. This is probably a LLVM bug.
634 */
635 result = ac_build_canonicalize(&ctx->ac, result,
636 instr->dest.dest.ssa.bit_size);
637 }
638 break;
639 case nir_op_ineg:
640 result = LLVMBuildNeg(ctx->ac.builder, src[0], "");
641 break;
642 case nir_op_inot:
643 result = LLVMBuildNot(ctx->ac.builder, src[0], "");
644 break;
645 case nir_op_iadd:
646 result = LLVMBuildAdd(ctx->ac.builder, src[0], src[1], "");
647 break;
648 case nir_op_fadd:
649 src[0] = ac_to_float(&ctx->ac, src[0]);
650 src[1] = ac_to_float(&ctx->ac, src[1]);
651 result = LLVMBuildFAdd(ctx->ac.builder, src[0], src[1], "");
652 break;
653 case nir_op_fsub:
654 src[0] = ac_to_float(&ctx->ac, src[0]);
655 src[1] = ac_to_float(&ctx->ac, src[1]);
656 result = LLVMBuildFSub(ctx->ac.builder, src[0], src[1], "");
657 break;
658 case nir_op_isub:
659 result = LLVMBuildSub(ctx->ac.builder, src[0], src[1], "");
660 break;
661 case nir_op_imul:
662 result = LLVMBuildMul(ctx->ac.builder, src[0], src[1], "");
663 break;
664 case nir_op_imod:
665 result = LLVMBuildSRem(ctx->ac.builder, src[0], src[1], "");
666 break;
667 case nir_op_umod:
668 result = LLVMBuildURem(ctx->ac.builder, src[0], src[1], "");
669 break;
670 case nir_op_fmod:
671 /* lower_fmod only lower 16-bit and 32-bit fmod */
672 assert(instr->dest.dest.ssa.bit_size == 64);
673 src[0] = ac_to_float(&ctx->ac, src[0]);
674 src[1] = ac_to_float(&ctx->ac, src[1]);
675 result = ac_build_fdiv(&ctx->ac, src[0], src[1]);
676 result = emit_intrin_1f_param(&ctx->ac, "llvm.floor",
677 ac_to_float_type(&ctx->ac, def_type), result);
678 result = LLVMBuildFMul(ctx->ac.builder, src[1] , result, "");
679 result = LLVMBuildFSub(ctx->ac.builder, src[0], result, "");
680 break;
681 case nir_op_irem:
682 result = LLVMBuildSRem(ctx->ac.builder, src[0], src[1], "");
683 break;
684 case nir_op_idiv:
685 result = LLVMBuildSDiv(ctx->ac.builder, src[0], src[1], "");
686 break;
687 case nir_op_udiv:
688 result = LLVMBuildUDiv(ctx->ac.builder, src[0], src[1], "");
689 break;
690 case nir_op_fmul:
691 src[0] = ac_to_float(&ctx->ac, src[0]);
692 src[1] = ac_to_float(&ctx->ac, src[1]);
693 result = LLVMBuildFMul(ctx->ac.builder, src[0], src[1], "");
694 break;
695 case nir_op_frcp:
696 result = emit_intrin_1f_param(&ctx->ac, "llvm.amdgcn.rcp",
697 ac_to_float_type(&ctx->ac, def_type), src[0]);
698 break;
699 case nir_op_iand:
700 result = LLVMBuildAnd(ctx->ac.builder, src[0], src[1], "");
701 break;
702 case nir_op_ior:
703 result = LLVMBuildOr(ctx->ac.builder, src[0], src[1], "");
704 break;
705 case nir_op_ixor:
706 result = LLVMBuildXor(ctx->ac.builder, src[0], src[1], "");
707 break;
708 case nir_op_ishl:
709 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) < ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
710 src[1] = LLVMBuildZExt(ctx->ac.builder, src[1],
711 LLVMTypeOf(src[0]), "");
712 else if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) > ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
713 src[1] = LLVMBuildTrunc(ctx->ac.builder, src[1],
714 LLVMTypeOf(src[0]), "");
715 result = LLVMBuildShl(ctx->ac.builder, src[0], src[1], "");
716 break;
717 case nir_op_ishr:
718 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) < ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
719 src[1] = LLVMBuildZExt(ctx->ac.builder, src[1],
720 LLVMTypeOf(src[0]), "");
721 else if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) > ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
722 src[1] = LLVMBuildTrunc(ctx->ac.builder, src[1],
723 LLVMTypeOf(src[0]), "");
724 result = LLVMBuildAShr(ctx->ac.builder, src[0], src[1], "");
725 break;
726 case nir_op_ushr:
727 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) < ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
728 src[1] = LLVMBuildZExt(ctx->ac.builder, src[1],
729 LLVMTypeOf(src[0]), "");
730 else if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) > ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
731 src[1] = LLVMBuildTrunc(ctx->ac.builder, src[1],
732 LLVMTypeOf(src[0]), "");
733 result = LLVMBuildLShr(ctx->ac.builder, src[0], src[1], "");
734 break;
735 case nir_op_ilt32:
736 result = emit_int_cmp(&ctx->ac, LLVMIntSLT, src[0], src[1]);
737 break;
738 case nir_op_ine32:
739 result = emit_int_cmp(&ctx->ac, LLVMIntNE, src[0], src[1]);
740 break;
741 case nir_op_ieq32:
742 result = emit_int_cmp(&ctx->ac, LLVMIntEQ, src[0], src[1]);
743 break;
744 case nir_op_ige32:
745 result = emit_int_cmp(&ctx->ac, LLVMIntSGE, src[0], src[1]);
746 break;
747 case nir_op_ult32:
748 result = emit_int_cmp(&ctx->ac, LLVMIntULT, src[0], src[1]);
749 break;
750 case nir_op_uge32:
751 result = emit_int_cmp(&ctx->ac, LLVMIntUGE, src[0], src[1]);
752 break;
753 case nir_op_feq32:
754 result = emit_float_cmp(&ctx->ac, LLVMRealOEQ, src[0], src[1]);
755 break;
756 case nir_op_fne32:
757 result = emit_float_cmp(&ctx->ac, LLVMRealUNE, src[0], src[1]);
758 break;
759 case nir_op_flt32:
760 result = emit_float_cmp(&ctx->ac, LLVMRealOLT, src[0], src[1]);
761 break;
762 case nir_op_fge32:
763 result = emit_float_cmp(&ctx->ac, LLVMRealOGE, src[0], src[1]);
764 break;
765 case nir_op_fabs:
766 result = emit_intrin_1f_param(&ctx->ac, "llvm.fabs",
767 ac_to_float_type(&ctx->ac, def_type), src[0]);
768 if (ctx->ac.float_mode == AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO) {
769 /* fabs will be optimized by backend compiler with sign
770 * bit removed via AND.
771 */
772 result = ac_build_canonicalize(&ctx->ac, result,
773 instr->dest.dest.ssa.bit_size);
774 }
775 break;
776 case nir_op_iabs:
777 result = emit_iabs(&ctx->ac, src[0]);
778 break;
779 case nir_op_imax:
780 result = ac_build_imax(&ctx->ac, src[0], src[1]);
781 break;
782 case nir_op_imin:
783 result = ac_build_imin(&ctx->ac, src[0], src[1]);
784 break;
785 case nir_op_umax:
786 result = ac_build_umax(&ctx->ac, src[0], src[1]);
787 break;
788 case nir_op_umin:
789 result = ac_build_umin(&ctx->ac, src[0], src[1]);
790 break;
791 case nir_op_isign:
792 result = ac_build_isign(&ctx->ac, src[0],
793 instr->dest.dest.ssa.bit_size);
794 break;
795 case nir_op_fsign:
796 src[0] = ac_to_float(&ctx->ac, src[0]);
797 result = ac_build_fsign(&ctx->ac, src[0],
798 instr->dest.dest.ssa.bit_size);
799 break;
800 case nir_op_ffloor:
801 result = emit_intrin_1f_param(&ctx->ac, "llvm.floor",
802 ac_to_float_type(&ctx->ac, def_type), src[0]);
803 break;
804 case nir_op_ftrunc:
805 result = emit_intrin_1f_param(&ctx->ac, "llvm.trunc",
806 ac_to_float_type(&ctx->ac, def_type), src[0]);
807 break;
808 case nir_op_fceil:
809 result = emit_intrin_1f_param(&ctx->ac, "llvm.ceil",
810 ac_to_float_type(&ctx->ac, def_type), src[0]);
811 break;
812 case nir_op_fround_even:
813 result = emit_intrin_1f_param(&ctx->ac, "llvm.rint",
814 ac_to_float_type(&ctx->ac, def_type),src[0]);
815 break;
816 case nir_op_ffract:
817 src[0] = ac_to_float(&ctx->ac, src[0]);
818 result = ac_build_fract(&ctx->ac, src[0],
819 instr->dest.dest.ssa.bit_size);
820 break;
821 case nir_op_fsin:
822 result = emit_intrin_1f_param(&ctx->ac, "llvm.sin",
823 ac_to_float_type(&ctx->ac, def_type), src[0]);
824 break;
825 case nir_op_fcos:
826 result = emit_intrin_1f_param(&ctx->ac, "llvm.cos",
827 ac_to_float_type(&ctx->ac, def_type), src[0]);
828 break;
829 case nir_op_fsqrt:
830 result = emit_intrin_1f_param(&ctx->ac, "llvm.sqrt",
831 ac_to_float_type(&ctx->ac, def_type), src[0]);
832 break;
833 case nir_op_fexp2:
834 result = emit_intrin_1f_param(&ctx->ac, "llvm.exp2",
835 ac_to_float_type(&ctx->ac, def_type), src[0]);
836 break;
837 case nir_op_flog2:
838 result = emit_intrin_1f_param(&ctx->ac, "llvm.log2",
839 ac_to_float_type(&ctx->ac, def_type), src[0]);
840 break;
841 case nir_op_frsq:
842 result = emit_intrin_1f_param(&ctx->ac, "llvm.amdgcn.rsq",
843 ac_to_float_type(&ctx->ac, def_type), src[0]);
844 break;
845 case nir_op_frexp_exp:
846 src[0] = ac_to_float(&ctx->ac, src[0]);
847 result = ac_build_frexp_exp(&ctx->ac, src[0],
848 ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])));
849 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) == 16)
850 result = LLVMBuildSExt(ctx->ac.builder, result,
851 ctx->ac.i32, "");
852 break;
853 case nir_op_frexp_sig:
854 src[0] = ac_to_float(&ctx->ac, src[0]);
855 result = ac_build_frexp_mant(&ctx->ac, src[0],
856 instr->dest.dest.ssa.bit_size);
857 break;
858 case nir_op_fpow:
859 result = emit_intrin_2f_param(&ctx->ac, "llvm.pow",
860 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
861 break;
862 case nir_op_fmax:
863 result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
864 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
865 if (ctx->ac.chip_class < GFX9 &&
866 instr->dest.dest.ssa.bit_size == 32) {
867 /* Only pre-GFX9 chips do not flush denorms. */
868 result = ac_build_canonicalize(&ctx->ac, result,
869 instr->dest.dest.ssa.bit_size);
870 }
871 break;
872 case nir_op_fmin:
873 result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
874 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
875 if (ctx->ac.chip_class < GFX9 &&
876 instr->dest.dest.ssa.bit_size == 32) {
877 /* Only pre-GFX9 chips do not flush denorms. */
878 result = ac_build_canonicalize(&ctx->ac, result,
879 instr->dest.dest.ssa.bit_size);
880 }
881 break;
882 case nir_op_ffma:
883 /* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
884 result = emit_intrin_3f_param(&ctx->ac, ctx->ac.chip_class >= GFX10 ? "llvm.fma" : "llvm.fmuladd",
885 ac_to_float_type(&ctx->ac, def_type), src[0], src[1], src[2]);
886 break;
887 case nir_op_ldexp:
888 src[0] = ac_to_float(&ctx->ac, src[0]);
889 if (ac_get_elem_bits(&ctx->ac, def_type) == 32)
890 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f32", ctx->ac.f32, src, 2, AC_FUNC_ATTR_READNONE);
891 else if (ac_get_elem_bits(&ctx->ac, def_type) == 16)
892 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f16", ctx->ac.f16, src, 2, AC_FUNC_ATTR_READNONE);
893 else
894 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f64", ctx->ac.f64, src, 2, AC_FUNC_ATTR_READNONE);
895 break;
896 case nir_op_bfm:
897 result = emit_bfm(&ctx->ac, src[0], src[1]);
898 break;
899 case nir_op_bitfield_select:
900 result = emit_bitfield_select(&ctx->ac, src[0], src[1], src[2]);
901 break;
902 case nir_op_ubfe:
903 result = ac_build_bfe(&ctx->ac, src[0], src[1], src[2], false);
904 break;
905 case nir_op_ibfe:
906 result = ac_build_bfe(&ctx->ac, src[0], src[1], src[2], true);
907 break;
908 case nir_op_bitfield_reverse:
909 result = ac_build_bitfield_reverse(&ctx->ac, src[0]);
910 break;
911 case nir_op_bit_count:
912 result = ac_build_bit_count(&ctx->ac, src[0]);
913 break;
914 case nir_op_vec2:
915 case nir_op_vec3:
916 case nir_op_vec4:
917 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
918 src[i] = ac_to_integer(&ctx->ac, src[i]);
919 result = ac_build_gather_values(&ctx->ac, src, num_components);
920 break;
921 case nir_op_f2i8:
922 case nir_op_f2i16:
923 case nir_op_f2i32:
924 case nir_op_f2i64:
925 src[0] = ac_to_float(&ctx->ac, src[0]);
926 result = LLVMBuildFPToSI(ctx->ac.builder, src[0], def_type, "");
927 break;
928 case nir_op_f2u8:
929 case nir_op_f2u16:
930 case nir_op_f2u32:
931 case nir_op_f2u64:
932 src[0] = ac_to_float(&ctx->ac, src[0]);
933 result = LLVMBuildFPToUI(ctx->ac.builder, src[0], def_type, "");
934 break;
935 case nir_op_i2f16:
936 case nir_op_i2f32:
937 case nir_op_i2f64:
938 result = LLVMBuildSIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
939 break;
940 case nir_op_u2f16:
941 case nir_op_u2f32:
942 case nir_op_u2f64:
943 result = LLVMBuildUIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
944 break;
945 case nir_op_f2f16_rtz:
946 src[0] = ac_to_float(&ctx->ac, src[0]);
947 if (LLVMTypeOf(src[0]) == ctx->ac.f64)
948 src[0] = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ctx->ac.f32, "");
949 LLVMValueRef param[2] = { src[0], ctx->ac.f32_0 };
950 result = ac_build_cvt_pkrtz_f16(&ctx->ac, param);
951 result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
952 break;
953 case nir_op_f2f16_rtne:
954 case nir_op_f2f16:
955 case nir_op_f2f32:
956 case nir_op_f2f64:
957 src[0] = ac_to_float(&ctx->ac, src[0]);
958 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
959 result = LLVMBuildFPExt(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
960 else
961 result = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
962 break;
963 case nir_op_u2u8:
964 case nir_op_u2u16:
965 case nir_op_u2u32:
966 case nir_op_u2u64:
967 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
968 result = LLVMBuildZExt(ctx->ac.builder, src[0], def_type, "");
969 else
970 result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
971 break;
972 case nir_op_i2i8:
973 case nir_op_i2i16:
974 case nir_op_i2i32:
975 case nir_op_i2i64:
976 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
977 result = LLVMBuildSExt(ctx->ac.builder, src[0], def_type, "");
978 else
979 result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
980 break;
981 case nir_op_b32csel:
982 result = emit_bcsel(&ctx->ac, src[0], src[1], src[2]);
983 break;
984 case nir_op_find_lsb:
985 result = ac_find_lsb(&ctx->ac, ctx->ac.i32, src[0]);
986 break;
987 case nir_op_ufind_msb:
988 result = ac_build_umsb(&ctx->ac, src[0], ctx->ac.i32);
989 break;
990 case nir_op_ifind_msb:
991 result = ac_build_imsb(&ctx->ac, src[0], ctx->ac.i32);
992 break;
993 case nir_op_uadd_carry:
994 result = emit_uint_carry(&ctx->ac, "llvm.uadd.with.overflow.i32", src[0], src[1]);
995 break;
996 case nir_op_usub_borrow:
997 result = emit_uint_carry(&ctx->ac, "llvm.usub.with.overflow.i32", src[0], src[1]);
998 break;
999 case nir_op_b2f16:
1000 case nir_op_b2f32:
1001 case nir_op_b2f64:
1002 result = emit_b2f(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
1003 break;
1004 case nir_op_f2b32:
1005 result = emit_f2b(&ctx->ac, src[0]);
1006 break;
1007 case nir_op_b2i8:
1008 case nir_op_b2i16:
1009 case nir_op_b2i32:
1010 case nir_op_b2i64:
1011 result = emit_b2i(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
1012 break;
1013 case nir_op_i2b32:
1014 result = emit_i2b(&ctx->ac, src[0]);
1015 break;
1016 case nir_op_fquantize2f16:
1017 result = emit_f2f16(&ctx->ac, src[0]);
1018 break;
1019 case nir_op_umul_high:
1020 result = emit_umul_high(&ctx->ac, src[0], src[1]);
1021 break;
1022 case nir_op_imul_high:
1023 result = emit_imul_high(&ctx->ac, src[0], src[1]);
1024 break;
1025 case nir_op_pack_half_2x16:
1026 result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pkrtz_f16);
1027 break;
1028 case nir_op_pack_snorm_2x16:
1029 result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pknorm_i16);
1030 break;
1031 case nir_op_pack_unorm_2x16:
1032 result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pknorm_u16);
1033 break;
1034 case nir_op_unpack_half_2x16:
1035 result = emit_unpack_half_2x16(&ctx->ac, src[0]);
1036 break;
1037 case nir_op_fddx:
1038 case nir_op_fddy:
1039 case nir_op_fddx_fine:
1040 case nir_op_fddy_fine:
1041 case nir_op_fddx_coarse:
1042 case nir_op_fddy_coarse:
1043 result = emit_ddxy(ctx, instr->op, src[0]);
1044 break;
1045
1046 case nir_op_unpack_64_2x32_split_x: {
1047 assert(ac_get_llvm_num_components(src[0]) == 1);
1048 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1049 ctx->ac.v2i32,
1050 "");
1051 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1052 ctx->ac.i32_0, "");
1053 break;
1054 }
1055
1056 case nir_op_unpack_64_2x32_split_y: {
1057 assert(ac_get_llvm_num_components(src[0]) == 1);
1058 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1059 ctx->ac.v2i32,
1060 "");
1061 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1062 ctx->ac.i32_1, "");
1063 break;
1064 }
1065
1066 case nir_op_pack_64_2x32_split: {
1067 LLVMValueRef tmp = ac_build_gather_values(&ctx->ac, src, 2);
1068 result = LLVMBuildBitCast(ctx->ac.builder, tmp, ctx->ac.i64, "");
1069 break;
1070 }
1071
1072 case nir_op_pack_32_2x16_split: {
1073 LLVMValueRef tmp = ac_build_gather_values(&ctx->ac, src, 2);
1074 result = LLVMBuildBitCast(ctx->ac.builder, tmp, ctx->ac.i32, "");
1075 break;
1076 }
1077
1078 case nir_op_unpack_32_2x16_split_x: {
1079 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1080 ctx->ac.v2i16,
1081 "");
1082 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1083 ctx->ac.i32_0, "");
1084 break;
1085 }
1086
1087 case nir_op_unpack_32_2x16_split_y: {
1088 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1089 ctx->ac.v2i16,
1090 "");
1091 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1092 ctx->ac.i32_1, "");
1093 break;
1094 }
1095
1096 case nir_op_cube_face_coord: {
1097 src[0] = ac_to_float(&ctx->ac, src[0]);
1098 LLVMValueRef results[2];
1099 LLVMValueRef in[3];
1100 for (unsigned chan = 0; chan < 3; chan++)
1101 in[chan] = ac_llvm_extract_elem(&ctx->ac, src[0], chan);
1102 results[0] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubesc",
1103 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1104 results[1] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubetc",
1105 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1106 LLVMValueRef ma = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubema",
1107 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1108 results[0] = ac_build_fdiv(&ctx->ac, results[0], ma);
1109 results[1] = ac_build_fdiv(&ctx->ac, results[1], ma);
1110 LLVMValueRef offset = LLVMConstReal(ctx->ac.f32, 0.5);
1111 results[0] = LLVMBuildFAdd(ctx->ac.builder, results[0], offset, "");
1112 results[1] = LLVMBuildFAdd(ctx->ac.builder, results[1], offset, "");
1113 result = ac_build_gather_values(&ctx->ac, results, 2);
1114 break;
1115 }
1116
1117 case nir_op_cube_face_index: {
1118 src[0] = ac_to_float(&ctx->ac, src[0]);
1119 LLVMValueRef in[3];
1120 for (unsigned chan = 0; chan < 3; chan++)
1121 in[chan] = ac_llvm_extract_elem(&ctx->ac, src[0], chan);
1122 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubeid",
1123 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1124 break;
1125 }
1126
1127 case nir_op_fmin3:
1128 result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
1129 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
1130 result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
1131 ac_to_float_type(&ctx->ac, def_type), result, src[2]);
1132 break;
1133 case nir_op_umin3:
1134 result = ac_build_umin(&ctx->ac, src[0], src[1]);
1135 result = ac_build_umin(&ctx->ac, result, src[2]);
1136 break;
1137 case nir_op_imin3:
1138 result = ac_build_imin(&ctx->ac, src[0], src[1]);
1139 result = ac_build_imin(&ctx->ac, result, src[2]);
1140 break;
1141 case nir_op_fmax3:
1142 result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
1143 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
1144 result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
1145 ac_to_float_type(&ctx->ac, def_type), result, src[2]);
1146 break;
1147 case nir_op_umax3:
1148 result = ac_build_umax(&ctx->ac, src[0], src[1]);
1149 result = ac_build_umax(&ctx->ac, result, src[2]);
1150 break;
1151 case nir_op_imax3:
1152 result = ac_build_imax(&ctx->ac, src[0], src[1]);
1153 result = ac_build_imax(&ctx->ac, result, src[2]);
1154 break;
1155 case nir_op_fmed3: {
1156 src[0] = ac_to_float(&ctx->ac, src[0]);
1157 src[1] = ac_to_float(&ctx->ac, src[1]);
1158 src[2] = ac_to_float(&ctx->ac, src[2]);
1159 result = ac_build_fmed3(&ctx->ac, src[0], src[1], src[2],
1160 instr->dest.dest.ssa.bit_size);
1161 break;
1162 }
1163 case nir_op_imed3: {
1164 LLVMValueRef tmp1 = ac_build_imin(&ctx->ac, src[0], src[1]);
1165 LLVMValueRef tmp2 = ac_build_imax(&ctx->ac, src[0], src[1]);
1166 tmp2 = ac_build_imin(&ctx->ac, tmp2, src[2]);
1167 result = ac_build_imax(&ctx->ac, tmp1, tmp2);
1168 break;
1169 }
1170 case nir_op_umed3: {
1171 LLVMValueRef tmp1 = ac_build_umin(&ctx->ac, src[0], src[1]);
1172 LLVMValueRef tmp2 = ac_build_umax(&ctx->ac, src[0], src[1]);
1173 tmp2 = ac_build_umin(&ctx->ac, tmp2, src[2]);
1174 result = ac_build_umax(&ctx->ac, tmp1, tmp2);
1175 break;
1176 }
1177
1178 default:
1179 fprintf(stderr, "Unknown NIR alu instr: ");
1180 nir_print_instr(&instr->instr, stderr);
1181 fprintf(stderr, "\n");
1182 abort();
1183 }
1184
1185 if (result) {
1186 assert(instr->dest.dest.is_ssa);
1187 result = ac_to_integer_or_pointer(&ctx->ac, result);
1188 ctx->ssa_defs[instr->dest.dest.ssa.index] = result;
1189 }
1190
1191 if (instr->exact)
1192 ac_restore_inexact_math(ctx->ac.builder, saved_inexact);
1193 }
1194
1195 static void visit_load_const(struct ac_nir_context *ctx,
1196 const nir_load_const_instr *instr)
1197 {
1198 LLVMValueRef values[4], value = NULL;
1199 LLVMTypeRef element_type =
1200 LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
1201
1202 for (unsigned i = 0; i < instr->def.num_components; ++i) {
1203 switch (instr->def.bit_size) {
1204 case 8:
1205 values[i] = LLVMConstInt(element_type,
1206 instr->value[i].u8, false);
1207 break;
1208 case 16:
1209 values[i] = LLVMConstInt(element_type,
1210 instr->value[i].u16, false);
1211 break;
1212 case 32:
1213 values[i] = LLVMConstInt(element_type,
1214 instr->value[i].u32, false);
1215 break;
1216 case 64:
1217 values[i] = LLVMConstInt(element_type,
1218 instr->value[i].u64, false);
1219 break;
1220 default:
1221 fprintf(stderr,
1222 "unsupported nir load_const bit_size: %d\n",
1223 instr->def.bit_size);
1224 abort();
1225 }
1226 }
1227 if (instr->def.num_components > 1) {
1228 value = LLVMConstVector(values, instr->def.num_components);
1229 } else
1230 value = values[0];
1231
1232 ctx->ssa_defs[instr->def.index] = value;
1233 }
1234
1235 static LLVMValueRef
1236 get_buffer_size(struct ac_nir_context *ctx, LLVMValueRef descriptor, bool in_elements)
1237 {
1238 LLVMValueRef size =
1239 LLVMBuildExtractElement(ctx->ac.builder, descriptor,
1240 LLVMConstInt(ctx->ac.i32, 2, false), "");
1241
1242 /* GFX8 only */
1243 if (ctx->ac.chip_class == GFX8 && in_elements) {
1244 /* On GFX8, the descriptor contains the size in bytes,
1245 * but TXQ must return the size in elements.
1246 * The stride is always non-zero for resources using TXQ.
1247 */
1248 LLVMValueRef stride =
1249 LLVMBuildExtractElement(ctx->ac.builder, descriptor,
1250 ctx->ac.i32_1, "");
1251 stride = LLVMBuildLShr(ctx->ac.builder, stride,
1252 LLVMConstInt(ctx->ac.i32, 16, false), "");
1253 stride = LLVMBuildAnd(ctx->ac.builder, stride,
1254 LLVMConstInt(ctx->ac.i32, 0x3fff, false), "");
1255
1256 size = LLVMBuildUDiv(ctx->ac.builder, size, stride, "");
1257 }
1258 return size;
1259 }
1260
1261 /* Gather4 should follow the same rules as bilinear filtering, but the hardware
1262 * incorrectly forces nearest filtering if the texture format is integer.
1263 * The only effect it has on Gather4, which always returns 4 texels for
1264 * bilinear filtering, is that the final coordinates are off by 0.5 of
1265 * the texel size.
1266 *
1267 * The workaround is to subtract 0.5 from the unnormalized coordinates,
1268 * or (0.5 / size) from the normalized coordinates.
1269 *
1270 * However, cube textures with 8_8_8_8 data formats require a different
1271 * workaround of overriding the num format to USCALED/SSCALED. This would lose
1272 * precision in 32-bit data formats, so it needs to be applied dynamically at
1273 * runtime. In this case, return an i1 value that indicates whether the
1274 * descriptor was overridden (and hence a fixup of the sampler result is needed).
1275 */
1276 static LLVMValueRef lower_gather4_integer(struct ac_llvm_context *ctx,
1277 nir_variable *var,
1278 struct ac_image_args *args,
1279 const nir_tex_instr *instr)
1280 {
1281 const struct glsl_type *type = glsl_without_array(var->type);
1282 enum glsl_base_type stype = glsl_get_sampler_result_type(type);
1283 LLVMValueRef wa_8888 = NULL;
1284 LLVMValueRef half_texel[2];
1285 LLVMValueRef result;
1286
1287 assert(stype == GLSL_TYPE_INT || stype == GLSL_TYPE_UINT);
1288
1289 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
1290 LLVMValueRef formats;
1291 LLVMValueRef data_format;
1292 LLVMValueRef wa_formats;
1293
1294 formats = LLVMBuildExtractElement(ctx->builder, args->resource, ctx->i32_1, "");
1295
1296 data_format = LLVMBuildLShr(ctx->builder, formats,
1297 LLVMConstInt(ctx->i32, 20, false), "");
1298 data_format = LLVMBuildAnd(ctx->builder, data_format,
1299 LLVMConstInt(ctx->i32, (1u << 6) - 1, false), "");
1300 wa_8888 = LLVMBuildICmp(
1301 ctx->builder, LLVMIntEQ, data_format,
1302 LLVMConstInt(ctx->i32, V_008F14_IMG_DATA_FORMAT_8_8_8_8, false),
1303 "");
1304
1305 uint32_t wa_num_format =
1306 stype == GLSL_TYPE_UINT ?
1307 S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_USCALED) :
1308 S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_SSCALED);
1309 wa_formats = LLVMBuildAnd(ctx->builder, formats,
1310 LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT, false),
1311 "");
1312 wa_formats = LLVMBuildOr(ctx->builder, wa_formats,
1313 LLVMConstInt(ctx->i32, wa_num_format, false), "");
1314
1315 formats = LLVMBuildSelect(ctx->builder, wa_8888, wa_formats, formats, "");
1316 args->resource = LLVMBuildInsertElement(
1317 ctx->builder, args->resource, formats, ctx->i32_1, "");
1318 }
1319
1320 if (instr->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
1321 assert(!wa_8888);
1322 half_texel[0] = half_texel[1] = LLVMConstReal(ctx->f32, -0.5);
1323 } else {
1324 struct ac_image_args resinfo = {};
1325 LLVMBasicBlockRef bbs[2];
1326
1327 LLVMValueRef unnorm = NULL;
1328 LLVMValueRef default_offset = ctx->f32_0;
1329 if (instr->sampler_dim == GLSL_SAMPLER_DIM_2D &&
1330 !instr->is_array) {
1331 /* In vulkan, whether the sampler uses unnormalized
1332 * coordinates or not is a dynamic property of the
1333 * sampler. Hence, to figure out whether or not we
1334 * need to divide by the texture size, we need to test
1335 * the sampler at runtime. This tests the bit set by
1336 * radv_init_sampler().
1337 */
1338 LLVMValueRef sampler0 =
1339 LLVMBuildExtractElement(ctx->builder, args->sampler, ctx->i32_0, "");
1340 sampler0 = LLVMBuildLShr(ctx->builder, sampler0,
1341 LLVMConstInt(ctx->i32, 15, false), "");
1342 sampler0 = LLVMBuildAnd(ctx->builder, sampler0, ctx->i32_1, "");
1343 unnorm = LLVMBuildICmp(ctx->builder, LLVMIntEQ, sampler0, ctx->i32_1, "");
1344 default_offset = LLVMConstReal(ctx->f32, -0.5);
1345 }
1346
1347 bbs[0] = LLVMGetInsertBlock(ctx->builder);
1348 if (wa_8888 || unnorm) {
1349 assert(!(wa_8888 && unnorm));
1350 LLVMValueRef not_needed = wa_8888 ? wa_8888 : unnorm;
1351 /* Skip the texture size query entirely if we don't need it. */
1352 ac_build_ifcc(ctx, LLVMBuildNot(ctx->builder, not_needed, ""), 2000);
1353 bbs[1] = LLVMGetInsertBlock(ctx->builder);
1354 }
1355
1356 /* Query the texture size. */
1357 resinfo.dim = ac_get_sampler_dim(ctx->chip_class, instr->sampler_dim, instr->is_array);
1358 resinfo.opcode = ac_image_get_resinfo;
1359 resinfo.dmask = 0xf;
1360 resinfo.lod = ctx->i32_0;
1361 resinfo.resource = args->resource;
1362 resinfo.attributes = AC_FUNC_ATTR_READNONE;
1363 LLVMValueRef size = ac_build_image_opcode(ctx, &resinfo);
1364
1365 /* Compute -0.5 / size. */
1366 for (unsigned c = 0; c < 2; c++) {
1367 half_texel[c] =
1368 LLVMBuildExtractElement(ctx->builder, size,
1369 LLVMConstInt(ctx->i32, c, 0), "");
1370 half_texel[c] = LLVMBuildUIToFP(ctx->builder, half_texel[c], ctx->f32, "");
1371 half_texel[c] = ac_build_fdiv(ctx, ctx->f32_1, half_texel[c]);
1372 half_texel[c] = LLVMBuildFMul(ctx->builder, half_texel[c],
1373 LLVMConstReal(ctx->f32, -0.5), "");
1374 }
1375
1376 if (wa_8888 || unnorm) {
1377 ac_build_endif(ctx, 2000);
1378
1379 for (unsigned c = 0; c < 2; c++) {
1380 LLVMValueRef values[2] = { default_offset, half_texel[c] };
1381 half_texel[c] = ac_build_phi(ctx, ctx->f32, 2,
1382 values, bbs);
1383 }
1384 }
1385 }
1386
1387 for (unsigned c = 0; c < 2; c++) {
1388 LLVMValueRef tmp;
1389 tmp = LLVMBuildBitCast(ctx->builder, args->coords[c], ctx->f32, "");
1390 args->coords[c] = LLVMBuildFAdd(ctx->builder, tmp, half_texel[c], "");
1391 }
1392
1393 args->attributes = AC_FUNC_ATTR_READNONE;
1394 result = ac_build_image_opcode(ctx, args);
1395
1396 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
1397 LLVMValueRef tmp, tmp2;
1398
1399 /* if the cube workaround is in place, f2i the result. */
1400 for (unsigned c = 0; c < 4; c++) {
1401 tmp = LLVMBuildExtractElement(ctx->builder, result, LLVMConstInt(ctx->i32, c, false), "");
1402 if (stype == GLSL_TYPE_UINT)
1403 tmp2 = LLVMBuildFPToUI(ctx->builder, tmp, ctx->i32, "");
1404 else
1405 tmp2 = LLVMBuildFPToSI(ctx->builder, tmp, ctx->i32, "");
1406 tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->i32, "");
1407 tmp2 = LLVMBuildBitCast(ctx->builder, tmp2, ctx->i32, "");
1408 tmp = LLVMBuildSelect(ctx->builder, wa_8888, tmp2, tmp, "");
1409 tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->f32, "");
1410 result = LLVMBuildInsertElement(ctx->builder, result, tmp, LLVMConstInt(ctx->i32, c, false), "");
1411 }
1412 }
1413 return result;
1414 }
1415
1416 static nir_deref_instr *get_tex_texture_deref(const nir_tex_instr *instr)
1417 {
1418 nir_deref_instr *texture_deref_instr = NULL;
1419
1420 for (unsigned i = 0; i < instr->num_srcs; i++) {
1421 switch (instr->src[i].src_type) {
1422 case nir_tex_src_texture_deref:
1423 texture_deref_instr = nir_src_as_deref(instr->src[i].src);
1424 break;
1425 default:
1426 break;
1427 }
1428 }
1429 return texture_deref_instr;
1430 }
1431
1432 static LLVMValueRef build_tex_intrinsic(struct ac_nir_context *ctx,
1433 const nir_tex_instr *instr,
1434 struct ac_image_args *args)
1435 {
1436 if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
1437 unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
1438
1439 assert(instr->dest.is_ssa);
1440 return ac_build_buffer_load_format(&ctx->ac,
1441 args->resource,
1442 args->coords[0],
1443 ctx->ac.i32_0,
1444 util_last_bit(mask),
1445 0, true,
1446 instr->dest.ssa.bit_size == 16);
1447 }
1448
1449 args->opcode = ac_image_sample;
1450
1451 switch (instr->op) {
1452 case nir_texop_txf:
1453 case nir_texop_txf_ms:
1454 case nir_texop_samples_identical:
1455 args->opcode = args->level_zero ||
1456 instr->sampler_dim == GLSL_SAMPLER_DIM_MS ?
1457 ac_image_load : ac_image_load_mip;
1458 args->level_zero = false;
1459 break;
1460 case nir_texop_txs:
1461 case nir_texop_query_levels:
1462 args->opcode = ac_image_get_resinfo;
1463 if (!args->lod)
1464 args->lod = ctx->ac.i32_0;
1465 args->level_zero = false;
1466 break;
1467 case nir_texop_tex:
1468 if (ctx->stage != MESA_SHADER_FRAGMENT) {
1469 assert(!args->lod);
1470 args->level_zero = true;
1471 }
1472 break;
1473 case nir_texop_tg4:
1474 args->opcode = ac_image_gather4;
1475 if (!args->lod && !args->bias)
1476 args->level_zero = true;
1477 break;
1478 case nir_texop_lod:
1479 args->opcode = ac_image_get_lod;
1480 break;
1481 case nir_texop_fragment_fetch:
1482 case nir_texop_fragment_mask_fetch:
1483 args->opcode = ac_image_load;
1484 args->level_zero = false;
1485 break;
1486 default:
1487 break;
1488 }
1489
1490 if (instr->op == nir_texop_tg4 && ctx->ac.chip_class <= GFX8) {
1491 nir_deref_instr *texture_deref_instr = get_tex_texture_deref(instr);
1492 nir_variable *var = nir_deref_instr_get_variable(texture_deref_instr);
1493 const struct glsl_type *type = glsl_without_array(var->type);
1494 enum glsl_base_type stype = glsl_get_sampler_result_type(type);
1495 if (stype == GLSL_TYPE_UINT || stype == GLSL_TYPE_INT) {
1496 return lower_gather4_integer(&ctx->ac, var, args, instr);
1497 }
1498 }
1499
1500 /* Fixup for GFX9 which allocates 1D textures as 2D. */
1501 if (instr->op == nir_texop_lod && ctx->ac.chip_class == GFX9) {
1502 if ((args->dim == ac_image_2darray ||
1503 args->dim == ac_image_2d) && !args->coords[1]) {
1504 args->coords[1] = ctx->ac.i32_0;
1505 }
1506 }
1507
1508 args->attributes = AC_FUNC_ATTR_READNONE;
1509 bool cs_derivs = ctx->stage == MESA_SHADER_COMPUTE &&
1510 ctx->info->cs.derivative_group != DERIVATIVE_GROUP_NONE;
1511 if (ctx->stage == MESA_SHADER_FRAGMENT || cs_derivs) {
1512 /* Prevent texture instructions with implicit derivatives from being
1513 * sinked into branches. */
1514 switch (instr->op) {
1515 case nir_texop_tex:
1516 case nir_texop_txb:
1517 case nir_texop_lod:
1518 args->attributes |= AC_FUNC_ATTR_CONVERGENT;
1519 break;
1520 default:
1521 break;
1522 }
1523 }
1524
1525 return ac_build_image_opcode(&ctx->ac, args);
1526 }
1527
1528 static LLVMValueRef visit_vulkan_resource_reindex(struct ac_nir_context *ctx,
1529 nir_intrinsic_instr *instr)
1530 {
1531 LLVMValueRef ptr = get_src(ctx, instr->src[0]);
1532 LLVMValueRef index = get_src(ctx, instr->src[1]);
1533
1534 LLVMValueRef result = LLVMBuildGEP(ctx->ac.builder, ptr, &index, 1, "");
1535 LLVMSetMetadata(result, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
1536 return result;
1537 }
1538
1539 static LLVMValueRef visit_load_push_constant(struct ac_nir_context *ctx,
1540 nir_intrinsic_instr *instr)
1541 {
1542 LLVMValueRef ptr, addr;
1543 LLVMValueRef src0 = get_src(ctx, instr->src[0]);
1544 unsigned index = nir_intrinsic_base(instr);
1545
1546 addr = LLVMConstInt(ctx->ac.i32, index, 0);
1547 addr = LLVMBuildAdd(ctx->ac.builder, addr, src0, "");
1548
1549 /* Load constant values from user SGPRS when possible, otherwise
1550 * fallback to the default path that loads directly from memory.
1551 */
1552 if (LLVMIsConstant(src0) &&
1553 instr->dest.ssa.bit_size == 32) {
1554 unsigned count = instr->dest.ssa.num_components;
1555 unsigned offset = index;
1556
1557 offset += LLVMConstIntGetZExtValue(src0);
1558 offset /= 4;
1559
1560 offset -= ctx->args->base_inline_push_consts;
1561
1562 unsigned num_inline_push_consts = ctx->args->num_inline_push_consts;
1563 if (offset + count <= num_inline_push_consts) {
1564 LLVMValueRef push_constants[num_inline_push_consts];
1565 for (unsigned i = 0; i < num_inline_push_consts; i++)
1566 push_constants[i] = ac_get_arg(&ctx->ac,
1567 ctx->args->inline_push_consts[i]);
1568 return ac_build_gather_values(&ctx->ac,
1569 push_constants + offset,
1570 count);
1571 }
1572 }
1573
1574 ptr = LLVMBuildGEP(ctx->ac.builder,
1575 ac_get_arg(&ctx->ac, ctx->args->push_constants), &addr, 1, "");
1576
1577 if (instr->dest.ssa.bit_size == 8) {
1578 unsigned load_dwords = instr->dest.ssa.num_components > 1 ? 2 : 1;
1579 LLVMTypeRef vec_type = LLVMVectorType(ctx->ac.i8, 4 * load_dwords);
1580 ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
1581 LLVMValueRef res = LLVMBuildLoad(ctx->ac.builder, ptr, "");
1582
1583 LLVMValueRef params[3];
1584 if (load_dwords > 1) {
1585 LLVMValueRef res_vec = LLVMBuildBitCast(ctx->ac.builder, res, ctx->ac.v2i32, "");
1586 params[0] = LLVMBuildExtractElement(ctx->ac.builder, res_vec, LLVMConstInt(ctx->ac.i32, 1, false), "");
1587 params[1] = LLVMBuildExtractElement(ctx->ac.builder, res_vec, LLVMConstInt(ctx->ac.i32, 0, false), "");
1588 } else {
1589 res = LLVMBuildBitCast(ctx->ac.builder, res, ctx->ac.i32, "");
1590 params[0] = ctx->ac.i32_0;
1591 params[1] = res;
1592 }
1593 params[2] = addr;
1594 res = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.alignbyte", ctx->ac.i32, params, 3, 0);
1595
1596 res = LLVMBuildTrunc(ctx->ac.builder, res, LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.num_components * 8), "");
1597 if (instr->dest.ssa.num_components > 1)
1598 res = LLVMBuildBitCast(ctx->ac.builder, res, LLVMVectorType(ctx->ac.i8, instr->dest.ssa.num_components), "");
1599 return res;
1600 } else if (instr->dest.ssa.bit_size == 16) {
1601 unsigned load_dwords = instr->dest.ssa.num_components / 2 + 1;
1602 LLVMTypeRef vec_type = LLVMVectorType(ctx->ac.i16, 2 * load_dwords);
1603 ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
1604 LLVMValueRef res = LLVMBuildLoad(ctx->ac.builder, ptr, "");
1605 res = LLVMBuildBitCast(ctx->ac.builder, res, vec_type, "");
1606 LLVMValueRef cond = LLVMBuildLShr(ctx->ac.builder, addr, ctx->ac.i32_1, "");
1607 cond = LLVMBuildTrunc(ctx->ac.builder, cond, ctx->ac.i1, "");
1608 LLVMValueRef mask[] = { LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
1609 LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
1610 LLVMConstInt(ctx->ac.i32, 4, false)};
1611 LLVMValueRef swizzle_aligned = LLVMConstVector(&mask[0], instr->dest.ssa.num_components);
1612 LLVMValueRef swizzle_unaligned = LLVMConstVector(&mask[1], instr->dest.ssa.num_components);
1613 LLVMValueRef shuffle_aligned = LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_aligned, "");
1614 LLVMValueRef shuffle_unaligned = LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_unaligned, "");
1615 res = LLVMBuildSelect(ctx->ac.builder, cond, shuffle_unaligned, shuffle_aligned, "");
1616 return LLVMBuildBitCast(ctx->ac.builder, res, get_def_type(ctx, &instr->dest.ssa), "");
1617 }
1618
1619 ptr = ac_cast_ptr(&ctx->ac, ptr, get_def_type(ctx, &instr->dest.ssa));
1620
1621 return LLVMBuildLoad(ctx->ac.builder, ptr, "");
1622 }
1623
1624 static LLVMValueRef visit_get_buffer_size(struct ac_nir_context *ctx,
1625 const nir_intrinsic_instr *instr)
1626 {
1627 LLVMValueRef index = get_src(ctx, instr->src[0]);
1628
1629 return get_buffer_size(ctx, ctx->abi->load_ssbo(ctx->abi, index, false), false);
1630 }
1631
1632 static uint32_t widen_mask(uint32_t mask, unsigned multiplier)
1633 {
1634 uint32_t new_mask = 0;
1635 for(unsigned i = 0; i < 32 && (1u << i) <= mask; ++i)
1636 if (mask & (1u << i))
1637 new_mask |= ((1u << multiplier) - 1u) << (i * multiplier);
1638 return new_mask;
1639 }
1640
1641 static LLVMValueRef extract_vector_range(struct ac_llvm_context *ctx, LLVMValueRef src,
1642 unsigned start, unsigned count)
1643 {
1644 LLVMValueRef mask[] = {
1645 ctx->i32_0, ctx->i32_1,
1646 LLVMConstInt(ctx->i32, 2, false), LLVMConstInt(ctx->i32, 3, false) };
1647
1648 unsigned src_elements = ac_get_llvm_num_components(src);
1649
1650 if (count == src_elements) {
1651 assert(start == 0);
1652 return src;
1653 } else if (count == 1) {
1654 assert(start < src_elements);
1655 return LLVMBuildExtractElement(ctx->builder, src, mask[start], "");
1656 } else {
1657 assert(start + count <= src_elements);
1658 assert(count <= 4);
1659 LLVMValueRef swizzle = LLVMConstVector(&mask[start], count);
1660 return LLVMBuildShuffleVector(ctx->builder, src, src, swizzle, "");
1661 }
1662 }
1663
1664 static unsigned get_cache_policy(struct ac_nir_context *ctx,
1665 enum gl_access_qualifier access,
1666 bool may_store_unaligned,
1667 bool writeonly_memory)
1668 {
1669 unsigned cache_policy = 0;
1670
1671 /* GFX6 has a TC L1 bug causing corruption of 8bit/16bit stores. All
1672 * store opcodes not aligned to a dword are affected. The only way to
1673 * get unaligned stores is through shader images.
1674 */
1675 if (((may_store_unaligned && ctx->ac.chip_class == GFX6) ||
1676 /* If this is write-only, don't keep data in L1 to prevent
1677 * evicting L1 cache lines that may be needed by other
1678 * instructions.
1679 */
1680 writeonly_memory ||
1681 access & (ACCESS_COHERENT | ACCESS_VOLATILE))) {
1682 cache_policy |= ac_glc;
1683 }
1684
1685 if (access & ACCESS_STREAM_CACHE_POLICY)
1686 cache_policy |= ac_slc | ac_glc;
1687
1688 return cache_policy;
1689 }
1690
1691 static LLVMValueRef enter_waterfall_ssbo(struct ac_nir_context *ctx,
1692 struct waterfall_context *wctx,
1693 const nir_intrinsic_instr *instr,
1694 nir_src src)
1695 {
1696 return enter_waterfall(ctx, wctx, get_src(ctx, src),
1697 nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
1698 }
1699
1700 static void visit_store_ssbo(struct ac_nir_context *ctx,
1701 nir_intrinsic_instr *instr)
1702 {
1703 if (ctx->ac.postponed_kill) {
1704 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
1705 ctx->ac.postponed_kill, "");
1706 ac_build_ifcc(&ctx->ac, cond, 7000);
1707 }
1708
1709 LLVMValueRef src_data = get_src(ctx, instr->src[0]);
1710 int elem_size_bytes = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src_data)) / 8;
1711 unsigned writemask = nir_intrinsic_write_mask(instr);
1712 enum gl_access_qualifier access = nir_intrinsic_access(instr);
1713 bool writeonly_memory = access & ACCESS_NON_READABLE;
1714 unsigned cache_policy = get_cache_policy(ctx, access, false, writeonly_memory);
1715
1716 struct waterfall_context wctx;
1717 LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[1]);
1718
1719 LLVMValueRef rsrc = ctx->abi->load_ssbo(ctx->abi, rsrc_base, true);
1720 LLVMValueRef base_data = src_data;
1721 base_data = ac_trim_vector(&ctx->ac, base_data, instr->num_components);
1722 LLVMValueRef base_offset = get_src(ctx, instr->src[2]);
1723
1724 while (writemask) {
1725 int start, count;
1726 LLVMValueRef data, offset;
1727 LLVMTypeRef data_type;
1728
1729 u_bit_scan_consecutive_range(&writemask, &start, &count);
1730
1731 /* Due to an LLVM limitation with LLVM < 9, split 3-element
1732 * writes into a 2-element and a 1-element write. */
1733 if (count == 3 &&
1734 (elem_size_bytes != 4 || !ac_has_vec3_support(ctx->ac.chip_class, false))) {
1735 writemask |= 1 << (start + 2);
1736 count = 2;
1737 }
1738 int num_bytes = count * elem_size_bytes; /* count in bytes */
1739
1740 /* we can only store 4 DWords at the same time.
1741 * can only happen for 64 Bit vectors. */
1742 if (num_bytes > 16) {
1743 writemask |= ((1u << (count - 2)) - 1u) << (start + 2);
1744 count = 2;
1745 num_bytes = 16;
1746 }
1747
1748 /* check alignment of 16 Bit stores */
1749 if (elem_size_bytes == 2 && num_bytes > 2 && (start % 2) == 1) {
1750 writemask |= ((1u << (count - 1)) - 1u) << (start + 1);
1751 count = 1;
1752 num_bytes = 2;
1753 }
1754
1755 /* Due to alignment issues, split stores of 8-bit/16-bit
1756 * vectors.
1757 */
1758 if (ctx->ac.chip_class == GFX6 && count > 1 && elem_size_bytes < 4) {
1759 writemask |= ((1u << (count - 1)) - 1u) << (start + 1);
1760 count = 1;
1761 num_bytes = elem_size_bytes;
1762 }
1763
1764 data = extract_vector_range(&ctx->ac, base_data, start, count);
1765
1766 offset = LLVMBuildAdd(ctx->ac.builder, base_offset,
1767 LLVMConstInt(ctx->ac.i32, start * elem_size_bytes, false), "");
1768
1769 if (num_bytes == 1) {
1770 ac_build_tbuffer_store_byte(&ctx->ac, rsrc, data,
1771 offset, ctx->ac.i32_0,
1772 cache_policy);
1773 } else if (num_bytes == 2) {
1774 ac_build_tbuffer_store_short(&ctx->ac, rsrc, data,
1775 offset, ctx->ac.i32_0,
1776 cache_policy);
1777 } else {
1778 int num_channels = num_bytes / 4;
1779
1780 switch (num_bytes) {
1781 case 16: /* v4f32 */
1782 data_type = ctx->ac.v4f32;
1783 break;
1784 case 12: /* v3f32 */
1785 data_type = ctx->ac.v3f32;
1786 break;
1787 case 8: /* v2f32 */
1788 data_type = ctx->ac.v2f32;
1789 break;
1790 case 4: /* f32 */
1791 data_type = ctx->ac.f32;
1792 break;
1793 default:
1794 unreachable("Malformed vector store.");
1795 }
1796 data = LLVMBuildBitCast(ctx->ac.builder, data, data_type, "");
1797
1798 ac_build_buffer_store_dword(&ctx->ac, rsrc, data,
1799 num_channels, offset,
1800 ctx->ac.i32_0, 0,
1801 cache_policy);
1802 }
1803 }
1804
1805 exit_waterfall(ctx, &wctx, NULL);
1806
1807 if (ctx->ac.postponed_kill)
1808 ac_build_endif(&ctx->ac, 7000);
1809 }
1810
1811 static LLVMValueRef emit_ssbo_comp_swap_64(struct ac_nir_context *ctx,
1812 LLVMValueRef descriptor,
1813 LLVMValueRef offset,
1814 LLVMValueRef compare,
1815 LLVMValueRef exchange)
1816 {
1817 LLVMBasicBlockRef start_block = NULL, then_block = NULL;
1818 if (ctx->abi->robust_buffer_access) {
1819 LLVMValueRef size = ac_llvm_extract_elem(&ctx->ac, descriptor, 2);
1820
1821 LLVMValueRef cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, offset, size, "");
1822 start_block = LLVMGetInsertBlock(ctx->ac.builder);
1823
1824 ac_build_ifcc(&ctx->ac, cond, -1);
1825
1826 then_block = LLVMGetInsertBlock(ctx->ac.builder);
1827 }
1828
1829 LLVMValueRef ptr_parts[2] = {
1830 ac_llvm_extract_elem(&ctx->ac, descriptor, 0),
1831 LLVMBuildAnd(ctx->ac.builder,
1832 ac_llvm_extract_elem(&ctx->ac, descriptor, 1),
1833 LLVMConstInt(ctx->ac.i32, 65535, 0), "")
1834 };
1835
1836 ptr_parts[1] = LLVMBuildTrunc(ctx->ac.builder, ptr_parts[1], ctx->ac.i16, "");
1837 ptr_parts[1] = LLVMBuildSExt(ctx->ac.builder, ptr_parts[1], ctx->ac.i32, "");
1838
1839 offset = LLVMBuildZExt(ctx->ac.builder, offset, ctx->ac.i64, "");
1840
1841 LLVMValueRef ptr = ac_build_gather_values(&ctx->ac, ptr_parts, 2);
1842 ptr = LLVMBuildBitCast(ctx->ac.builder, ptr, ctx->ac.i64, "");
1843 ptr = LLVMBuildAdd(ctx->ac.builder, ptr, offset, "");
1844 ptr = LLVMBuildIntToPtr(ctx->ac.builder, ptr, LLVMPointerType(ctx->ac.i64, AC_ADDR_SPACE_GLOBAL), "");
1845
1846 LLVMValueRef result = ac_build_atomic_cmp_xchg(&ctx->ac, ptr, compare, exchange, "singlethread-one-as");
1847 result = LLVMBuildExtractValue(ctx->ac.builder, result, 0, "");
1848
1849 if (ctx->abi->robust_buffer_access) {
1850 ac_build_endif(&ctx->ac, -1);
1851
1852 LLVMBasicBlockRef incoming_blocks[2] = {
1853 start_block,
1854 then_block,
1855 };
1856
1857 LLVMValueRef incoming_values[2] = {
1858 LLVMConstInt(ctx->ac.i64, 0, 0),
1859 result,
1860 };
1861 LLVMValueRef ret = LLVMBuildPhi(ctx->ac.builder, ctx->ac.i64, "");
1862 LLVMAddIncoming(ret, incoming_values, incoming_blocks, 2);
1863 return ret;
1864 } else {
1865 return result;
1866 }
1867 }
1868
1869 static LLVMValueRef visit_atomic_ssbo(struct ac_nir_context *ctx,
1870 nir_intrinsic_instr *instr)
1871 {
1872 if (ctx->ac.postponed_kill) {
1873 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
1874 ctx->ac.postponed_kill, "");
1875 ac_build_ifcc(&ctx->ac, cond, 7001);
1876 }
1877
1878 LLVMTypeRef return_type = LLVMTypeOf(get_src(ctx, instr->src[2]));
1879 const char *op;
1880 char name[64], type[8];
1881 LLVMValueRef params[6], descriptor;
1882 LLVMValueRef result;
1883 int arg_count = 0;
1884
1885 struct waterfall_context wctx;
1886 LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[0]);
1887
1888 switch (instr->intrinsic) {
1889 case nir_intrinsic_ssbo_atomic_add:
1890 op = "add";
1891 break;
1892 case nir_intrinsic_ssbo_atomic_imin:
1893 op = "smin";
1894 break;
1895 case nir_intrinsic_ssbo_atomic_umin:
1896 op = "umin";
1897 break;
1898 case nir_intrinsic_ssbo_atomic_imax:
1899 op = "smax";
1900 break;
1901 case nir_intrinsic_ssbo_atomic_umax:
1902 op = "umax";
1903 break;
1904 case nir_intrinsic_ssbo_atomic_and:
1905 op = "and";
1906 break;
1907 case nir_intrinsic_ssbo_atomic_or:
1908 op = "or";
1909 break;
1910 case nir_intrinsic_ssbo_atomic_xor:
1911 op = "xor";
1912 break;
1913 case nir_intrinsic_ssbo_atomic_exchange:
1914 op = "swap";
1915 break;
1916 case nir_intrinsic_ssbo_atomic_comp_swap:
1917 op = "cmpswap";
1918 break;
1919 default:
1920 abort();
1921 }
1922
1923 descriptor = ctx->abi->load_ssbo(ctx->abi,
1924 rsrc_base,
1925 true);
1926
1927 if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap &&
1928 return_type == ctx->ac.i64) {
1929 result = emit_ssbo_comp_swap_64(ctx, descriptor,
1930 get_src(ctx, instr->src[1]),
1931 get_src(ctx, instr->src[2]),
1932 get_src(ctx, instr->src[3]));
1933 } else {
1934 if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap) {
1935 params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[3]), 0);
1936 }
1937 params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
1938 params[arg_count++] = descriptor;
1939
1940 if (LLVM_VERSION_MAJOR >= 9) {
1941 /* XXX: The new raw/struct atomic intrinsics are buggy with
1942 * LLVM 8, see r358579.
1943 */
1944 params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
1945 params[arg_count++] = ctx->ac.i32_0; /* soffset */
1946 params[arg_count++] = ctx->ac.i32_0; /* slc */
1947
1948 ac_build_type_name_for_intr(return_type, type, sizeof(type));
1949 snprintf(name, sizeof(name),
1950 "llvm.amdgcn.raw.buffer.atomic.%s.%s", op, type);
1951 } else {
1952 params[arg_count++] = ctx->ac.i32_0; /* vindex */
1953 params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
1954 params[arg_count++] = ctx->ac.i1false; /* slc */
1955
1956 assert(return_type == ctx->ac.i32);
1957 snprintf(name, sizeof(name),
1958 "llvm.amdgcn.buffer.atomic.%s", op);
1959 }
1960
1961 result = ac_build_intrinsic(&ctx->ac, name, return_type, params,
1962 arg_count, 0);
1963 }
1964
1965 result = exit_waterfall(ctx, &wctx, result);
1966 if (ctx->ac.postponed_kill)
1967 ac_build_endif(&ctx->ac, 7001);
1968 return result;
1969 }
1970
1971 static LLVMValueRef visit_load_buffer(struct ac_nir_context *ctx,
1972 nir_intrinsic_instr *instr)
1973 {
1974 struct waterfall_context wctx;
1975 LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[0]);
1976
1977 int elem_size_bytes = instr->dest.ssa.bit_size / 8;
1978 int num_components = instr->num_components;
1979 enum gl_access_qualifier access = nir_intrinsic_access(instr);
1980 unsigned cache_policy = get_cache_policy(ctx, access, false, false);
1981
1982 LLVMValueRef offset = get_src(ctx, instr->src[1]);
1983 LLVMValueRef rsrc = ctx->abi->load_ssbo(ctx->abi, rsrc_base, false);
1984 LLVMValueRef vindex = ctx->ac.i32_0;
1985
1986 LLVMTypeRef def_type = get_def_type(ctx, &instr->dest.ssa);
1987 LLVMTypeRef def_elem_type = num_components > 1 ? LLVMGetElementType(def_type) : def_type;
1988
1989 LLVMValueRef results[4];
1990 for (int i = 0; i < num_components;) {
1991 int num_elems = num_components - i;
1992 if (elem_size_bytes < 4 && nir_intrinsic_align(instr) % 4 != 0)
1993 num_elems = 1;
1994 if (num_elems * elem_size_bytes > 16)
1995 num_elems = 16 / elem_size_bytes;
1996 int load_bytes = num_elems * elem_size_bytes;
1997
1998 LLVMValueRef immoffset = LLVMConstInt(ctx->ac.i32, i * elem_size_bytes, false);
1999
2000 LLVMValueRef ret;
2001
2002 if (load_bytes == 1) {
2003 ret = ac_build_tbuffer_load_byte(&ctx->ac,
2004 rsrc,
2005 offset,
2006 ctx->ac.i32_0,
2007 immoffset,
2008 cache_policy);
2009 } else if (load_bytes == 2) {
2010 ret = ac_build_tbuffer_load_short(&ctx->ac,
2011 rsrc,
2012 offset,
2013 ctx->ac.i32_0,
2014 immoffset,
2015 cache_policy);
2016 } else {
2017 int num_channels = util_next_power_of_two(load_bytes) / 4;
2018 bool can_speculate = access & ACCESS_CAN_REORDER;
2019
2020 ret = ac_build_buffer_load(&ctx->ac, rsrc, num_channels,
2021 vindex, offset, immoffset, 0,
2022 cache_policy, can_speculate, false);
2023 }
2024
2025 LLVMTypeRef byte_vec = LLVMVectorType(ctx->ac.i8, ac_get_type_size(LLVMTypeOf(ret)));
2026 ret = LLVMBuildBitCast(ctx->ac.builder, ret, byte_vec, "");
2027 ret = ac_trim_vector(&ctx->ac, ret, load_bytes);
2028
2029 LLVMTypeRef ret_type = LLVMVectorType(def_elem_type, num_elems);
2030 ret = LLVMBuildBitCast(ctx->ac.builder, ret, ret_type, "");
2031
2032 for (unsigned j = 0; j < num_elems; j++) {
2033 results[i + j] = LLVMBuildExtractElement(ctx->ac.builder, ret, LLVMConstInt(ctx->ac.i32, j, false), "");
2034 }
2035 i += num_elems;
2036 }
2037
2038 LLVMValueRef ret = ac_build_gather_values(&ctx->ac, results, num_components);
2039 return exit_waterfall(ctx, &wctx, ret);
2040 }
2041
2042 static LLVMValueRef enter_waterfall_ubo(struct ac_nir_context *ctx,
2043 struct waterfall_context *wctx,
2044 const nir_intrinsic_instr *instr)
2045 {
2046 return enter_waterfall(ctx, wctx, get_src(ctx, instr->src[0]),
2047 nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
2048 }
2049
2050 static LLVMValueRef visit_load_ubo_buffer(struct ac_nir_context *ctx,
2051 nir_intrinsic_instr *instr)
2052 {
2053 struct waterfall_context wctx;
2054 LLVMValueRef rsrc_base = enter_waterfall_ubo(ctx, &wctx, instr);
2055
2056 LLVMValueRef ret;
2057 LLVMValueRef rsrc = rsrc_base;
2058 LLVMValueRef offset = get_src(ctx, instr->src[1]);
2059 int num_components = instr->num_components;
2060
2061 if (ctx->abi->load_ubo)
2062 rsrc = ctx->abi->load_ubo(ctx->abi, rsrc);
2063
2064 if (instr->dest.ssa.bit_size == 64)
2065 num_components *= 2;
2066
2067 if (instr->dest.ssa.bit_size == 16 || instr->dest.ssa.bit_size == 8) {
2068 unsigned load_bytes = instr->dest.ssa.bit_size / 8;
2069 LLVMValueRef results[num_components];
2070 for (unsigned i = 0; i < num_components; ++i) {
2071 LLVMValueRef immoffset = LLVMConstInt(ctx->ac.i32,
2072 load_bytes * i, 0);
2073
2074 if (load_bytes == 1) {
2075 results[i] = ac_build_tbuffer_load_byte(&ctx->ac,
2076 rsrc,
2077 offset,
2078 ctx->ac.i32_0,
2079 immoffset,
2080 0);
2081 } else {
2082 assert(load_bytes == 2);
2083 results[i] = ac_build_tbuffer_load_short(&ctx->ac,
2084 rsrc,
2085 offset,
2086 ctx->ac.i32_0,
2087 immoffset,
2088 0);
2089 }
2090 }
2091 ret = ac_build_gather_values(&ctx->ac, results, num_components);
2092 } else {
2093 ret = ac_build_buffer_load(&ctx->ac, rsrc, num_components, NULL, offset,
2094 NULL, 0, 0, true, true);
2095
2096 ret = ac_trim_vector(&ctx->ac, ret, num_components);
2097 }
2098
2099 ret = LLVMBuildBitCast(ctx->ac.builder, ret,
2100 get_def_type(ctx, &instr->dest.ssa), "");
2101
2102 return exit_waterfall(ctx, &wctx, ret);
2103 }
2104
2105 static void
2106 get_deref_offset(struct ac_nir_context *ctx, nir_deref_instr *instr,
2107 bool vs_in, unsigned *vertex_index_out,
2108 LLVMValueRef *vertex_index_ref,
2109 unsigned *const_out, LLVMValueRef *indir_out)
2110 {
2111 nir_variable *var = nir_deref_instr_get_variable(instr);
2112 nir_deref_path path;
2113 unsigned idx_lvl = 1;
2114
2115 nir_deref_path_init(&path, instr, NULL);
2116
2117 if (vertex_index_out != NULL || vertex_index_ref != NULL) {
2118 if (vertex_index_ref) {
2119 *vertex_index_ref = get_src(ctx, path.path[idx_lvl]->arr.index);
2120 if (vertex_index_out)
2121 *vertex_index_out = 0;
2122 } else {
2123 *vertex_index_out = nir_src_as_uint(path.path[idx_lvl]->arr.index);
2124 }
2125 ++idx_lvl;
2126 }
2127
2128 uint32_t const_offset = 0;
2129 LLVMValueRef offset = NULL;
2130
2131 if (var->data.compact) {
2132 assert(instr->deref_type == nir_deref_type_array);
2133 const_offset = nir_src_as_uint(instr->arr.index);
2134 goto out;
2135 }
2136
2137 for (; path.path[idx_lvl]; ++idx_lvl) {
2138 const struct glsl_type *parent_type = path.path[idx_lvl - 1]->type;
2139 if (path.path[idx_lvl]->deref_type == nir_deref_type_struct) {
2140 unsigned index = path.path[idx_lvl]->strct.index;
2141
2142 for (unsigned i = 0; i < index; i++) {
2143 const struct glsl_type *ft = glsl_get_struct_field(parent_type, i);
2144 const_offset += glsl_count_attribute_slots(ft, vs_in);
2145 }
2146 } else if(path.path[idx_lvl]->deref_type == nir_deref_type_array) {
2147 unsigned size = glsl_count_attribute_slots(path.path[idx_lvl]->type, vs_in);
2148 if (nir_src_is_const(path.path[idx_lvl]->arr.index)) {
2149 const_offset += size *
2150 nir_src_as_uint(path.path[idx_lvl]->arr.index);
2151 } else {
2152 LLVMValueRef array_off = LLVMBuildMul(ctx->ac.builder, LLVMConstInt(ctx->ac.i32, size, 0),
2153 get_src(ctx, path.path[idx_lvl]->arr.index), "");
2154 if (offset)
2155 offset = LLVMBuildAdd(ctx->ac.builder, offset, array_off, "");
2156 else
2157 offset = array_off;
2158 }
2159 } else
2160 unreachable("Uhandled deref type in get_deref_instr_offset");
2161 }
2162
2163 out:
2164 nir_deref_path_finish(&path);
2165
2166 if (const_offset && offset)
2167 offset = LLVMBuildAdd(ctx->ac.builder, offset,
2168 LLVMConstInt(ctx->ac.i32, const_offset, 0),
2169 "");
2170
2171 *const_out = const_offset;
2172 *indir_out = offset;
2173 }
2174
2175 static LLVMValueRef load_tess_varyings(struct ac_nir_context *ctx,
2176 nir_intrinsic_instr *instr,
2177 bool load_inputs)
2178 {
2179 LLVMValueRef result;
2180 LLVMValueRef vertex_index = NULL;
2181 LLVMValueRef indir_index = NULL;
2182 unsigned const_index = 0;
2183
2184 nir_variable *var = nir_deref_instr_get_variable(nir_instr_as_deref(instr->src[0].ssa->parent_instr));
2185
2186 unsigned location = var->data.location;
2187 unsigned driver_location = var->data.driver_location;
2188 const bool is_patch = var->data.patch ||
2189 var->data.location == VARYING_SLOT_TESS_LEVEL_INNER ||
2190 var->data.location == VARYING_SLOT_TESS_LEVEL_OUTER;
2191 const bool is_compact = var->data.compact;
2192
2193 get_deref_offset(ctx, nir_instr_as_deref(instr->src[0].ssa->parent_instr),
2194 false, NULL, is_patch ? NULL : &vertex_index,
2195 &const_index, &indir_index);
2196
2197 LLVMTypeRef dest_type = get_def_type(ctx, &instr->dest.ssa);
2198
2199 LLVMTypeRef src_component_type;
2200 if (LLVMGetTypeKind(dest_type) == LLVMVectorTypeKind)
2201 src_component_type = LLVMGetElementType(dest_type);
2202 else
2203 src_component_type = dest_type;
2204
2205 result = ctx->abi->load_tess_varyings(ctx->abi, src_component_type,
2206 vertex_index, indir_index,
2207 const_index, location, driver_location,
2208 var->data.location_frac,
2209 instr->num_components,
2210 is_patch, is_compact, load_inputs);
2211 if (instr->dest.ssa.bit_size == 16) {
2212 result = ac_to_integer(&ctx->ac, result);
2213 result = LLVMBuildTrunc(ctx->ac.builder, result, dest_type, "");
2214 }
2215 return LLVMBuildBitCast(ctx->ac.builder, result, dest_type, "");
2216 }
2217
2218 static unsigned
2219 type_scalar_size_bytes(const struct glsl_type *type)
2220 {
2221 assert(glsl_type_is_vector_or_scalar(type) ||
2222 glsl_type_is_matrix(type));
2223 return glsl_type_is_boolean(type) ? 4 : glsl_get_bit_size(type) / 8;
2224 }
2225
2226 static LLVMValueRef visit_load_var(struct ac_nir_context *ctx,
2227 nir_intrinsic_instr *instr)
2228 {
2229 nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2230 nir_variable *var = nir_deref_instr_get_variable(deref);
2231
2232 LLVMValueRef values[8];
2233 int idx = 0;
2234 int ve = instr->dest.ssa.num_components;
2235 unsigned comp = 0;
2236 LLVMValueRef indir_index;
2237 LLVMValueRef ret;
2238 unsigned const_index;
2239 unsigned stride = 4;
2240 int mode = deref->mode;
2241
2242 if (var) {
2243 bool vs_in = ctx->stage == MESA_SHADER_VERTEX &&
2244 var->data.mode == nir_var_shader_in;
2245 idx = var->data.driver_location;
2246 comp = var->data.location_frac;
2247 mode = var->data.mode;
2248
2249 get_deref_offset(ctx, deref, vs_in, NULL, NULL,
2250 &const_index, &indir_index);
2251
2252 if (var->data.compact) {
2253 stride = 1;
2254 const_index += comp;
2255 comp = 0;
2256 }
2257 }
2258
2259 if (instr->dest.ssa.bit_size == 64 &&
2260 (deref->mode == nir_var_shader_in ||
2261 deref->mode == nir_var_shader_out ||
2262 deref->mode == nir_var_function_temp))
2263 ve *= 2;
2264
2265 switch (mode) {
2266 case nir_var_shader_in:
2267 if (ctx->stage == MESA_SHADER_TESS_CTRL ||
2268 ctx->stage == MESA_SHADER_TESS_EVAL) {
2269 return load_tess_varyings(ctx, instr, true);
2270 }
2271
2272 if (ctx->stage == MESA_SHADER_GEOMETRY) {
2273 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
2274 LLVMValueRef indir_index;
2275 unsigned const_index, vertex_index;
2276 get_deref_offset(ctx, deref, false, &vertex_index, NULL,
2277 &const_index, &indir_index);
2278 assert(indir_index == NULL);
2279
2280 return ctx->abi->load_inputs(ctx->abi, var->data.location,
2281 var->data.driver_location,
2282 var->data.location_frac,
2283 instr->num_components, vertex_index, const_index, type);
2284 }
2285
2286 for (unsigned chan = comp; chan < ve + comp; chan++) {
2287 if (indir_index) {
2288 unsigned count = glsl_count_attribute_slots(
2289 var->type,
2290 ctx->stage == MESA_SHADER_VERTEX);
2291 count -= chan / 4;
2292 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2293 &ctx->ac, ctx->abi->inputs + idx + chan, count,
2294 stride, false, true);
2295
2296 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2297 tmp_vec,
2298 indir_index, "");
2299 } else
2300 values[chan] = ctx->abi->inputs[idx + chan + const_index * stride];
2301 }
2302 break;
2303 case nir_var_function_temp:
2304 for (unsigned chan = 0; chan < ve; chan++) {
2305 if (indir_index) {
2306 unsigned count = glsl_count_attribute_slots(
2307 var->type, false);
2308 count -= chan / 4;
2309 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2310 &ctx->ac, ctx->locals + idx + chan, count,
2311 stride, true, true);
2312
2313 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2314 tmp_vec,
2315 indir_index, "");
2316 } else {
2317 values[chan] = LLVMBuildLoad(ctx->ac.builder, ctx->locals[idx + chan + const_index * stride], "");
2318 }
2319 }
2320 break;
2321 case nir_var_shader_out:
2322 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
2323 return load_tess_varyings(ctx, instr, false);
2324 }
2325
2326 if (ctx->stage == MESA_SHADER_FRAGMENT &&
2327 var->data.fb_fetch_output &&
2328 ctx->abi->emit_fbfetch)
2329 return ctx->abi->emit_fbfetch(ctx->abi);
2330
2331 for (unsigned chan = comp; chan < ve + comp; chan++) {
2332 if (indir_index) {
2333 unsigned count = glsl_count_attribute_slots(
2334 var->type, false);
2335 count -= chan / 4;
2336 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2337 &ctx->ac, ctx->abi->outputs + idx + chan, count,
2338 stride, true, true);
2339
2340 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2341 tmp_vec,
2342 indir_index, "");
2343 } else {
2344 values[chan] = LLVMBuildLoad(ctx->ac.builder,
2345 ctx->abi->outputs[idx + chan + const_index * stride],
2346 "");
2347 }
2348 }
2349 break;
2350 case nir_var_mem_global: {
2351 LLVMValueRef address = get_src(ctx, instr->src[0]);
2352 LLVMTypeRef result_type = get_def_type(ctx, &instr->dest.ssa);
2353 unsigned explicit_stride = glsl_get_explicit_stride(deref->type);
2354 unsigned natural_stride = type_scalar_size_bytes(deref->type);
2355 unsigned stride = explicit_stride ? explicit_stride : natural_stride;
2356 int elem_size_bytes = ac_get_elem_bits(&ctx->ac, result_type) / 8;
2357 bool split_loads = ctx->ac.chip_class == GFX6 && elem_size_bytes < 4;
2358
2359 if (stride != natural_stride || split_loads) {
2360 if (LLVMGetTypeKind(result_type) == LLVMVectorTypeKind)
2361 result_type = LLVMGetElementType(result_type);
2362
2363 LLVMTypeRef ptr_type = LLVMPointerType(result_type,
2364 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2365 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2366
2367 for (unsigned i = 0; i < instr->dest.ssa.num_components; ++i) {
2368 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, i * stride / natural_stride, 0);
2369 values[i] = LLVMBuildLoad(ctx->ac.builder,
2370 ac_build_gep_ptr(&ctx->ac, address, offset), "");
2371 }
2372 return ac_build_gather_values(&ctx->ac, values, instr->dest.ssa.num_components);
2373 } else {
2374 LLVMTypeRef ptr_type = LLVMPointerType(result_type,
2375 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2376 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2377 LLVMValueRef val = LLVMBuildLoad(ctx->ac.builder, address, "");
2378 return val;
2379 }
2380 }
2381 default:
2382 unreachable("unhandle variable mode");
2383 }
2384 ret = ac_build_varying_gather_values(&ctx->ac, values, ve, comp);
2385 return LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->dest.ssa), "");
2386 }
2387
2388 static void
2389 visit_store_var(struct ac_nir_context *ctx,
2390 nir_intrinsic_instr *instr)
2391 {
2392 if (ctx->ac.postponed_kill) {
2393 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
2394 ctx->ac.postponed_kill, "");
2395 ac_build_ifcc(&ctx->ac, cond, 7002);
2396 }
2397
2398 nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2399 nir_variable *var = nir_deref_instr_get_variable(deref);
2400
2401 LLVMValueRef temp_ptr, value;
2402 int idx = 0;
2403 unsigned comp = 0;
2404 LLVMValueRef src = ac_to_float(&ctx->ac, get_src(ctx, instr->src[1]));
2405 int writemask = instr->const_index[0];
2406 LLVMValueRef indir_index;
2407 unsigned const_index;
2408
2409 if (var) {
2410 get_deref_offset(ctx, deref, false,
2411 NULL, NULL, &const_index, &indir_index);
2412 idx = var->data.driver_location;
2413 comp = var->data.location_frac;
2414
2415 if (var->data.compact) {
2416 const_index += comp;
2417 comp = 0;
2418 }
2419 }
2420
2421 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src)) == 64 &&
2422 (deref->mode == nir_var_shader_out ||
2423 deref->mode == nir_var_function_temp)) {
2424
2425 src = LLVMBuildBitCast(ctx->ac.builder, src,
2426 LLVMVectorType(ctx->ac.f32, ac_get_llvm_num_components(src) * 2),
2427 "");
2428
2429 writemask = widen_mask(writemask, 2);
2430 }
2431
2432 writemask = writemask << comp;
2433
2434 switch (deref->mode) {
2435 case nir_var_shader_out:
2436
2437 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
2438 LLVMValueRef vertex_index = NULL;
2439 LLVMValueRef indir_index = NULL;
2440 unsigned const_index = 0;
2441 const bool is_patch = var->data.patch ||
2442 var->data.location == VARYING_SLOT_TESS_LEVEL_INNER ||
2443 var->data.location == VARYING_SLOT_TESS_LEVEL_OUTER;
2444
2445 get_deref_offset(ctx, deref, false, NULL,
2446 is_patch ? NULL : &vertex_index,
2447 &const_index, &indir_index);
2448
2449 ctx->abi->store_tcs_outputs(ctx->abi, var,
2450 vertex_index, indir_index,
2451 const_index, src, writemask);
2452 break;
2453 }
2454
2455 for (unsigned chan = 0; chan < 8; chan++) {
2456 int stride = 4;
2457 if (!(writemask & (1 << chan)))
2458 continue;
2459
2460 value = ac_llvm_extract_elem(&ctx->ac, src, chan - comp);
2461
2462 if (var->data.compact)
2463 stride = 1;
2464 if (indir_index) {
2465 unsigned count = glsl_count_attribute_slots(
2466 var->type, false);
2467 count -= chan / 4;
2468 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2469 &ctx->ac, ctx->abi->outputs + idx + chan, count,
2470 stride, true, true);
2471
2472 tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
2473 value, indir_index, "");
2474 build_store_values_extended(&ctx->ac, ctx->abi->outputs + idx + chan,
2475 count, stride, tmp_vec);
2476
2477 } else {
2478 temp_ptr = ctx->abi->outputs[idx + chan + const_index * stride];
2479
2480 LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
2481 }
2482 }
2483 break;
2484 case nir_var_function_temp:
2485 for (unsigned chan = 0; chan < 8; chan++) {
2486 if (!(writemask & (1 << chan)))
2487 continue;
2488
2489 value = ac_llvm_extract_elem(&ctx->ac, src, chan);
2490 if (indir_index) {
2491 unsigned count = glsl_count_attribute_slots(
2492 var->type, false);
2493 count -= chan / 4;
2494 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2495 &ctx->ac, ctx->locals + idx + chan, count,
2496 4, true, true);
2497
2498 tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
2499 value, indir_index, "");
2500 build_store_values_extended(&ctx->ac, ctx->locals + idx + chan,
2501 count, 4, tmp_vec);
2502 } else {
2503 temp_ptr = ctx->locals[idx + chan + const_index * 4];
2504
2505 LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
2506 }
2507 }
2508 break;
2509
2510 case nir_var_mem_global: {
2511 int writemask = instr->const_index[0];
2512 LLVMValueRef address = get_src(ctx, instr->src[0]);
2513 LLVMValueRef val = get_src(ctx, instr->src[1]);
2514
2515 unsigned explicit_stride = glsl_get_explicit_stride(deref->type);
2516 unsigned natural_stride = type_scalar_size_bytes(deref->type);
2517 unsigned stride = explicit_stride ? explicit_stride : natural_stride;
2518 int elem_size_bytes = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(val)) / 8;
2519 bool split_stores = ctx->ac.chip_class == GFX6 && elem_size_bytes < 4;
2520
2521 LLVMTypeRef ptr_type = LLVMPointerType(LLVMTypeOf(val),
2522 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2523 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2524
2525 if (writemask == (1u << ac_get_llvm_num_components(val)) - 1 &&
2526 stride == natural_stride && !split_stores) {
2527 LLVMTypeRef ptr_type = LLVMPointerType(LLVMTypeOf(val),
2528 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2529 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2530
2531 val = LLVMBuildBitCast(ctx->ac.builder, val,
2532 LLVMGetElementType(LLVMTypeOf(address)), "");
2533 LLVMBuildStore(ctx->ac.builder, val, address);
2534 } else {
2535 LLVMTypeRef val_type = LLVMTypeOf(val);
2536 if (LLVMGetTypeKind(LLVMTypeOf(val)) == LLVMVectorTypeKind)
2537 val_type = LLVMGetElementType(val_type);
2538
2539 LLVMTypeRef ptr_type = LLVMPointerType(val_type,
2540 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2541 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2542 for (unsigned chan = 0; chan < 4; chan++) {
2543 if (!(writemask & (1 << chan)))
2544 continue;
2545
2546 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, chan * stride / natural_stride, 0);
2547
2548 LLVMValueRef ptr = ac_build_gep_ptr(&ctx->ac, address, offset);
2549 LLVMValueRef src = ac_llvm_extract_elem(&ctx->ac, val,
2550 chan);
2551 src = LLVMBuildBitCast(ctx->ac.builder, src,
2552 LLVMGetElementType(LLVMTypeOf(ptr)), "");
2553 LLVMBuildStore(ctx->ac.builder, src, ptr);
2554 }
2555 }
2556 break;
2557 }
2558 default:
2559 abort();
2560 break;
2561 }
2562
2563 if (ctx->ac.postponed_kill)
2564 ac_build_endif(&ctx->ac, 7002);
2565 }
2566
2567 static int image_type_to_components_count(enum glsl_sampler_dim dim, bool array)
2568 {
2569 switch (dim) {
2570 case GLSL_SAMPLER_DIM_BUF:
2571 return 1;
2572 case GLSL_SAMPLER_DIM_1D:
2573 return array ? 2 : 1;
2574 case GLSL_SAMPLER_DIM_2D:
2575 return array ? 3 : 2;
2576 case GLSL_SAMPLER_DIM_MS:
2577 return array ? 4 : 3;
2578 case GLSL_SAMPLER_DIM_3D:
2579 case GLSL_SAMPLER_DIM_CUBE:
2580 return 3;
2581 case GLSL_SAMPLER_DIM_RECT:
2582 case GLSL_SAMPLER_DIM_SUBPASS:
2583 return 2;
2584 case GLSL_SAMPLER_DIM_SUBPASS_MS:
2585 return 3;
2586 default:
2587 break;
2588 }
2589 return 0;
2590 }
2591
2592 static LLVMValueRef adjust_sample_index_using_fmask(struct ac_llvm_context *ctx,
2593 LLVMValueRef coord_x, LLVMValueRef coord_y,
2594 LLVMValueRef coord_z,
2595 LLVMValueRef sample_index,
2596 LLVMValueRef fmask_desc_ptr)
2597 {
2598 unsigned sample_chan = coord_z ? 3 : 2;
2599 LLVMValueRef addr[4] = {coord_x, coord_y, coord_z};
2600 addr[sample_chan] = sample_index;
2601
2602 ac_apply_fmask_to_sample(ctx, fmask_desc_ptr, addr, coord_z != NULL);
2603 return addr[sample_chan];
2604 }
2605
2606 static nir_deref_instr *get_image_deref(const nir_intrinsic_instr *instr)
2607 {
2608 assert(instr->src[0].is_ssa);
2609 return nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2610 }
2611
2612 static LLVMValueRef get_image_descriptor(struct ac_nir_context *ctx,
2613 const nir_intrinsic_instr *instr,
2614 LLVMValueRef dynamic_index,
2615 enum ac_descriptor_type desc_type,
2616 bool write)
2617 {
2618 nir_deref_instr *deref_instr =
2619 instr->src[0].ssa->parent_instr->type == nir_instr_type_deref ?
2620 nir_instr_as_deref(instr->src[0].ssa->parent_instr) : NULL;
2621
2622 return get_sampler_desc(ctx, deref_instr, desc_type, &instr->instr, dynamic_index, true, write);
2623 }
2624
2625 static void get_image_coords(struct ac_nir_context *ctx,
2626 const nir_intrinsic_instr *instr,
2627 LLVMValueRef dynamic_desc_index,
2628 struct ac_image_args *args,
2629 enum glsl_sampler_dim dim,
2630 bool is_array)
2631 {
2632 LLVMValueRef src0 = get_src(ctx, instr->src[1]);
2633 LLVMValueRef masks[] = {
2634 LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
2635 LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
2636 };
2637 LLVMValueRef sample_index = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
2638
2639 int count;
2640 ASSERTED bool add_frag_pos = (dim == GLSL_SAMPLER_DIM_SUBPASS ||
2641 dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2642 bool is_ms = (dim == GLSL_SAMPLER_DIM_MS ||
2643 dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2644 bool gfx9_1d = ctx->ac.chip_class == GFX9 && dim == GLSL_SAMPLER_DIM_1D;
2645 assert(!add_frag_pos && "Input attachments should be lowered by this point.");
2646 count = image_type_to_components_count(dim, is_array);
2647
2648 if (is_ms && (instr->intrinsic == nir_intrinsic_image_deref_load ||
2649 instr->intrinsic == nir_intrinsic_bindless_image_load)) {
2650 LLVMValueRef fmask_load_address[3];
2651
2652 fmask_load_address[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
2653 fmask_load_address[1] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[1], "");
2654 if (is_array)
2655 fmask_load_address[2] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[2], "");
2656 else
2657 fmask_load_address[2] = NULL;
2658
2659 sample_index = adjust_sample_index_using_fmask(&ctx->ac,
2660 fmask_load_address[0],
2661 fmask_load_address[1],
2662 fmask_load_address[2],
2663 sample_index,
2664 get_sampler_desc(ctx, nir_instr_as_deref(instr->src[0].ssa->parent_instr),
2665 AC_DESC_FMASK, &instr->instr, dynamic_desc_index, true, false));
2666 }
2667 if (count == 1 && !gfx9_1d) {
2668 if (instr->src[1].ssa->num_components)
2669 args->coords[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
2670 else
2671 args->coords[0] = src0;
2672 } else {
2673 int chan;
2674 if (is_ms)
2675 count--;
2676 for (chan = 0; chan < count; ++chan) {
2677 args->coords[chan] = ac_llvm_extract_elem(&ctx->ac, src0, chan);
2678 }
2679
2680 if (gfx9_1d) {
2681 if (is_array) {
2682 args->coords[2] = args->coords[1];
2683 args->coords[1] = ctx->ac.i32_0;
2684 } else
2685 args->coords[1] = ctx->ac.i32_0;
2686 count++;
2687 }
2688 if (ctx->ac.chip_class == GFX9 &&
2689 dim == GLSL_SAMPLER_DIM_2D &&
2690 !is_array) {
2691 /* The hw can't bind a slice of a 3D image as a 2D
2692 * image, because it ignores BASE_ARRAY if the target
2693 * is 3D. The workaround is to read BASE_ARRAY and set
2694 * it as the 3rd address operand for all 2D images.
2695 */
2696 LLVMValueRef first_layer, const5, mask;
2697
2698 const5 = LLVMConstInt(ctx->ac.i32, 5, 0);
2699 mask = LLVMConstInt(ctx->ac.i32, S_008F24_BASE_ARRAY(~0), 0);
2700 first_layer = LLVMBuildExtractElement(ctx->ac.builder, args->resource, const5, "");
2701 first_layer = LLVMBuildAnd(ctx->ac.builder, first_layer, mask, "");
2702
2703 args->coords[count] = first_layer;
2704 count++;
2705 }
2706
2707
2708 if (is_ms) {
2709 args->coords[count] = sample_index;
2710 count++;
2711 }
2712 }
2713 }
2714
2715 static LLVMValueRef get_image_buffer_descriptor(struct ac_nir_context *ctx,
2716 const nir_intrinsic_instr *instr,
2717 LLVMValueRef dynamic_index,
2718 bool write, bool atomic)
2719 {
2720 LLVMValueRef rsrc = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_BUFFER, write);
2721 if (ctx->ac.chip_class == GFX9 && LLVM_VERSION_MAJOR < 9 && atomic) {
2722 LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 2, 0), "");
2723 LLVMValueRef stride = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 1, 0), "");
2724 stride = LLVMBuildLShr(ctx->ac.builder, stride, LLVMConstInt(ctx->ac.i32, 16, 0), "");
2725
2726 LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->ac.builder,
2727 LLVMBuildICmp(ctx->ac.builder, LLVMIntUGT, elem_count, stride, ""),
2728 elem_count, stride, "");
2729
2730 rsrc = LLVMBuildInsertElement(ctx->ac.builder, rsrc, new_elem_count,
2731 LLVMConstInt(ctx->ac.i32, 2, 0), "");
2732 }
2733 return rsrc;
2734 }
2735
2736 static LLVMValueRef enter_waterfall_image(struct ac_nir_context *ctx,
2737 struct waterfall_context *wctx,
2738 const nir_intrinsic_instr *instr)
2739 {
2740 nir_deref_instr *deref_instr = NULL;
2741
2742 if (instr->src[0].ssa->parent_instr->type == nir_instr_type_deref)
2743 deref_instr = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2744
2745 LLVMValueRef value = get_sampler_desc_index(ctx, deref_instr, &instr->instr, true);
2746 return enter_waterfall(ctx, wctx, value, nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
2747 }
2748
2749 static LLVMValueRef visit_image_load(struct ac_nir_context *ctx,
2750 const nir_intrinsic_instr *instr,
2751 bool bindless)
2752 {
2753 LLVMValueRef res;
2754
2755 enum glsl_sampler_dim dim;
2756 enum gl_access_qualifier access;
2757 bool is_array;
2758 if (bindless) {
2759 dim = nir_intrinsic_image_dim(instr);
2760 access = nir_intrinsic_access(instr);
2761 is_array = nir_intrinsic_image_array(instr);
2762 } else {
2763 const nir_deref_instr *image_deref = get_image_deref(instr);