ac/nir: use llvm.amdgcn.rcp for nir_op_frcp
[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
55 LLVMValueRef main_function;
56 LLVMBasicBlockRef continue_block;
57 LLVMBasicBlockRef break_block;
58
59 int num_locals;
60 LLVMValueRef *locals;
61 };
62
63 static LLVMValueRef get_sampler_desc_index(struct ac_nir_context *ctx,
64 nir_deref_instr *deref_instr,
65 const nir_instr *instr,
66 bool image);
67
68 static LLVMValueRef get_sampler_desc(struct ac_nir_context *ctx,
69 nir_deref_instr *deref_instr,
70 enum ac_descriptor_type desc_type,
71 const nir_instr *instr,
72 LLVMValueRef index,
73 bool image, bool write);
74
75 static void
76 build_store_values_extended(struct ac_llvm_context *ac,
77 LLVMValueRef *values,
78 unsigned value_count,
79 unsigned value_stride,
80 LLVMValueRef vec)
81 {
82 LLVMBuilderRef builder = ac->builder;
83 unsigned i;
84
85 for (i = 0; i < value_count; i++) {
86 LLVMValueRef ptr = values[i * value_stride];
87 LLVMValueRef index = LLVMConstInt(ac->i32, i, false);
88 LLVMValueRef value = LLVMBuildExtractElement(builder, vec, index, "");
89 LLVMBuildStore(builder, value, ptr);
90 }
91 }
92
93 static LLVMTypeRef get_def_type(struct ac_nir_context *ctx,
94 const nir_ssa_def *def)
95 {
96 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, def->bit_size);
97 if (def->num_components > 1) {
98 type = LLVMVectorType(type, def->num_components);
99 }
100 return type;
101 }
102
103 static LLVMValueRef get_src(struct ac_nir_context *nir, nir_src src)
104 {
105 assert(src.is_ssa);
106 return nir->ssa_defs[src.ssa->index];
107 }
108
109 static LLVMValueRef
110 get_memory_ptr(struct ac_nir_context *ctx, nir_src src, unsigned bit_size)
111 {
112 LLVMValueRef ptr = get_src(ctx, src);
113 ptr = LLVMBuildGEP(ctx->ac.builder, ctx->ac.lds, &ptr, 1, "");
114 int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
115
116 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, bit_size);
117
118 return LLVMBuildBitCast(ctx->ac.builder, ptr,
119 LLVMPointerType(type, addr_space), "");
120 }
121
122 static LLVMBasicBlockRef get_block(struct ac_nir_context *nir,
123 const struct nir_block *b)
124 {
125 struct hash_entry *entry = _mesa_hash_table_search(nir->defs, b);
126 return (LLVMBasicBlockRef)entry->data;
127 }
128
129 static LLVMValueRef get_alu_src(struct ac_nir_context *ctx,
130 nir_alu_src src,
131 unsigned num_components)
132 {
133 LLVMValueRef value = get_src(ctx, src.src);
134 bool need_swizzle = false;
135
136 assert(value);
137 unsigned src_components = ac_get_llvm_num_components(value);
138 for (unsigned i = 0; i < num_components; ++i) {
139 assert(src.swizzle[i] < src_components);
140 if (src.swizzle[i] != i)
141 need_swizzle = true;
142 }
143
144 if (need_swizzle || num_components != src_components) {
145 LLVMValueRef masks[] = {
146 LLVMConstInt(ctx->ac.i32, src.swizzle[0], false),
147 LLVMConstInt(ctx->ac.i32, src.swizzle[1], false),
148 LLVMConstInt(ctx->ac.i32, src.swizzle[2], false),
149 LLVMConstInt(ctx->ac.i32, src.swizzle[3], false)};
150
151 if (src_components > 1 && num_components == 1) {
152 value = LLVMBuildExtractElement(ctx->ac.builder, value,
153 masks[0], "");
154 } else if (src_components == 1 && num_components > 1) {
155 LLVMValueRef values[] = {value, value, value, value};
156 value = ac_build_gather_values(&ctx->ac, values, num_components);
157 } else {
158 LLVMValueRef swizzle = LLVMConstVector(masks, num_components);
159 value = LLVMBuildShuffleVector(ctx->ac.builder, value, value,
160 swizzle, "");
161 }
162 }
163 assert(!src.negate);
164 assert(!src.abs);
165 return value;
166 }
167
168 static LLVMValueRef emit_int_cmp(struct ac_llvm_context *ctx,
169 LLVMIntPredicate pred, LLVMValueRef src0,
170 LLVMValueRef src1)
171 {
172 LLVMValueRef result = LLVMBuildICmp(ctx->builder, pred, src0, src1, "");
173 return LLVMBuildSelect(ctx->builder, result,
174 LLVMConstInt(ctx->i32, 0xFFFFFFFF, false),
175 ctx->i32_0, "");
176 }
177
178 static LLVMValueRef emit_float_cmp(struct ac_llvm_context *ctx,
179 LLVMRealPredicate pred, LLVMValueRef src0,
180 LLVMValueRef src1)
181 {
182 LLVMValueRef result;
183 src0 = ac_to_float(ctx, src0);
184 src1 = ac_to_float(ctx, src1);
185 result = LLVMBuildFCmp(ctx->builder, pred, src0, src1, "");
186 return LLVMBuildSelect(ctx->builder, result,
187 LLVMConstInt(ctx->i32, 0xFFFFFFFF, false),
188 ctx->i32_0, "");
189 }
190
191 static LLVMValueRef emit_intrin_1f_param(struct ac_llvm_context *ctx,
192 const char *intrin,
193 LLVMTypeRef result_type,
194 LLVMValueRef src0)
195 {
196 char name[64];
197 LLVMValueRef params[] = {
198 ac_to_float(ctx, src0),
199 };
200
201 ASSERTED const int length = snprintf(name, sizeof(name), "%s.f%d", intrin,
202 ac_get_elem_bits(ctx, result_type));
203 assert(length < sizeof(name));
204 return ac_build_intrinsic(ctx, name, result_type, params, 1, AC_FUNC_ATTR_READNONE);
205 }
206
207 static LLVMValueRef emit_intrin_2f_param(struct ac_llvm_context *ctx,
208 const char *intrin,
209 LLVMTypeRef result_type,
210 LLVMValueRef src0, LLVMValueRef src1)
211 {
212 char name[64];
213 LLVMValueRef params[] = {
214 ac_to_float(ctx, src0),
215 ac_to_float(ctx, src1),
216 };
217
218 ASSERTED const int length = snprintf(name, sizeof(name), "%s.f%d", intrin,
219 ac_get_elem_bits(ctx, result_type));
220 assert(length < sizeof(name));
221 return ac_build_intrinsic(ctx, name, result_type, params, 2, AC_FUNC_ATTR_READNONE);
222 }
223
224 static LLVMValueRef emit_intrin_3f_param(struct ac_llvm_context *ctx,
225 const char *intrin,
226 LLVMTypeRef result_type,
227 LLVMValueRef src0, LLVMValueRef src1, LLVMValueRef src2)
228 {
229 char name[64];
230 LLVMValueRef params[] = {
231 ac_to_float(ctx, src0),
232 ac_to_float(ctx, src1),
233 ac_to_float(ctx, src2),
234 };
235
236 ASSERTED const int length = snprintf(name, sizeof(name), "%s.f%d", intrin,
237 ac_get_elem_bits(ctx, result_type));
238 assert(length < sizeof(name));
239 return ac_build_intrinsic(ctx, name, result_type, params, 3, AC_FUNC_ATTR_READNONE);
240 }
241
242 static LLVMValueRef emit_bcsel(struct ac_llvm_context *ctx,
243 LLVMValueRef src0, LLVMValueRef src1, LLVMValueRef src2)
244 {
245 LLVMTypeRef src1_type = LLVMTypeOf(src1);
246 LLVMTypeRef src2_type = LLVMTypeOf(src2);
247
248 assert(LLVMGetTypeKind(LLVMTypeOf(src0)) != LLVMVectorTypeKind);
249
250 if (LLVMGetTypeKind(src1_type) == LLVMPointerTypeKind &&
251 LLVMGetTypeKind(src2_type) != LLVMPointerTypeKind) {
252 src2 = LLVMBuildIntToPtr(ctx->builder, src2, src1_type, "");
253 } else if (LLVMGetTypeKind(src2_type) == LLVMPointerTypeKind &&
254 LLVMGetTypeKind(src1_type) != LLVMPointerTypeKind) {
255 src1 = LLVMBuildIntToPtr(ctx->builder, src1, src2_type, "");
256 }
257
258 LLVMValueRef v = LLVMBuildICmp(ctx->builder, LLVMIntNE, src0,
259 ctx->i32_0, "");
260 return LLVMBuildSelect(ctx->builder, v,
261 ac_to_integer_or_pointer(ctx, src1),
262 ac_to_integer_or_pointer(ctx, src2), "");
263 }
264
265 static LLVMValueRef emit_iabs(struct ac_llvm_context *ctx,
266 LLVMValueRef src0)
267 {
268 return ac_build_imax(ctx, src0, LLVMBuildNeg(ctx->builder, src0, ""));
269 }
270
271 static LLVMValueRef emit_uint_carry(struct ac_llvm_context *ctx,
272 const char *intrin,
273 LLVMValueRef src0, LLVMValueRef src1)
274 {
275 LLVMTypeRef ret_type;
276 LLVMTypeRef types[] = { ctx->i32, ctx->i1 };
277 LLVMValueRef res;
278 LLVMValueRef params[] = { src0, src1 };
279 ret_type = LLVMStructTypeInContext(ctx->context, types,
280 2, true);
281
282 res = ac_build_intrinsic(ctx, intrin, ret_type,
283 params, 2, AC_FUNC_ATTR_READNONE);
284
285 res = LLVMBuildExtractValue(ctx->builder, res, 1, "");
286 res = LLVMBuildZExt(ctx->builder, res, ctx->i32, "");
287 return res;
288 }
289
290 static LLVMValueRef emit_b2f(struct ac_llvm_context *ctx,
291 LLVMValueRef src0,
292 unsigned bitsize)
293 {
294 LLVMValueRef result = LLVMBuildAnd(ctx->builder, src0,
295 LLVMBuildBitCast(ctx->builder, LLVMConstReal(ctx->f32, 1.0), ctx->i32, ""),
296 "");
297 result = LLVMBuildBitCast(ctx->builder, result, ctx->f32, "");
298
299 switch (bitsize) {
300 case 16:
301 return LLVMBuildFPTrunc(ctx->builder, result, ctx->f16, "");
302 case 32:
303 return result;
304 case 64:
305 return LLVMBuildFPExt(ctx->builder, result, ctx->f64, "");
306 default:
307 unreachable("Unsupported bit size.");
308 }
309 }
310
311 static LLVMValueRef emit_f2b(struct ac_llvm_context *ctx,
312 LLVMValueRef src0)
313 {
314 src0 = ac_to_float(ctx, src0);
315 LLVMValueRef zero = LLVMConstNull(LLVMTypeOf(src0));
316 return LLVMBuildSExt(ctx->builder,
317 LLVMBuildFCmp(ctx->builder, LLVMRealUNE, src0, zero, ""),
318 ctx->i32, "");
319 }
320
321 static LLVMValueRef emit_b2i(struct ac_llvm_context *ctx,
322 LLVMValueRef src0,
323 unsigned bitsize)
324 {
325 LLVMValueRef result = LLVMBuildAnd(ctx->builder, src0, ctx->i32_1, "");
326
327 switch (bitsize) {
328 case 8:
329 return LLVMBuildTrunc(ctx->builder, result, ctx->i8, "");
330 case 16:
331 return LLVMBuildTrunc(ctx->builder, result, ctx->i16, "");
332 case 32:
333 return result;
334 case 64:
335 return LLVMBuildZExt(ctx->builder, result, ctx->i64, "");
336 default:
337 unreachable("Unsupported bit size.");
338 }
339 }
340
341 static LLVMValueRef emit_i2b(struct ac_llvm_context *ctx,
342 LLVMValueRef src0)
343 {
344 LLVMValueRef zero = LLVMConstNull(LLVMTypeOf(src0));
345 return LLVMBuildSExt(ctx->builder,
346 LLVMBuildICmp(ctx->builder, LLVMIntNE, src0, zero, ""),
347 ctx->i32, "");
348 }
349
350 static LLVMValueRef emit_f2f16(struct ac_llvm_context *ctx,
351 LLVMValueRef src0)
352 {
353 LLVMValueRef result;
354 LLVMValueRef cond = NULL;
355
356 src0 = ac_to_float(ctx, src0);
357 result = LLVMBuildFPTrunc(ctx->builder, src0, ctx->f16, "");
358
359 if (ctx->chip_class >= GFX8) {
360 LLVMValueRef args[2];
361 /* Check if the result is a denormal - and flush to 0 if so. */
362 args[0] = result;
363 args[1] = LLVMConstInt(ctx->i32, N_SUBNORMAL | P_SUBNORMAL, false);
364 cond = ac_build_intrinsic(ctx, "llvm.amdgcn.class.f16", ctx->i1, args, 2, AC_FUNC_ATTR_READNONE);
365 }
366
367 /* need to convert back up to f32 */
368 result = LLVMBuildFPExt(ctx->builder, result, ctx->f32, "");
369
370 if (ctx->chip_class >= GFX8)
371 result = LLVMBuildSelect(ctx->builder, cond, ctx->f32_0, result, "");
372 else {
373 /* for GFX6-GFX7 */
374 /* 0x38800000 is smallest half float value (2^-14) in 32-bit float,
375 * so compare the result and flush to 0 if it's smaller.
376 */
377 LLVMValueRef temp, cond2;
378 temp = emit_intrin_1f_param(ctx, "llvm.fabs", ctx->f32, result);
379 cond = LLVMBuildFCmp(ctx->builder, LLVMRealOGT,
380 LLVMBuildBitCast(ctx->builder, LLVMConstInt(ctx->i32, 0x38800000, false), ctx->f32, ""),
381 temp, "");
382 cond2 = LLVMBuildFCmp(ctx->builder, LLVMRealONE,
383 temp, ctx->f32_0, "");
384 cond = LLVMBuildAnd(ctx->builder, cond, cond2, "");
385 result = LLVMBuildSelect(ctx->builder, cond, ctx->f32_0, result, "");
386 }
387 return result;
388 }
389
390 static LLVMValueRef emit_umul_high(struct ac_llvm_context *ctx,
391 LLVMValueRef src0, LLVMValueRef src1)
392 {
393 LLVMValueRef dst64, result;
394 src0 = LLVMBuildZExt(ctx->builder, src0, ctx->i64, "");
395 src1 = LLVMBuildZExt(ctx->builder, src1, ctx->i64, "");
396
397 dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
398 dst64 = LLVMBuildLShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
399 result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
400 return result;
401 }
402
403 static LLVMValueRef emit_imul_high(struct ac_llvm_context *ctx,
404 LLVMValueRef src0, LLVMValueRef src1)
405 {
406 LLVMValueRef dst64, result;
407 src0 = LLVMBuildSExt(ctx->builder, src0, ctx->i64, "");
408 src1 = LLVMBuildSExt(ctx->builder, src1, ctx->i64, "");
409
410 dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
411 dst64 = LLVMBuildAShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
412 result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
413 return result;
414 }
415
416 static LLVMValueRef emit_bfm(struct ac_llvm_context *ctx,
417 LLVMValueRef bits, LLVMValueRef offset)
418 {
419 /* mask = ((1 << bits) - 1) << offset */
420 return LLVMBuildShl(ctx->builder,
421 LLVMBuildSub(ctx->builder,
422 LLVMBuildShl(ctx->builder,
423 ctx->i32_1,
424 bits, ""),
425 ctx->i32_1, ""),
426 offset, "");
427 }
428
429 static LLVMValueRef emit_bitfield_select(struct ac_llvm_context *ctx,
430 LLVMValueRef mask, LLVMValueRef insert,
431 LLVMValueRef base)
432 {
433 /* Calculate:
434 * (mask & insert) | (~mask & base) = base ^ (mask & (insert ^ base))
435 * Use the right-hand side, which the LLVM backend can convert to V_BFI.
436 */
437 return LLVMBuildXor(ctx->builder, base,
438 LLVMBuildAnd(ctx->builder, mask,
439 LLVMBuildXor(ctx->builder, insert, base, ""), ""), "");
440 }
441
442 static LLVMValueRef emit_pack_2x16(struct ac_llvm_context *ctx,
443 LLVMValueRef src0,
444 LLVMValueRef (*pack)(struct ac_llvm_context *ctx,
445 LLVMValueRef args[2]))
446 {
447 LLVMValueRef comp[2];
448
449 src0 = ac_to_float(ctx, src0);
450 comp[0] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_0, "");
451 comp[1] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_1, "");
452
453 return LLVMBuildBitCast(ctx->builder, pack(ctx, comp), ctx->i32, "");
454 }
455
456 static LLVMValueRef emit_unpack_half_2x16(struct ac_llvm_context *ctx,
457 LLVMValueRef src0)
458 {
459 LLVMValueRef const16 = LLVMConstInt(ctx->i32, 16, false);
460 LLVMValueRef temps[2], val;
461 int i;
462
463 for (i = 0; i < 2; i++) {
464 val = i == 1 ? LLVMBuildLShr(ctx->builder, src0, const16, "") : src0;
465 val = LLVMBuildTrunc(ctx->builder, val, ctx->i16, "");
466 val = LLVMBuildBitCast(ctx->builder, val, ctx->f16, "");
467 temps[i] = LLVMBuildFPExt(ctx->builder, val, ctx->f32, "");
468 }
469 return ac_build_gather_values(ctx, temps, 2);
470 }
471
472 static LLVMValueRef emit_ddxy(struct ac_nir_context *ctx,
473 nir_op op,
474 LLVMValueRef src0)
475 {
476 unsigned mask;
477 int idx;
478 LLVMValueRef result;
479
480 if (op == nir_op_fddx_fine)
481 mask = AC_TID_MASK_LEFT;
482 else if (op == nir_op_fddy_fine)
483 mask = AC_TID_MASK_TOP;
484 else
485 mask = AC_TID_MASK_TOP_LEFT;
486
487 /* for DDX we want to next X pixel, DDY next Y pixel. */
488 if (op == nir_op_fddx_fine ||
489 op == nir_op_fddx_coarse ||
490 op == nir_op_fddx)
491 idx = 1;
492 else
493 idx = 2;
494
495 result = ac_build_ddxy(&ctx->ac, mask, idx, src0);
496 return result;
497 }
498
499 struct waterfall_context {
500 LLVMBasicBlockRef phi_bb[2];
501 bool use_waterfall;
502 };
503
504 /* To deal with divergent descriptors we can create a loop that handles all
505 * lanes with the same descriptor on a given iteration (henceforth a
506 * waterfall loop).
507 *
508 * These helper create the begin and end of the loop leaving the caller
509 * to implement the body.
510 *
511 * params:
512 * - ctx is the usal nir context
513 * - wctx is a temporary struct containing some loop info. Can be left uninitialized.
514 * - value is the possibly divergent value for which we built the loop
515 * - divergent is whether value is actually divergent. If false we just pass
516 * things through.
517 */
518 static LLVMValueRef enter_waterfall(struct ac_nir_context *ctx,
519 struct waterfall_context *wctx,
520 LLVMValueRef value, bool divergent)
521 {
522 /* If the app claims the value is divergent but it is constant we can
523 * end up with a dynamic index of NULL. */
524 if (!value)
525 divergent = false;
526
527 wctx->use_waterfall = divergent;
528 if (!divergent)
529 return value;
530
531 ac_build_bgnloop(&ctx->ac, 6000);
532
533 LLVMValueRef scalar_value = ac_build_readlane(&ctx->ac, value, NULL);
534
535 LLVMValueRef active = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, value,
536 scalar_value, "uniform_active");
537
538 wctx->phi_bb[0] = LLVMGetInsertBlock(ctx->ac.builder);
539 ac_build_ifcc(&ctx->ac, active, 6001);
540
541 return scalar_value;
542 }
543
544 static LLVMValueRef exit_waterfall(struct ac_nir_context *ctx,
545 struct waterfall_context *wctx,
546 LLVMValueRef value)
547 {
548 LLVMValueRef ret = NULL;
549 LLVMValueRef phi_src[2];
550 LLVMValueRef cc_phi_src[2] = {
551 LLVMConstInt(ctx->ac.i32, 0, false),
552 LLVMConstInt(ctx->ac.i32, 0xffffffff, false),
553 };
554
555 if (!wctx->use_waterfall)
556 return value;
557
558 wctx->phi_bb[1] = LLVMGetInsertBlock(ctx->ac.builder);
559
560 ac_build_endif(&ctx->ac, 6001);
561
562 if (value) {
563 phi_src[0] = LLVMGetUndef(LLVMTypeOf(value));
564 phi_src[1] = value;
565
566 ret = ac_build_phi(&ctx->ac, LLVMTypeOf(value), 2, phi_src, wctx->phi_bb);
567 }
568
569 /*
570 * By using the optimization barrier on the exit decision, we decouple
571 * the operations from the break, and hence avoid LLVM hoisting the
572 * opteration into the break block.
573 */
574 LLVMValueRef cc = ac_build_phi(&ctx->ac, ctx->ac.i32, 2, cc_phi_src, wctx->phi_bb);
575 ac_build_optimization_barrier(&ctx->ac, &cc);
576
577 LLVMValueRef active = LLVMBuildICmp(ctx->ac.builder, LLVMIntNE, cc, ctx->ac.i32_0, "uniform_active2");
578 ac_build_ifcc(&ctx->ac, active, 6002);
579 ac_build_break(&ctx->ac);
580 ac_build_endif(&ctx->ac, 6002);
581
582 ac_build_endloop(&ctx->ac, 6000);
583 return ret;
584 }
585
586 static void visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
587 {
588 LLVMValueRef src[4], result = NULL;
589 unsigned num_components = instr->dest.dest.ssa.num_components;
590 unsigned src_components;
591 LLVMTypeRef def_type = get_def_type(ctx, &instr->dest.dest.ssa);
592
593 assert(nir_op_infos[instr->op].num_inputs <= ARRAY_SIZE(src));
594 switch (instr->op) {
595 case nir_op_vec2:
596 case nir_op_vec3:
597 case nir_op_vec4:
598 src_components = 1;
599 break;
600 case nir_op_pack_half_2x16:
601 case nir_op_pack_snorm_2x16:
602 case nir_op_pack_unorm_2x16:
603 src_components = 2;
604 break;
605 case nir_op_unpack_half_2x16:
606 src_components = 1;
607 break;
608 case nir_op_cube_face_coord:
609 case nir_op_cube_face_index:
610 src_components = 3;
611 break;
612 default:
613 src_components = num_components;
614 break;
615 }
616 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
617 src[i] = get_alu_src(ctx, instr->src[i], src_components);
618
619 switch (instr->op) {
620 case nir_op_mov:
621 result = src[0];
622 break;
623 case nir_op_fneg:
624 src[0] = ac_to_float(&ctx->ac, src[0]);
625 result = LLVMBuildFNeg(ctx->ac.builder, src[0], "");
626 if (ctx->ac.float_mode == AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO) {
627 /* fneg will be optimized by backend compiler with sign
628 * bit removed via XOR. This is probably a LLVM bug.
629 */
630 result = ac_build_canonicalize(&ctx->ac, result,
631 instr->dest.dest.ssa.bit_size);
632 }
633 break;
634 case nir_op_ineg:
635 result = LLVMBuildNeg(ctx->ac.builder, src[0], "");
636 break;
637 case nir_op_inot:
638 result = LLVMBuildNot(ctx->ac.builder, src[0], "");
639 break;
640 case nir_op_iadd:
641 result = LLVMBuildAdd(ctx->ac.builder, src[0], src[1], "");
642 break;
643 case nir_op_fadd:
644 src[0] = ac_to_float(&ctx->ac, src[0]);
645 src[1] = ac_to_float(&ctx->ac, src[1]);
646 result = LLVMBuildFAdd(ctx->ac.builder, src[0], src[1], "");
647 break;
648 case nir_op_fsub:
649 src[0] = ac_to_float(&ctx->ac, src[0]);
650 src[1] = ac_to_float(&ctx->ac, src[1]);
651 result = LLVMBuildFSub(ctx->ac.builder, src[0], src[1], "");
652 break;
653 case nir_op_isub:
654 result = LLVMBuildSub(ctx->ac.builder, src[0], src[1], "");
655 break;
656 case nir_op_imul:
657 result = LLVMBuildMul(ctx->ac.builder, src[0], src[1], "");
658 break;
659 case nir_op_imod:
660 result = LLVMBuildSRem(ctx->ac.builder, src[0], src[1], "");
661 break;
662 case nir_op_umod:
663 result = LLVMBuildURem(ctx->ac.builder, src[0], src[1], "");
664 break;
665 case nir_op_fmod:
666 /* lower_fmod only lower 16-bit and 32-bit fmod */
667 assert(instr->dest.dest.ssa.bit_size == 64);
668 src[0] = ac_to_float(&ctx->ac, src[0]);
669 src[1] = ac_to_float(&ctx->ac, src[1]);
670 result = ac_build_fdiv(&ctx->ac, src[0], src[1]);
671 result = emit_intrin_1f_param(&ctx->ac, "llvm.floor",
672 ac_to_float_type(&ctx->ac, def_type), result);
673 result = LLVMBuildFMul(ctx->ac.builder, src[1] , result, "");
674 result = LLVMBuildFSub(ctx->ac.builder, src[0], result, "");
675 break;
676 case nir_op_irem:
677 result = LLVMBuildSRem(ctx->ac.builder, src[0], src[1], "");
678 break;
679 case nir_op_idiv:
680 result = LLVMBuildSDiv(ctx->ac.builder, src[0], src[1], "");
681 break;
682 case nir_op_udiv:
683 result = LLVMBuildUDiv(ctx->ac.builder, src[0], src[1], "");
684 break;
685 case nir_op_fmul:
686 src[0] = ac_to_float(&ctx->ac, src[0]);
687 src[1] = ac_to_float(&ctx->ac, src[1]);
688 result = LLVMBuildFMul(ctx->ac.builder, src[0], src[1], "");
689 break;
690 case nir_op_frcp:
691 result = emit_intrin_1f_param(&ctx->ac, "llvm.amdgcn.rcp",
692 ac_to_float_type(&ctx->ac, def_type), src[0]);
693 break;
694 case nir_op_iand:
695 result = LLVMBuildAnd(ctx->ac.builder, src[0], src[1], "");
696 break;
697 case nir_op_ior:
698 result = LLVMBuildOr(ctx->ac.builder, src[0], src[1], "");
699 break;
700 case nir_op_ixor:
701 result = LLVMBuildXor(ctx->ac.builder, src[0], src[1], "");
702 break;
703 case nir_op_ishl:
704 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) < ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
705 src[1] = LLVMBuildZExt(ctx->ac.builder, src[1],
706 LLVMTypeOf(src[0]), "");
707 else if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) > ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
708 src[1] = LLVMBuildTrunc(ctx->ac.builder, src[1],
709 LLVMTypeOf(src[0]), "");
710 result = LLVMBuildShl(ctx->ac.builder, src[0], src[1], "");
711 break;
712 case nir_op_ishr:
713 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) < ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
714 src[1] = LLVMBuildZExt(ctx->ac.builder, src[1],
715 LLVMTypeOf(src[0]), "");
716 else if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) > ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
717 src[1] = LLVMBuildTrunc(ctx->ac.builder, src[1],
718 LLVMTypeOf(src[0]), "");
719 result = LLVMBuildAShr(ctx->ac.builder, src[0], src[1], "");
720 break;
721 case nir_op_ushr:
722 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) < ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
723 src[1] = LLVMBuildZExt(ctx->ac.builder, src[1],
724 LLVMTypeOf(src[0]), "");
725 else if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) > ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
726 src[1] = LLVMBuildTrunc(ctx->ac.builder, src[1],
727 LLVMTypeOf(src[0]), "");
728 result = LLVMBuildLShr(ctx->ac.builder, src[0], src[1], "");
729 break;
730 case nir_op_ilt32:
731 result = emit_int_cmp(&ctx->ac, LLVMIntSLT, src[0], src[1]);
732 break;
733 case nir_op_ine32:
734 result = emit_int_cmp(&ctx->ac, LLVMIntNE, src[0], src[1]);
735 break;
736 case nir_op_ieq32:
737 result = emit_int_cmp(&ctx->ac, LLVMIntEQ, src[0], src[1]);
738 break;
739 case nir_op_ige32:
740 result = emit_int_cmp(&ctx->ac, LLVMIntSGE, src[0], src[1]);
741 break;
742 case nir_op_ult32:
743 result = emit_int_cmp(&ctx->ac, LLVMIntULT, src[0], src[1]);
744 break;
745 case nir_op_uge32:
746 result = emit_int_cmp(&ctx->ac, LLVMIntUGE, src[0], src[1]);
747 break;
748 case nir_op_feq32:
749 result = emit_float_cmp(&ctx->ac, LLVMRealOEQ, src[0], src[1]);
750 break;
751 case nir_op_fne32:
752 result = emit_float_cmp(&ctx->ac, LLVMRealUNE, src[0], src[1]);
753 break;
754 case nir_op_flt32:
755 result = emit_float_cmp(&ctx->ac, LLVMRealOLT, src[0], src[1]);
756 break;
757 case nir_op_fge32:
758 result = emit_float_cmp(&ctx->ac, LLVMRealOGE, src[0], src[1]);
759 break;
760 case nir_op_fabs:
761 result = emit_intrin_1f_param(&ctx->ac, "llvm.fabs",
762 ac_to_float_type(&ctx->ac, def_type), src[0]);
763 if (ctx->ac.float_mode == AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO) {
764 /* fabs will be optimized by backend compiler with sign
765 * bit removed via AND.
766 */
767 result = ac_build_canonicalize(&ctx->ac, result,
768 instr->dest.dest.ssa.bit_size);
769 }
770 break;
771 case nir_op_iabs:
772 result = emit_iabs(&ctx->ac, src[0]);
773 break;
774 case nir_op_imax:
775 result = ac_build_imax(&ctx->ac, src[0], src[1]);
776 break;
777 case nir_op_imin:
778 result = ac_build_imin(&ctx->ac, src[0], src[1]);
779 break;
780 case nir_op_umax:
781 result = ac_build_umax(&ctx->ac, src[0], src[1]);
782 break;
783 case nir_op_umin:
784 result = ac_build_umin(&ctx->ac, src[0], src[1]);
785 break;
786 case nir_op_isign:
787 result = ac_build_isign(&ctx->ac, src[0],
788 instr->dest.dest.ssa.bit_size);
789 break;
790 case nir_op_fsign:
791 src[0] = ac_to_float(&ctx->ac, src[0]);
792 result = ac_build_fsign(&ctx->ac, src[0],
793 instr->dest.dest.ssa.bit_size);
794 break;
795 case nir_op_ffloor:
796 result = emit_intrin_1f_param(&ctx->ac, "llvm.floor",
797 ac_to_float_type(&ctx->ac, def_type), src[0]);
798 break;
799 case nir_op_ftrunc:
800 result = emit_intrin_1f_param(&ctx->ac, "llvm.trunc",
801 ac_to_float_type(&ctx->ac, def_type), src[0]);
802 break;
803 case nir_op_fceil:
804 result = emit_intrin_1f_param(&ctx->ac, "llvm.ceil",
805 ac_to_float_type(&ctx->ac, def_type), src[0]);
806 break;
807 case nir_op_fround_even:
808 result = emit_intrin_1f_param(&ctx->ac, "llvm.rint",
809 ac_to_float_type(&ctx->ac, def_type),src[0]);
810 break;
811 case nir_op_ffract:
812 src[0] = ac_to_float(&ctx->ac, src[0]);
813 result = ac_build_fract(&ctx->ac, src[0],
814 instr->dest.dest.ssa.bit_size);
815 break;
816 case nir_op_fsin:
817 result = emit_intrin_1f_param(&ctx->ac, "llvm.sin",
818 ac_to_float_type(&ctx->ac, def_type), src[0]);
819 break;
820 case nir_op_fcos:
821 result = emit_intrin_1f_param(&ctx->ac, "llvm.cos",
822 ac_to_float_type(&ctx->ac, def_type), src[0]);
823 break;
824 case nir_op_fsqrt:
825 result = emit_intrin_1f_param(&ctx->ac, "llvm.sqrt",
826 ac_to_float_type(&ctx->ac, def_type), src[0]);
827 break;
828 case nir_op_fexp2:
829 result = emit_intrin_1f_param(&ctx->ac, "llvm.exp2",
830 ac_to_float_type(&ctx->ac, def_type), src[0]);
831 break;
832 case nir_op_flog2:
833 result = emit_intrin_1f_param(&ctx->ac, "llvm.log2",
834 ac_to_float_type(&ctx->ac, def_type), src[0]);
835 break;
836 case nir_op_frsq:
837 result = emit_intrin_1f_param(&ctx->ac, "llvm.sqrt",
838 ac_to_float_type(&ctx->ac, def_type), src[0]);
839 result = ac_build_fdiv(&ctx->ac, LLVMConstReal(LLVMTypeOf(result), 1.0), result);
840 break;
841 case nir_op_frexp_exp:
842 src[0] = ac_to_float(&ctx->ac, src[0]);
843 result = ac_build_frexp_exp(&ctx->ac, src[0],
844 ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])));
845 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) == 16)
846 result = LLVMBuildSExt(ctx->ac.builder, result,
847 ctx->ac.i32, "");
848 break;
849 case nir_op_frexp_sig:
850 src[0] = ac_to_float(&ctx->ac, src[0]);
851 result = ac_build_frexp_mant(&ctx->ac, src[0],
852 instr->dest.dest.ssa.bit_size);
853 break;
854 case nir_op_fpow:
855 result = emit_intrin_2f_param(&ctx->ac, "llvm.pow",
856 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
857 break;
858 case nir_op_fmax:
859 result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
860 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
861 if (ctx->ac.chip_class < GFX9 &&
862 instr->dest.dest.ssa.bit_size == 32) {
863 /* Only pre-GFX9 chips do not flush denorms. */
864 result = ac_build_canonicalize(&ctx->ac, result,
865 instr->dest.dest.ssa.bit_size);
866 }
867 break;
868 case nir_op_fmin:
869 result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
870 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
871 if (ctx->ac.chip_class < GFX9 &&
872 instr->dest.dest.ssa.bit_size == 32) {
873 /* Only pre-GFX9 chips do not flush denorms. */
874 result = ac_build_canonicalize(&ctx->ac, result,
875 instr->dest.dest.ssa.bit_size);
876 }
877 break;
878 case nir_op_ffma:
879 /* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
880 result = emit_intrin_3f_param(&ctx->ac, ctx->ac.chip_class >= GFX10 ? "llvm.fma" : "llvm.fmuladd",
881 ac_to_float_type(&ctx->ac, def_type), src[0], src[1], src[2]);
882 break;
883 case nir_op_ldexp:
884 src[0] = ac_to_float(&ctx->ac, src[0]);
885 if (ac_get_elem_bits(&ctx->ac, def_type) == 32)
886 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f32", ctx->ac.f32, src, 2, AC_FUNC_ATTR_READNONE);
887 else if (ac_get_elem_bits(&ctx->ac, def_type) == 16)
888 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f16", ctx->ac.f16, src, 2, AC_FUNC_ATTR_READNONE);
889 else
890 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f64", ctx->ac.f64, src, 2, AC_FUNC_ATTR_READNONE);
891 break;
892 case nir_op_bfm:
893 result = emit_bfm(&ctx->ac, src[0], src[1]);
894 break;
895 case nir_op_bitfield_select:
896 result = emit_bitfield_select(&ctx->ac, src[0], src[1], src[2]);
897 break;
898 case nir_op_ubfe:
899 result = ac_build_bfe(&ctx->ac, src[0], src[1], src[2], false);
900 break;
901 case nir_op_ibfe:
902 result = ac_build_bfe(&ctx->ac, src[0], src[1], src[2], true);
903 break;
904 case nir_op_bitfield_reverse:
905 result = ac_build_bitfield_reverse(&ctx->ac, src[0]);
906 break;
907 case nir_op_bit_count:
908 result = ac_build_bit_count(&ctx->ac, src[0]);
909 break;
910 case nir_op_vec2:
911 case nir_op_vec3:
912 case nir_op_vec4:
913 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
914 src[i] = ac_to_integer(&ctx->ac, src[i]);
915 result = ac_build_gather_values(&ctx->ac, src, num_components);
916 break;
917 case nir_op_f2i8:
918 case nir_op_f2i16:
919 case nir_op_f2i32:
920 case nir_op_f2i64:
921 src[0] = ac_to_float(&ctx->ac, src[0]);
922 result = LLVMBuildFPToSI(ctx->ac.builder, src[0], def_type, "");
923 break;
924 case nir_op_f2u8:
925 case nir_op_f2u16:
926 case nir_op_f2u32:
927 case nir_op_f2u64:
928 src[0] = ac_to_float(&ctx->ac, src[0]);
929 result = LLVMBuildFPToUI(ctx->ac.builder, src[0], def_type, "");
930 break;
931 case nir_op_i2f16:
932 case nir_op_i2f32:
933 case nir_op_i2f64:
934 result = LLVMBuildSIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
935 break;
936 case nir_op_u2f16:
937 case nir_op_u2f32:
938 case nir_op_u2f64:
939 result = LLVMBuildUIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
940 break;
941 case nir_op_f2f16_rtz:
942 src[0] = ac_to_float(&ctx->ac, src[0]);
943 if (LLVMTypeOf(src[0]) == ctx->ac.f64)
944 src[0] = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ctx->ac.f32, "");
945 LLVMValueRef param[2] = { src[0], ctx->ac.f32_0 };
946 result = ac_build_cvt_pkrtz_f16(&ctx->ac, param);
947 result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
948 break;
949 case nir_op_f2f16_rtne:
950 case nir_op_f2f16:
951 case nir_op_f2f32:
952 case nir_op_f2f64:
953 src[0] = ac_to_float(&ctx->ac, src[0]);
954 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
955 result = LLVMBuildFPExt(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
956 else
957 result = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
958 break;
959 case nir_op_u2u8:
960 case nir_op_u2u16:
961 case nir_op_u2u32:
962 case nir_op_u2u64:
963 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
964 result = LLVMBuildZExt(ctx->ac.builder, src[0], def_type, "");
965 else
966 result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
967 break;
968 case nir_op_i2i8:
969 case nir_op_i2i16:
970 case nir_op_i2i32:
971 case nir_op_i2i64:
972 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
973 result = LLVMBuildSExt(ctx->ac.builder, src[0], def_type, "");
974 else
975 result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
976 break;
977 case nir_op_b32csel:
978 result = emit_bcsel(&ctx->ac, src[0], src[1], src[2]);
979 break;
980 case nir_op_find_lsb:
981 result = ac_find_lsb(&ctx->ac, ctx->ac.i32, src[0]);
982 break;
983 case nir_op_ufind_msb:
984 result = ac_build_umsb(&ctx->ac, src[0], ctx->ac.i32);
985 break;
986 case nir_op_ifind_msb:
987 result = ac_build_imsb(&ctx->ac, src[0], ctx->ac.i32);
988 break;
989 case nir_op_uadd_carry:
990 result = emit_uint_carry(&ctx->ac, "llvm.uadd.with.overflow.i32", src[0], src[1]);
991 break;
992 case nir_op_usub_borrow:
993 result = emit_uint_carry(&ctx->ac, "llvm.usub.with.overflow.i32", src[0], src[1]);
994 break;
995 case nir_op_b2f16:
996 case nir_op_b2f32:
997 case nir_op_b2f64:
998 result = emit_b2f(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
999 break;
1000 case nir_op_f2b32:
1001 result = emit_f2b(&ctx->ac, src[0]);
1002 break;
1003 case nir_op_b2i8:
1004 case nir_op_b2i16:
1005 case nir_op_b2i32:
1006 case nir_op_b2i64:
1007 result = emit_b2i(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
1008 break;
1009 case nir_op_i2b32:
1010 result = emit_i2b(&ctx->ac, src[0]);
1011 break;
1012 case nir_op_fquantize2f16:
1013 result = emit_f2f16(&ctx->ac, src[0]);
1014 break;
1015 case nir_op_umul_high:
1016 result = emit_umul_high(&ctx->ac, src[0], src[1]);
1017 break;
1018 case nir_op_imul_high:
1019 result = emit_imul_high(&ctx->ac, src[0], src[1]);
1020 break;
1021 case nir_op_pack_half_2x16:
1022 result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pkrtz_f16);
1023 break;
1024 case nir_op_pack_snorm_2x16:
1025 result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pknorm_i16);
1026 break;
1027 case nir_op_pack_unorm_2x16:
1028 result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pknorm_u16);
1029 break;
1030 case nir_op_unpack_half_2x16:
1031 result = emit_unpack_half_2x16(&ctx->ac, src[0]);
1032 break;
1033 case nir_op_fddx:
1034 case nir_op_fddy:
1035 case nir_op_fddx_fine:
1036 case nir_op_fddy_fine:
1037 case nir_op_fddx_coarse:
1038 case nir_op_fddy_coarse:
1039 result = emit_ddxy(ctx, instr->op, src[0]);
1040 break;
1041
1042 case nir_op_unpack_64_2x32_split_x: {
1043 assert(ac_get_llvm_num_components(src[0]) == 1);
1044 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1045 ctx->ac.v2i32,
1046 "");
1047 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1048 ctx->ac.i32_0, "");
1049 break;
1050 }
1051
1052 case nir_op_unpack_64_2x32_split_y: {
1053 assert(ac_get_llvm_num_components(src[0]) == 1);
1054 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1055 ctx->ac.v2i32,
1056 "");
1057 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1058 ctx->ac.i32_1, "");
1059 break;
1060 }
1061
1062 case nir_op_pack_64_2x32_split: {
1063 LLVMValueRef tmp = ac_build_gather_values(&ctx->ac, src, 2);
1064 result = LLVMBuildBitCast(ctx->ac.builder, tmp, ctx->ac.i64, "");
1065 break;
1066 }
1067
1068 case nir_op_pack_32_2x16_split: {
1069 LLVMValueRef tmp = ac_build_gather_values(&ctx->ac, src, 2);
1070 result = LLVMBuildBitCast(ctx->ac.builder, tmp, ctx->ac.i32, "");
1071 break;
1072 }
1073
1074 case nir_op_unpack_32_2x16_split_x: {
1075 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1076 ctx->ac.v2i16,
1077 "");
1078 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1079 ctx->ac.i32_0, "");
1080 break;
1081 }
1082
1083 case nir_op_unpack_32_2x16_split_y: {
1084 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1085 ctx->ac.v2i16,
1086 "");
1087 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1088 ctx->ac.i32_1, "");
1089 break;
1090 }
1091
1092 case nir_op_cube_face_coord: {
1093 src[0] = ac_to_float(&ctx->ac, src[0]);
1094 LLVMValueRef results[2];
1095 LLVMValueRef in[3];
1096 for (unsigned chan = 0; chan < 3; chan++)
1097 in[chan] = ac_llvm_extract_elem(&ctx->ac, src[0], chan);
1098 results[0] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubesc",
1099 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1100 results[1] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubetc",
1101 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1102 LLVMValueRef ma = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubema",
1103 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1104 results[0] = ac_build_fdiv(&ctx->ac, results[0], ma);
1105 results[1] = ac_build_fdiv(&ctx->ac, results[1], ma);
1106 LLVMValueRef offset = LLVMConstReal(ctx->ac.f32, 0.5);
1107 results[0] = LLVMBuildFAdd(ctx->ac.builder, results[0], offset, "");
1108 results[1] = LLVMBuildFAdd(ctx->ac.builder, results[1], offset, "");
1109 result = ac_build_gather_values(&ctx->ac, results, 2);
1110 break;
1111 }
1112
1113 case nir_op_cube_face_index: {
1114 src[0] = ac_to_float(&ctx->ac, src[0]);
1115 LLVMValueRef in[3];
1116 for (unsigned chan = 0; chan < 3; chan++)
1117 in[chan] = ac_llvm_extract_elem(&ctx->ac, src[0], chan);
1118 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubeid",
1119 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1120 break;
1121 }
1122
1123 case nir_op_fmin3:
1124 result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
1125 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
1126 result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
1127 ac_to_float_type(&ctx->ac, def_type), result, src[2]);
1128 break;
1129 case nir_op_umin3:
1130 result = ac_build_umin(&ctx->ac, src[0], src[1]);
1131 result = ac_build_umin(&ctx->ac, result, src[2]);
1132 break;
1133 case nir_op_imin3:
1134 result = ac_build_imin(&ctx->ac, src[0], src[1]);
1135 result = ac_build_imin(&ctx->ac, result, src[2]);
1136 break;
1137 case nir_op_fmax3:
1138 result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
1139 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
1140 result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
1141 ac_to_float_type(&ctx->ac, def_type), result, src[2]);
1142 break;
1143 case nir_op_umax3:
1144 result = ac_build_umax(&ctx->ac, src[0], src[1]);
1145 result = ac_build_umax(&ctx->ac, result, src[2]);
1146 break;
1147 case nir_op_imax3:
1148 result = ac_build_imax(&ctx->ac, src[0], src[1]);
1149 result = ac_build_imax(&ctx->ac, result, src[2]);
1150 break;
1151 case nir_op_fmed3: {
1152 src[0] = ac_to_float(&ctx->ac, src[0]);
1153 src[1] = ac_to_float(&ctx->ac, src[1]);
1154 src[2] = ac_to_float(&ctx->ac, src[2]);
1155 result = ac_build_fmed3(&ctx->ac, src[0], src[1], src[2],
1156 instr->dest.dest.ssa.bit_size);
1157 break;
1158 }
1159 case nir_op_imed3: {
1160 LLVMValueRef tmp1 = ac_build_imin(&ctx->ac, src[0], src[1]);
1161 LLVMValueRef tmp2 = ac_build_imax(&ctx->ac, src[0], src[1]);
1162 tmp2 = ac_build_imin(&ctx->ac, tmp2, src[2]);
1163 result = ac_build_imax(&ctx->ac, tmp1, tmp2);
1164 break;
1165 }
1166 case nir_op_umed3: {
1167 LLVMValueRef tmp1 = ac_build_umin(&ctx->ac, src[0], src[1]);
1168 LLVMValueRef tmp2 = ac_build_umax(&ctx->ac, src[0], src[1]);
1169 tmp2 = ac_build_umin(&ctx->ac, tmp2, src[2]);
1170 result = ac_build_umax(&ctx->ac, tmp1, tmp2);
1171 break;
1172 }
1173
1174 default:
1175 fprintf(stderr, "Unknown NIR alu instr: ");
1176 nir_print_instr(&instr->instr, stderr);
1177 fprintf(stderr, "\n");
1178 abort();
1179 }
1180
1181 if (result) {
1182 assert(instr->dest.dest.is_ssa);
1183 result = ac_to_integer_or_pointer(&ctx->ac, result);
1184 ctx->ssa_defs[instr->dest.dest.ssa.index] = result;
1185 }
1186 }
1187
1188 static void visit_load_const(struct ac_nir_context *ctx,
1189 const nir_load_const_instr *instr)
1190 {
1191 LLVMValueRef values[4], value = NULL;
1192 LLVMTypeRef element_type =
1193 LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
1194
1195 for (unsigned i = 0; i < instr->def.num_components; ++i) {
1196 switch (instr->def.bit_size) {
1197 case 8:
1198 values[i] = LLVMConstInt(element_type,
1199 instr->value[i].u8, false);
1200 break;
1201 case 16:
1202 values[i] = LLVMConstInt(element_type,
1203 instr->value[i].u16, false);
1204 break;
1205 case 32:
1206 values[i] = LLVMConstInt(element_type,
1207 instr->value[i].u32, false);
1208 break;
1209 case 64:
1210 values[i] = LLVMConstInt(element_type,
1211 instr->value[i].u64, false);
1212 break;
1213 default:
1214 fprintf(stderr,
1215 "unsupported nir load_const bit_size: %d\n",
1216 instr->def.bit_size);
1217 abort();
1218 }
1219 }
1220 if (instr->def.num_components > 1) {
1221 value = LLVMConstVector(values, instr->def.num_components);
1222 } else
1223 value = values[0];
1224
1225 ctx->ssa_defs[instr->def.index] = value;
1226 }
1227
1228 static LLVMValueRef
1229 get_buffer_size(struct ac_nir_context *ctx, LLVMValueRef descriptor, bool in_elements)
1230 {
1231 LLVMValueRef size =
1232 LLVMBuildExtractElement(ctx->ac.builder, descriptor,
1233 LLVMConstInt(ctx->ac.i32, 2, false), "");
1234
1235 /* GFX8 only */
1236 if (ctx->ac.chip_class == GFX8 && in_elements) {
1237 /* On GFX8, the descriptor contains the size in bytes,
1238 * but TXQ must return the size in elements.
1239 * The stride is always non-zero for resources using TXQ.
1240 */
1241 LLVMValueRef stride =
1242 LLVMBuildExtractElement(ctx->ac.builder, descriptor,
1243 ctx->ac.i32_1, "");
1244 stride = LLVMBuildLShr(ctx->ac.builder, stride,
1245 LLVMConstInt(ctx->ac.i32, 16, false), "");
1246 stride = LLVMBuildAnd(ctx->ac.builder, stride,
1247 LLVMConstInt(ctx->ac.i32, 0x3fff, false), "");
1248
1249 size = LLVMBuildUDiv(ctx->ac.builder, size, stride, "");
1250 }
1251 return size;
1252 }
1253
1254 /* Gather4 should follow the same rules as bilinear filtering, but the hardware
1255 * incorrectly forces nearest filtering if the texture format is integer.
1256 * The only effect it has on Gather4, which always returns 4 texels for
1257 * bilinear filtering, is that the final coordinates are off by 0.5 of
1258 * the texel size.
1259 *
1260 * The workaround is to subtract 0.5 from the unnormalized coordinates,
1261 * or (0.5 / size) from the normalized coordinates.
1262 *
1263 * However, cube textures with 8_8_8_8 data formats require a different
1264 * workaround of overriding the num format to USCALED/SSCALED. This would lose
1265 * precision in 32-bit data formats, so it needs to be applied dynamically at
1266 * runtime. In this case, return an i1 value that indicates whether the
1267 * descriptor was overridden (and hence a fixup of the sampler result is needed).
1268 */
1269 static LLVMValueRef lower_gather4_integer(struct ac_llvm_context *ctx,
1270 nir_variable *var,
1271 struct ac_image_args *args,
1272 const nir_tex_instr *instr)
1273 {
1274 const struct glsl_type *type = glsl_without_array(var->type);
1275 enum glsl_base_type stype = glsl_get_sampler_result_type(type);
1276 LLVMValueRef wa_8888 = NULL;
1277 LLVMValueRef half_texel[2];
1278 LLVMValueRef result;
1279
1280 assert(stype == GLSL_TYPE_INT || stype == GLSL_TYPE_UINT);
1281
1282 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
1283 LLVMValueRef formats;
1284 LLVMValueRef data_format;
1285 LLVMValueRef wa_formats;
1286
1287 formats = LLVMBuildExtractElement(ctx->builder, args->resource, ctx->i32_1, "");
1288
1289 data_format = LLVMBuildLShr(ctx->builder, formats,
1290 LLVMConstInt(ctx->i32, 20, false), "");
1291 data_format = LLVMBuildAnd(ctx->builder, data_format,
1292 LLVMConstInt(ctx->i32, (1u << 6) - 1, false), "");
1293 wa_8888 = LLVMBuildICmp(
1294 ctx->builder, LLVMIntEQ, data_format,
1295 LLVMConstInt(ctx->i32, V_008F14_IMG_DATA_FORMAT_8_8_8_8, false),
1296 "");
1297
1298 uint32_t wa_num_format =
1299 stype == GLSL_TYPE_UINT ?
1300 S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_USCALED) :
1301 S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_SSCALED);
1302 wa_formats = LLVMBuildAnd(ctx->builder, formats,
1303 LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT, false),
1304 "");
1305 wa_formats = LLVMBuildOr(ctx->builder, wa_formats,
1306 LLVMConstInt(ctx->i32, wa_num_format, false), "");
1307
1308 formats = LLVMBuildSelect(ctx->builder, wa_8888, wa_formats, formats, "");
1309 args->resource = LLVMBuildInsertElement(
1310 ctx->builder, args->resource, formats, ctx->i32_1, "");
1311 }
1312
1313 if (instr->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
1314 assert(!wa_8888);
1315 half_texel[0] = half_texel[1] = LLVMConstReal(ctx->f32, -0.5);
1316 } else {
1317 struct ac_image_args resinfo = {};
1318 LLVMBasicBlockRef bbs[2];
1319
1320 LLVMValueRef unnorm = NULL;
1321 LLVMValueRef default_offset = ctx->f32_0;
1322 if (instr->sampler_dim == GLSL_SAMPLER_DIM_2D &&
1323 !instr->is_array) {
1324 /* In vulkan, whether the sampler uses unnormalized
1325 * coordinates or not is a dynamic property of the
1326 * sampler. Hence, to figure out whether or not we
1327 * need to divide by the texture size, we need to test
1328 * the sampler at runtime. This tests the bit set by
1329 * radv_init_sampler().
1330 */
1331 LLVMValueRef sampler0 =
1332 LLVMBuildExtractElement(ctx->builder, args->sampler, ctx->i32_0, "");
1333 sampler0 = LLVMBuildLShr(ctx->builder, sampler0,
1334 LLVMConstInt(ctx->i32, 15, false), "");
1335 sampler0 = LLVMBuildAnd(ctx->builder, sampler0, ctx->i32_1, "");
1336 unnorm = LLVMBuildICmp(ctx->builder, LLVMIntEQ, sampler0, ctx->i32_1, "");
1337 default_offset = LLVMConstReal(ctx->f32, -0.5);
1338 }
1339
1340 bbs[0] = LLVMGetInsertBlock(ctx->builder);
1341 if (wa_8888 || unnorm) {
1342 assert(!(wa_8888 && unnorm));
1343 LLVMValueRef not_needed = wa_8888 ? wa_8888 : unnorm;
1344 /* Skip the texture size query entirely if we don't need it. */
1345 ac_build_ifcc(ctx, LLVMBuildNot(ctx->builder, not_needed, ""), 2000);
1346 bbs[1] = LLVMGetInsertBlock(ctx->builder);
1347 }
1348
1349 /* Query the texture size. */
1350 resinfo.dim = ac_get_sampler_dim(ctx->chip_class, instr->sampler_dim, instr->is_array);
1351 resinfo.opcode = ac_image_get_resinfo;
1352 resinfo.dmask = 0xf;
1353 resinfo.lod = ctx->i32_0;
1354 resinfo.resource = args->resource;
1355 resinfo.attributes = AC_FUNC_ATTR_READNONE;
1356 LLVMValueRef size = ac_build_image_opcode(ctx, &resinfo);
1357
1358 /* Compute -0.5 / size. */
1359 for (unsigned c = 0; c < 2; c++) {
1360 half_texel[c] =
1361 LLVMBuildExtractElement(ctx->builder, size,
1362 LLVMConstInt(ctx->i32, c, 0), "");
1363 half_texel[c] = LLVMBuildUIToFP(ctx->builder, half_texel[c], ctx->f32, "");
1364 half_texel[c] = ac_build_fdiv(ctx, ctx->f32_1, half_texel[c]);
1365 half_texel[c] = LLVMBuildFMul(ctx->builder, half_texel[c],
1366 LLVMConstReal(ctx->f32, -0.5), "");
1367 }
1368
1369 if (wa_8888 || unnorm) {
1370 ac_build_endif(ctx, 2000);
1371
1372 for (unsigned c = 0; c < 2; c++) {
1373 LLVMValueRef values[2] = { default_offset, half_texel[c] };
1374 half_texel[c] = ac_build_phi(ctx, ctx->f32, 2,
1375 values, bbs);
1376 }
1377 }
1378 }
1379
1380 for (unsigned c = 0; c < 2; c++) {
1381 LLVMValueRef tmp;
1382 tmp = LLVMBuildBitCast(ctx->builder, args->coords[c], ctx->f32, "");
1383 args->coords[c] = LLVMBuildFAdd(ctx->builder, tmp, half_texel[c], "");
1384 }
1385
1386 args->attributes = AC_FUNC_ATTR_READNONE;
1387 result = ac_build_image_opcode(ctx, args);
1388
1389 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
1390 LLVMValueRef tmp, tmp2;
1391
1392 /* if the cube workaround is in place, f2i the result. */
1393 for (unsigned c = 0; c < 4; c++) {
1394 tmp = LLVMBuildExtractElement(ctx->builder, result, LLVMConstInt(ctx->i32, c, false), "");
1395 if (stype == GLSL_TYPE_UINT)
1396 tmp2 = LLVMBuildFPToUI(ctx->builder, tmp, ctx->i32, "");
1397 else
1398 tmp2 = LLVMBuildFPToSI(ctx->builder, tmp, ctx->i32, "");
1399 tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->i32, "");
1400 tmp2 = LLVMBuildBitCast(ctx->builder, tmp2, ctx->i32, "");
1401 tmp = LLVMBuildSelect(ctx->builder, wa_8888, tmp2, tmp, "");
1402 tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->f32, "");
1403 result = LLVMBuildInsertElement(ctx->builder, result, tmp, LLVMConstInt(ctx->i32, c, false), "");
1404 }
1405 }
1406 return result;
1407 }
1408
1409 static nir_deref_instr *get_tex_texture_deref(const nir_tex_instr *instr)
1410 {
1411 nir_deref_instr *texture_deref_instr = NULL;
1412
1413 for (unsigned i = 0; i < instr->num_srcs; i++) {
1414 switch (instr->src[i].src_type) {
1415 case nir_tex_src_texture_deref:
1416 texture_deref_instr = nir_src_as_deref(instr->src[i].src);
1417 break;
1418 default:
1419 break;
1420 }
1421 }
1422 return texture_deref_instr;
1423 }
1424
1425 static LLVMValueRef build_tex_intrinsic(struct ac_nir_context *ctx,
1426 const nir_tex_instr *instr,
1427 struct ac_image_args *args)
1428 {
1429 if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
1430 unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
1431
1432 return ac_build_buffer_load_format(&ctx->ac,
1433 args->resource,
1434 args->coords[0],
1435 ctx->ac.i32_0,
1436 util_last_bit(mask),
1437 0, true);
1438 }
1439
1440 args->opcode = ac_image_sample;
1441
1442 switch (instr->op) {
1443 case nir_texop_txf:
1444 case nir_texop_txf_ms:
1445 case nir_texop_samples_identical:
1446 args->opcode = args->level_zero ||
1447 instr->sampler_dim == GLSL_SAMPLER_DIM_MS ?
1448 ac_image_load : ac_image_load_mip;
1449 args->level_zero = false;
1450 break;
1451 case nir_texop_txs:
1452 case nir_texop_query_levels:
1453 args->opcode = ac_image_get_resinfo;
1454 if (!args->lod)
1455 args->lod = ctx->ac.i32_0;
1456 args->level_zero = false;
1457 break;
1458 case nir_texop_tex:
1459 if (ctx->stage != MESA_SHADER_FRAGMENT) {
1460 assert(!args->lod);
1461 args->level_zero = true;
1462 }
1463 break;
1464 case nir_texop_tg4:
1465 args->opcode = ac_image_gather4;
1466 args->level_zero = true;
1467 break;
1468 case nir_texop_lod:
1469 args->opcode = ac_image_get_lod;
1470 break;
1471 case nir_texop_fragment_fetch:
1472 case nir_texop_fragment_mask_fetch:
1473 args->opcode = ac_image_load;
1474 args->level_zero = false;
1475 break;
1476 default:
1477 break;
1478 }
1479
1480 if (instr->op == nir_texop_tg4 && ctx->ac.chip_class <= GFX8) {
1481 nir_deref_instr *texture_deref_instr = get_tex_texture_deref(instr);
1482 nir_variable *var = nir_deref_instr_get_variable(texture_deref_instr);
1483 const struct glsl_type *type = glsl_without_array(var->type);
1484 enum glsl_base_type stype = glsl_get_sampler_result_type(type);
1485 if (stype == GLSL_TYPE_UINT || stype == GLSL_TYPE_INT) {
1486 return lower_gather4_integer(&ctx->ac, var, args, instr);
1487 }
1488 }
1489
1490 /* Fixup for GFX9 which allocates 1D textures as 2D. */
1491 if (instr->op == nir_texop_lod && ctx->ac.chip_class == GFX9) {
1492 if ((args->dim == ac_image_2darray ||
1493 args->dim == ac_image_2d) && !args->coords[1]) {
1494 args->coords[1] = ctx->ac.i32_0;
1495 }
1496 }
1497
1498 args->attributes = AC_FUNC_ATTR_READNONE;
1499 bool cs_derivs = ctx->stage == MESA_SHADER_COMPUTE &&
1500 ctx->info->cs.derivative_group != DERIVATIVE_GROUP_NONE;
1501 if (ctx->stage == MESA_SHADER_FRAGMENT || cs_derivs) {
1502 /* Prevent texture instructions with implicit derivatives from being
1503 * sinked into branches. */
1504 switch (instr->op) {
1505 case nir_texop_tex:
1506 case nir_texop_txb:
1507 case nir_texop_lod:
1508 args->attributes |= AC_FUNC_ATTR_CONVERGENT;
1509 break;
1510 default:
1511 break;
1512 }
1513 }
1514
1515 return ac_build_image_opcode(&ctx->ac, args);
1516 }
1517
1518 static LLVMValueRef visit_vulkan_resource_reindex(struct ac_nir_context *ctx,
1519 nir_intrinsic_instr *instr)
1520 {
1521 LLVMValueRef ptr = get_src(ctx, instr->src[0]);
1522 LLVMValueRef index = get_src(ctx, instr->src[1]);
1523
1524 LLVMValueRef result = LLVMBuildGEP(ctx->ac.builder, ptr, &index, 1, "");
1525 LLVMSetMetadata(result, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
1526 return result;
1527 }
1528
1529 static LLVMValueRef visit_load_push_constant(struct ac_nir_context *ctx,
1530 nir_intrinsic_instr *instr)
1531 {
1532 LLVMValueRef ptr, addr;
1533 LLVMValueRef src0 = get_src(ctx, instr->src[0]);
1534 unsigned index = nir_intrinsic_base(instr);
1535
1536 addr = LLVMConstInt(ctx->ac.i32, index, 0);
1537 addr = LLVMBuildAdd(ctx->ac.builder, addr, src0, "");
1538
1539 /* Load constant values from user SGPRS when possible, otherwise
1540 * fallback to the default path that loads directly from memory.
1541 */
1542 if (LLVMIsConstant(src0) &&
1543 instr->dest.ssa.bit_size == 32) {
1544 unsigned count = instr->dest.ssa.num_components;
1545 unsigned offset = index;
1546
1547 offset += LLVMConstIntGetZExtValue(src0);
1548 offset /= 4;
1549
1550 offset -= ctx->args->base_inline_push_consts;
1551
1552 unsigned num_inline_push_consts = ctx->args->num_inline_push_consts;
1553 if (offset + count <= num_inline_push_consts) {
1554 LLVMValueRef push_constants[num_inline_push_consts];
1555 for (unsigned i = 0; i < num_inline_push_consts; i++)
1556 push_constants[i] = ac_get_arg(&ctx->ac,
1557 ctx->args->inline_push_consts[i]);
1558 return ac_build_gather_values(&ctx->ac,
1559 push_constants + offset,
1560 count);
1561 }
1562 }
1563
1564 ptr = LLVMBuildGEP(ctx->ac.builder,
1565 ac_get_arg(&ctx->ac, ctx->args->push_constants), &addr, 1, "");
1566
1567 if (instr->dest.ssa.bit_size == 8) {
1568 unsigned load_dwords = instr->dest.ssa.num_components > 1 ? 2 : 1;
1569 LLVMTypeRef vec_type = LLVMVectorType(LLVMInt8TypeInContext(ctx->ac.context), 4 * load_dwords);
1570 ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
1571 LLVMValueRef res = LLVMBuildLoad(ctx->ac.builder, ptr, "");
1572
1573 LLVMValueRef params[3];
1574 if (load_dwords > 1) {
1575 LLVMValueRef res_vec = LLVMBuildBitCast(ctx->ac.builder, res, LLVMVectorType(ctx->ac.i32, 2), "");
1576 params[0] = LLVMBuildExtractElement(ctx->ac.builder, res_vec, LLVMConstInt(ctx->ac.i32, 1, false), "");
1577 params[1] = LLVMBuildExtractElement(ctx->ac.builder, res_vec, LLVMConstInt(ctx->ac.i32, 0, false), "");
1578 } else {
1579 res = LLVMBuildBitCast(ctx->ac.builder, res, ctx->ac.i32, "");
1580 params[0] = ctx->ac.i32_0;
1581 params[1] = res;
1582 }
1583 params[2] = addr;
1584 res = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.alignbyte", ctx->ac.i32, params, 3, 0);
1585
1586 res = LLVMBuildTrunc(ctx->ac.builder, res, LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.num_components * 8), "");
1587 if (instr->dest.ssa.num_components > 1)
1588 res = LLVMBuildBitCast(ctx->ac.builder, res, LLVMVectorType(LLVMInt8TypeInContext(ctx->ac.context), instr->dest.ssa.num_components), "");
1589 return res;
1590 } else if (instr->dest.ssa.bit_size == 16) {
1591 unsigned load_dwords = instr->dest.ssa.num_components / 2 + 1;
1592 LLVMTypeRef vec_type = LLVMVectorType(LLVMInt16TypeInContext(ctx->ac.context), 2 * load_dwords);
1593 ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
1594 LLVMValueRef res = LLVMBuildLoad(ctx->ac.builder, ptr, "");
1595 res = LLVMBuildBitCast(ctx->ac.builder, res, vec_type, "");
1596 LLVMValueRef cond = LLVMBuildLShr(ctx->ac.builder, addr, ctx->ac.i32_1, "");
1597 cond = LLVMBuildTrunc(ctx->ac.builder, cond, ctx->ac.i1, "");
1598 LLVMValueRef mask[] = { LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
1599 LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
1600 LLVMConstInt(ctx->ac.i32, 4, false)};
1601 LLVMValueRef swizzle_aligned = LLVMConstVector(&mask[0], instr->dest.ssa.num_components);
1602 LLVMValueRef swizzle_unaligned = LLVMConstVector(&mask[1], instr->dest.ssa.num_components);
1603 LLVMValueRef shuffle_aligned = LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_aligned, "");
1604 LLVMValueRef shuffle_unaligned = LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_unaligned, "");
1605 res = LLVMBuildSelect(ctx->ac.builder, cond, shuffle_unaligned, shuffle_aligned, "");
1606 return LLVMBuildBitCast(ctx->ac.builder, res, get_def_type(ctx, &instr->dest.ssa), "");
1607 }
1608
1609 ptr = ac_cast_ptr(&ctx->ac, ptr, get_def_type(ctx, &instr->dest.ssa));
1610
1611 return LLVMBuildLoad(ctx->ac.builder, ptr, "");
1612 }
1613
1614 static LLVMValueRef visit_get_buffer_size(struct ac_nir_context *ctx,
1615 const nir_intrinsic_instr *instr)
1616 {
1617 LLVMValueRef index = get_src(ctx, instr->src[0]);
1618
1619 return get_buffer_size(ctx, ctx->abi->load_ssbo(ctx->abi, index, false), false);
1620 }
1621
1622 static uint32_t widen_mask(uint32_t mask, unsigned multiplier)
1623 {
1624 uint32_t new_mask = 0;
1625 for(unsigned i = 0; i < 32 && (1u << i) <= mask; ++i)
1626 if (mask & (1u << i))
1627 new_mask |= ((1u << multiplier) - 1u) << (i * multiplier);
1628 return new_mask;
1629 }
1630
1631 static LLVMValueRef extract_vector_range(struct ac_llvm_context *ctx, LLVMValueRef src,
1632 unsigned start, unsigned count)
1633 {
1634 LLVMValueRef mask[] = {
1635 ctx->i32_0, ctx->i32_1,
1636 LLVMConstInt(ctx->i32, 2, false), LLVMConstInt(ctx->i32, 3, false) };
1637
1638 unsigned src_elements = ac_get_llvm_num_components(src);
1639
1640 if (count == src_elements) {
1641 assert(start == 0);
1642 return src;
1643 } else if (count == 1) {
1644 assert(start < src_elements);
1645 return LLVMBuildExtractElement(ctx->builder, src, mask[start], "");
1646 } else {
1647 assert(start + count <= src_elements);
1648 assert(count <= 4);
1649 LLVMValueRef swizzle = LLVMConstVector(&mask[start], count);
1650 return LLVMBuildShuffleVector(ctx->builder, src, src, swizzle, "");
1651 }
1652 }
1653
1654 static unsigned get_cache_policy(struct ac_nir_context *ctx,
1655 enum gl_access_qualifier access,
1656 bool may_store_unaligned,
1657 bool writeonly_memory)
1658 {
1659 unsigned cache_policy = 0;
1660
1661 /* GFX6 has a TC L1 bug causing corruption of 8bit/16bit stores. All
1662 * store opcodes not aligned to a dword are affected. The only way to
1663 * get unaligned stores is through shader images.
1664 */
1665 if (((may_store_unaligned && ctx->ac.chip_class == GFX6) ||
1666 /* If this is write-only, don't keep data in L1 to prevent
1667 * evicting L1 cache lines that may be needed by other
1668 * instructions.
1669 */
1670 writeonly_memory ||
1671 access & (ACCESS_COHERENT | ACCESS_VOLATILE))) {
1672 cache_policy |= ac_glc;
1673 }
1674
1675 if (access & ACCESS_STREAM_CACHE_POLICY)
1676 cache_policy |= ac_slc;
1677
1678 return cache_policy;
1679 }
1680
1681 static LLVMValueRef enter_waterfall_ssbo(struct ac_nir_context *ctx,
1682 struct waterfall_context *wctx,
1683 const nir_intrinsic_instr *instr,
1684 nir_src src)
1685 {
1686 return enter_waterfall(ctx, wctx, get_src(ctx, src),
1687 nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
1688 }
1689
1690 static void visit_store_ssbo(struct ac_nir_context *ctx,
1691 nir_intrinsic_instr *instr)
1692 {
1693 if (ctx->ac.postponed_kill) {
1694 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
1695 ctx->ac.postponed_kill, "");
1696 ac_build_ifcc(&ctx->ac, cond, 7000);
1697 }
1698
1699 LLVMValueRef src_data = get_src(ctx, instr->src[0]);
1700 int elem_size_bytes = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src_data)) / 8;
1701 unsigned writemask = nir_intrinsic_write_mask(instr);
1702 enum gl_access_qualifier access = nir_intrinsic_access(instr);
1703 bool writeonly_memory = access & ACCESS_NON_READABLE;
1704 unsigned cache_policy = get_cache_policy(ctx, access, false, writeonly_memory);
1705
1706 struct waterfall_context wctx;
1707 LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[1]);
1708
1709 LLVMValueRef rsrc = ctx->abi->load_ssbo(ctx->abi, rsrc_base, true);
1710 LLVMValueRef base_data = src_data;
1711 base_data = ac_trim_vector(&ctx->ac, base_data, instr->num_components);
1712 LLVMValueRef base_offset = get_src(ctx, instr->src[2]);
1713
1714 while (writemask) {
1715 int start, count;
1716 LLVMValueRef data, offset;
1717 LLVMTypeRef data_type;
1718
1719 u_bit_scan_consecutive_range(&writemask, &start, &count);
1720
1721 /* Due to an LLVM limitation with LLVM < 9, split 3-element
1722 * writes into a 2-element and a 1-element write. */
1723 if (count == 3 &&
1724 (elem_size_bytes != 4 || !ac_has_vec3_support(ctx->ac.chip_class, false))) {
1725 writemask |= 1 << (start + 2);
1726 count = 2;
1727 }
1728 int num_bytes = count * elem_size_bytes; /* count in bytes */
1729
1730 /* we can only store 4 DWords at the same time.
1731 * can only happen for 64 Bit vectors. */
1732 if (num_bytes > 16) {
1733 writemask |= ((1u << (count - 2)) - 1u) << (start + 2);
1734 count = 2;
1735 num_bytes = 16;
1736 }
1737
1738 /* check alignment of 16 Bit stores */
1739 if (elem_size_bytes == 2 && num_bytes > 2 && (start % 2) == 1) {
1740 writemask |= ((1u << (count - 1)) - 1u) << (start + 1);
1741 count = 1;
1742 num_bytes = 2;
1743 }
1744 data = extract_vector_range(&ctx->ac, base_data, start, count);
1745
1746 offset = LLVMBuildAdd(ctx->ac.builder, base_offset,
1747 LLVMConstInt(ctx->ac.i32, start * elem_size_bytes, false), "");
1748
1749 if (num_bytes == 1) {
1750 ac_build_tbuffer_store_byte(&ctx->ac, rsrc, data,
1751 offset, ctx->ac.i32_0,
1752 cache_policy);
1753 } else if (num_bytes == 2) {
1754 ac_build_tbuffer_store_short(&ctx->ac, rsrc, data,
1755 offset, ctx->ac.i32_0,
1756 cache_policy);
1757 } else {
1758 int num_channels = num_bytes / 4;
1759
1760 switch (num_bytes) {
1761 case 16: /* v4f32 */
1762 data_type = ctx->ac.v4f32;
1763 break;
1764 case 12: /* v3f32 */
1765 data_type = ctx->ac.v3f32;
1766 break;
1767 case 8: /* v2f32 */
1768 data_type = ctx->ac.v2f32;
1769 break;
1770 case 4: /* f32 */
1771 data_type = ctx->ac.f32;
1772 break;
1773 default:
1774 unreachable("Malformed vector store.");
1775 }
1776 data = LLVMBuildBitCast(ctx->ac.builder, data, data_type, "");
1777
1778 ac_build_buffer_store_dword(&ctx->ac, rsrc, data,
1779 num_channels, offset,
1780 ctx->ac.i32_0, 0,
1781 cache_policy);
1782 }
1783 }
1784
1785 exit_waterfall(ctx, &wctx, NULL);
1786
1787 if (ctx->ac.postponed_kill)
1788 ac_build_endif(&ctx->ac, 7000);
1789 }
1790
1791 static LLVMValueRef emit_ssbo_comp_swap_64(struct ac_nir_context *ctx,
1792 LLVMValueRef descriptor,
1793 LLVMValueRef offset,
1794 LLVMValueRef compare,
1795 LLVMValueRef exchange)
1796 {
1797 LLVMBasicBlockRef start_block = NULL, then_block = NULL;
1798 if (ctx->abi->robust_buffer_access) {
1799 LLVMValueRef size = ac_llvm_extract_elem(&ctx->ac, descriptor, 2);
1800
1801 LLVMValueRef cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, offset, size, "");
1802 start_block = LLVMGetInsertBlock(ctx->ac.builder);
1803
1804 ac_build_ifcc(&ctx->ac, cond, -1);
1805
1806 then_block = LLVMGetInsertBlock(ctx->ac.builder);
1807 }
1808
1809 LLVMValueRef ptr_parts[2] = {
1810 ac_llvm_extract_elem(&ctx->ac, descriptor, 0),
1811 LLVMBuildAnd(ctx->ac.builder,
1812 ac_llvm_extract_elem(&ctx->ac, descriptor, 1),
1813 LLVMConstInt(ctx->ac.i32, 65535, 0), "")
1814 };
1815
1816 ptr_parts[1] = LLVMBuildTrunc(ctx->ac.builder, ptr_parts[1], ctx->ac.i16, "");
1817 ptr_parts[1] = LLVMBuildSExt(ctx->ac.builder, ptr_parts[1], ctx->ac.i32, "");
1818
1819 offset = LLVMBuildZExt(ctx->ac.builder, offset, ctx->ac.i64, "");
1820
1821 LLVMValueRef ptr = ac_build_gather_values(&ctx->ac, ptr_parts, 2);
1822 ptr = LLVMBuildBitCast(ctx->ac.builder, ptr, ctx->ac.i64, "");
1823 ptr = LLVMBuildAdd(ctx->ac.builder, ptr, offset, "");
1824 ptr = LLVMBuildIntToPtr(ctx->ac.builder, ptr, LLVMPointerType(ctx->ac.i64, AC_ADDR_SPACE_GLOBAL), "");
1825
1826 LLVMValueRef result = ac_build_atomic_cmp_xchg(&ctx->ac, ptr, compare, exchange, "singlethread-one-as");
1827 result = LLVMBuildExtractValue(ctx->ac.builder, result, 0, "");
1828
1829 if (ctx->abi->robust_buffer_access) {
1830 ac_build_endif(&ctx->ac, -1);
1831
1832 LLVMBasicBlockRef incoming_blocks[2] = {
1833 start_block,
1834 then_block,
1835 };
1836
1837 LLVMValueRef incoming_values[2] = {
1838 LLVMConstInt(ctx->ac.i64, 0, 0),
1839 result,
1840 };
1841 LLVMValueRef ret = LLVMBuildPhi(ctx->ac.builder, ctx->ac.i64, "");
1842 LLVMAddIncoming(ret, incoming_values, incoming_blocks, 2);
1843 return ret;
1844 } else {
1845 return result;
1846 }
1847 }
1848
1849 static LLVMValueRef visit_atomic_ssbo(struct ac_nir_context *ctx,
1850 nir_intrinsic_instr *instr)
1851 {
1852 if (ctx->ac.postponed_kill) {
1853 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
1854 ctx->ac.postponed_kill, "");
1855 ac_build_ifcc(&ctx->ac, cond, 7001);
1856 }
1857
1858 LLVMTypeRef return_type = LLVMTypeOf(get_src(ctx, instr->src[2]));
1859 const char *op;
1860 char name[64], type[8];
1861 LLVMValueRef params[6], descriptor;
1862 LLVMValueRef result;
1863 int arg_count = 0;
1864
1865 struct waterfall_context wctx;
1866 LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[0]);
1867
1868 switch (instr->intrinsic) {
1869 case nir_intrinsic_ssbo_atomic_add:
1870 op = "add";
1871 break;
1872 case nir_intrinsic_ssbo_atomic_imin:
1873 op = "smin";
1874 break;
1875 case nir_intrinsic_ssbo_atomic_umin:
1876 op = "umin";
1877 break;
1878 case nir_intrinsic_ssbo_atomic_imax:
1879 op = "smax";
1880 break;
1881 case nir_intrinsic_ssbo_atomic_umax:
1882 op = "umax";
1883 break;
1884 case nir_intrinsic_ssbo_atomic_and:
1885 op = "and";
1886 break;
1887 case nir_intrinsic_ssbo_atomic_or:
1888 op = "or";
1889 break;
1890 case nir_intrinsic_ssbo_atomic_xor:
1891 op = "xor";
1892 break;
1893 case nir_intrinsic_ssbo_atomic_exchange:
1894 op = "swap";
1895 break;
1896 case nir_intrinsic_ssbo_atomic_comp_swap:
1897 op = "cmpswap";
1898 break;
1899 default:
1900 abort();
1901 }
1902
1903 descriptor = ctx->abi->load_ssbo(ctx->abi,
1904 rsrc_base,
1905 true);
1906
1907 if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap &&
1908 return_type == ctx->ac.i64) {
1909 result = emit_ssbo_comp_swap_64(ctx, descriptor,
1910 get_src(ctx, instr->src[1]),
1911 get_src(ctx, instr->src[2]),
1912 get_src(ctx, instr->src[3]));
1913 } else {
1914 if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap) {
1915 params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[3]), 0);
1916 }
1917 params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
1918 params[arg_count++] = descriptor;
1919
1920 if (LLVM_VERSION_MAJOR >= 9) {
1921 /* XXX: The new raw/struct atomic intrinsics are buggy with
1922 * LLVM 8, see r358579.
1923 */
1924 params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
1925 params[arg_count++] = ctx->ac.i32_0; /* soffset */
1926 params[arg_count++] = ctx->ac.i32_0; /* slc */
1927
1928 ac_build_type_name_for_intr(return_type, type, sizeof(type));
1929 snprintf(name, sizeof(name),
1930 "llvm.amdgcn.raw.buffer.atomic.%s.%s", op, type);
1931 } else {
1932 params[arg_count++] = ctx->ac.i32_0; /* vindex */
1933 params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
1934 params[arg_count++] = ctx->ac.i1false; /* slc */
1935
1936 assert(return_type == ctx->ac.i32);
1937 snprintf(name, sizeof(name),
1938 "llvm.amdgcn.buffer.atomic.%s", op);
1939 }
1940
1941 result = ac_build_intrinsic(&ctx->ac, name, return_type, params,
1942 arg_count, 0);
1943 }
1944
1945 result = exit_waterfall(ctx, &wctx, result);
1946 if (ctx->ac.postponed_kill)
1947 ac_build_endif(&ctx->ac, 7001);
1948 return result;
1949 }
1950
1951 static LLVMValueRef visit_load_buffer(struct ac_nir_context *ctx,
1952 nir_intrinsic_instr *instr)
1953 {
1954 struct waterfall_context wctx;
1955 LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[0]);
1956
1957 int elem_size_bytes = instr->dest.ssa.bit_size / 8;
1958 int num_components = instr->num_components;
1959 enum gl_access_qualifier access = nir_intrinsic_access(instr);
1960 unsigned cache_policy = get_cache_policy(ctx, access, false, false);
1961
1962 LLVMValueRef offset = get_src(ctx, instr->src[1]);
1963 LLVMValueRef rsrc = ctx->abi->load_ssbo(ctx->abi, rsrc_base, false);
1964 LLVMValueRef vindex = ctx->ac.i32_0;
1965
1966 LLVMTypeRef def_type = get_def_type(ctx, &instr->dest.ssa);
1967 LLVMTypeRef def_elem_type = num_components > 1 ? LLVMGetElementType(def_type) : def_type;
1968
1969 LLVMValueRef results[4];
1970 for (int i = 0; i < num_components;) {
1971 int num_elems = num_components - i;
1972 if (elem_size_bytes < 4 && nir_intrinsic_align(instr) % 4 != 0)
1973 num_elems = 1;
1974 if (num_elems * elem_size_bytes > 16)
1975 num_elems = 16 / elem_size_bytes;
1976 int load_bytes = num_elems * elem_size_bytes;
1977
1978 LLVMValueRef immoffset = LLVMConstInt(ctx->ac.i32, i * elem_size_bytes, false);
1979
1980 LLVMValueRef ret;
1981
1982 if (load_bytes == 1) {
1983 ret = ac_build_tbuffer_load_byte(&ctx->ac,
1984 rsrc,
1985 offset,
1986 ctx->ac.i32_0,
1987 immoffset,
1988 cache_policy);
1989 } else if (load_bytes == 2) {
1990 ret = ac_build_tbuffer_load_short(&ctx->ac,
1991 rsrc,
1992 offset,
1993 ctx->ac.i32_0,
1994 immoffset,
1995 cache_policy);
1996 } else {
1997 int num_channels = util_next_power_of_two(load_bytes) / 4;
1998 bool can_speculate = access & ACCESS_CAN_REORDER;
1999
2000 ret = ac_build_buffer_load(&ctx->ac, rsrc, num_channels,
2001 vindex, offset, immoffset, 0,
2002 cache_policy, can_speculate, false);
2003 }
2004
2005 LLVMTypeRef byte_vec = LLVMVectorType(ctx->ac.i8, ac_get_type_size(LLVMTypeOf(ret)));
2006 ret = LLVMBuildBitCast(ctx->ac.builder, ret, byte_vec, "");
2007 ret = ac_trim_vector(&ctx->ac, ret, load_bytes);
2008
2009 LLVMTypeRef ret_type = LLVMVectorType(def_elem_type, num_elems);
2010 ret = LLVMBuildBitCast(ctx->ac.builder, ret, ret_type, "");
2011
2012 for (unsigned j = 0; j < num_elems; j++) {
2013 results[i + j] = LLVMBuildExtractElement(ctx->ac.builder, ret, LLVMConstInt(ctx->ac.i32, j, false), "");
2014 }
2015 i += num_elems;
2016 }
2017
2018 LLVMValueRef ret = ac_build_gather_values(&ctx->ac, results, num_components);
2019 return exit_waterfall(ctx, &wctx, ret);
2020 }
2021
2022 static LLVMValueRef enter_waterfall_ubo(struct ac_nir_context *ctx,
2023 struct waterfall_context *wctx,
2024 const nir_intrinsic_instr *instr)
2025 {
2026 return enter_waterfall(ctx, wctx, get_src(ctx, instr->src[0]),
2027 nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
2028 }
2029
2030 static LLVMValueRef visit_load_ubo_buffer(struct ac_nir_context *ctx,
2031 nir_intrinsic_instr *instr)
2032 {
2033 struct waterfall_context wctx;
2034 LLVMValueRef rsrc_base = enter_waterfall_ubo(ctx, &wctx, instr);
2035
2036 LLVMValueRef ret;
2037 LLVMValueRef rsrc = rsrc_base;
2038 LLVMValueRef offset = get_src(ctx, instr->src[1]);
2039 int num_components = instr->num_components;
2040
2041 if (ctx->abi->load_ubo)
2042 rsrc = ctx->abi->load_ubo(ctx->abi, rsrc);
2043
2044 if (instr->dest.ssa.bit_size == 64)
2045 num_components *= 2;
2046
2047 if (instr->dest.ssa.bit_size == 16 || instr->dest.ssa.bit_size == 8) {
2048 unsigned load_bytes = instr->dest.ssa.bit_size / 8;
2049 LLVMValueRef results[num_components];
2050 for (unsigned i = 0; i < num_components; ++i) {
2051 LLVMValueRef immoffset = LLVMConstInt(ctx->ac.i32,
2052 load_bytes * i, 0);
2053
2054 if (load_bytes == 1) {
2055 results[i] = ac_build_tbuffer_load_byte(&ctx->ac,
2056 rsrc,
2057 offset,
2058 ctx->ac.i32_0,
2059 immoffset,
2060 0);
2061 } else {
2062 assert(load_bytes == 2);
2063 results[i] = ac_build_tbuffer_load_short(&ctx->ac,
2064 rsrc,
2065 offset,
2066 ctx->ac.i32_0,
2067 immoffset,
2068 0);
2069 }
2070 }
2071 ret = ac_build_gather_values(&ctx->ac, results, num_components);
2072 } else {
2073 ret = ac_build_buffer_load(&ctx->ac, rsrc, num_components, NULL, offset,
2074 NULL, 0, 0, true, true);
2075
2076 ret = ac_trim_vector(&ctx->ac, ret, num_components);
2077 }
2078
2079 ret = LLVMBuildBitCast(ctx->ac.builder, ret,
2080 get_def_type(ctx, &instr->dest.ssa), "");
2081
2082 return exit_waterfall(ctx, &wctx, ret);
2083 }
2084
2085 static void
2086 get_deref_offset(struct ac_nir_context *ctx, nir_deref_instr *instr,
2087 bool vs_in, unsigned *vertex_index_out,
2088 LLVMValueRef *vertex_index_ref,
2089 unsigned *const_out, LLVMValueRef *indir_out)
2090 {
2091 nir_variable *var = nir_deref_instr_get_variable(instr);
2092 nir_deref_path path;
2093 unsigned idx_lvl = 1;
2094
2095 nir_deref_path_init(&path, instr, NULL);
2096
2097 if (vertex_index_out != NULL || vertex_index_ref != NULL) {
2098 if (vertex_index_ref) {
2099 *vertex_index_ref = get_src(ctx, path.path[idx_lvl]->arr.index);
2100 if (vertex_index_out)
2101 *vertex_index_out = 0;
2102 } else {
2103 *vertex_index_out = nir_src_as_uint(path.path[idx_lvl]->arr.index);
2104 }
2105 ++idx_lvl;
2106 }
2107
2108 uint32_t const_offset = 0;
2109 LLVMValueRef offset = NULL;
2110
2111 if (var->data.compact) {
2112 assert(instr->deref_type == nir_deref_type_array);
2113 const_offset = nir_src_as_uint(instr->arr.index);
2114 goto out;
2115 }
2116
2117 for (; path.path[idx_lvl]; ++idx_lvl) {
2118 const struct glsl_type *parent_type = path.path[idx_lvl - 1]->type;
2119 if (path.path[idx_lvl]->deref_type == nir_deref_type_struct) {
2120 unsigned index = path.path[idx_lvl]->strct.index;
2121
2122 for (unsigned i = 0; i < index; i++) {
2123 const struct glsl_type *ft = glsl_get_struct_field(parent_type, i);
2124 const_offset += glsl_count_attribute_slots(ft, vs_in);
2125 }
2126 } else if(path.path[idx_lvl]->deref_type == nir_deref_type_array) {
2127 unsigned size = glsl_count_attribute_slots(path.path[idx_lvl]->type, vs_in);
2128 if (nir_src_is_const(path.path[idx_lvl]->arr.index)) {
2129 const_offset += size *
2130 nir_src_as_uint(path.path[idx_lvl]->arr.index);
2131 } else {
2132 LLVMValueRef array_off = LLVMBuildMul(ctx->ac.builder, LLVMConstInt(ctx->ac.i32, size, 0),
2133 get_src(ctx, path.path[idx_lvl]->arr.index), "");
2134 if (offset)
2135 offset = LLVMBuildAdd(ctx->ac.builder, offset, array_off, "");
2136 else
2137 offset = array_off;
2138 }
2139 } else
2140 unreachable("Uhandled deref type in get_deref_instr_offset");
2141 }
2142
2143 out:
2144 nir_deref_path_finish(&path);
2145
2146 if (const_offset && offset)
2147 offset = LLVMBuildAdd(ctx->ac.builder, offset,
2148 LLVMConstInt(ctx->ac.i32, const_offset, 0),
2149 "");
2150
2151 *const_out = const_offset;
2152 *indir_out = offset;
2153 }
2154
2155 static LLVMValueRef load_tess_varyings(struct ac_nir_context *ctx,
2156 nir_intrinsic_instr *instr,
2157 bool load_inputs)
2158 {
2159 LLVMValueRef result;
2160 LLVMValueRef vertex_index = NULL;
2161 LLVMValueRef indir_index = NULL;
2162 unsigned const_index = 0;
2163
2164 nir_variable *var = nir_deref_instr_get_variable(nir_instr_as_deref(instr->src[0].ssa->parent_instr));
2165
2166 unsigned location = var->data.location;
2167 unsigned driver_location = var->data.driver_location;
2168 const bool is_patch = var->data.patch ||
2169 var->data.location == VARYING_SLOT_TESS_LEVEL_INNER ||
2170 var->data.location == VARYING_SLOT_TESS_LEVEL_OUTER;
2171 const bool is_compact = var->data.compact;
2172
2173 get_deref_offset(ctx, nir_instr_as_deref(instr->src[0].ssa->parent_instr),
2174 false, NULL, is_patch ? NULL : &vertex_index,
2175 &const_index, &indir_index);
2176
2177 LLVMTypeRef dest_type = get_def_type(ctx, &instr->dest.ssa);
2178
2179 LLVMTypeRef src_component_type;
2180 if (LLVMGetTypeKind(dest_type) == LLVMVectorTypeKind)
2181 src_component_type = LLVMGetElementType(dest_type);
2182 else
2183 src_component_type = dest_type;
2184
2185 result = ctx->abi->load_tess_varyings(ctx->abi, src_component_type,
2186 vertex_index, indir_index,
2187 const_index, location, driver_location,
2188 var->data.location_frac,
2189 instr->num_components,
2190 is_patch, is_compact, load_inputs);
2191 if (instr->dest.ssa.bit_size == 16) {
2192 result = ac_to_integer(&ctx->ac, result);
2193 result = LLVMBuildTrunc(ctx->ac.builder, result, dest_type, "");
2194 }
2195 return LLVMBuildBitCast(ctx->ac.builder, result, dest_type, "");
2196 }
2197
2198 static unsigned
2199 type_scalar_size_bytes(const struct glsl_type *type)
2200 {
2201 assert(glsl_type_is_vector_or_scalar(type) ||
2202 glsl_type_is_matrix(type));
2203 return glsl_type_is_boolean(type) ? 4 : glsl_get_bit_size(type) / 8;
2204 }
2205
2206 static LLVMValueRef visit_load_var(struct ac_nir_context *ctx,
2207 nir_intrinsic_instr *instr)
2208 {
2209 nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2210 nir_variable *var = nir_deref_instr_get_variable(deref);
2211
2212 LLVMValueRef values[8];
2213 int idx = 0;
2214 int ve = instr->dest.ssa.num_components;
2215 unsigned comp = 0;
2216 LLVMValueRef indir_index;
2217 LLVMValueRef ret;
2218 unsigned const_index;
2219 unsigned stride = 4;
2220 int mode = deref->mode;
2221
2222 if (var) {
2223 bool vs_in = ctx->stage == MESA_SHADER_VERTEX &&
2224 var->data.mode == nir_var_shader_in;
2225 idx = var->data.driver_location;
2226 comp = var->data.location_frac;
2227 mode = var->data.mode;
2228
2229 get_deref_offset(ctx, deref, vs_in, NULL, NULL,
2230 &const_index, &indir_index);
2231
2232 if (var->data.compact) {
2233 stride = 1;
2234 const_index += comp;
2235 comp = 0;
2236 }
2237 }
2238
2239 if (instr->dest.ssa.bit_size == 64 &&
2240 (deref->mode == nir_var_shader_in ||
2241 deref->mode == nir_var_shader_out ||
2242 deref->mode == nir_var_function_temp))
2243 ve *= 2;
2244
2245 switch (mode) {
2246 case nir_var_shader_in:
2247 if (ctx->stage == MESA_SHADER_TESS_CTRL ||
2248 ctx->stage == MESA_SHADER_TESS_EVAL) {
2249 return load_tess_varyings(ctx, instr, true);
2250 }
2251
2252 if (ctx->stage == MESA_SHADER_GEOMETRY) {
2253 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
2254 LLVMValueRef indir_index;
2255 unsigned const_index, vertex_index;
2256 get_deref_offset(ctx, deref, false, &vertex_index, NULL,
2257 &const_index, &indir_index);
2258 assert(indir_index == NULL);
2259
2260 return ctx->abi->load_inputs(ctx->abi, var->data.location,
2261 var->data.driver_location,
2262 var->data.location_frac,
2263 instr->num_components, vertex_index, const_index, type);
2264 }
2265
2266 for (unsigned chan = comp; chan < ve + comp; chan++) {
2267 if (indir_index) {
2268 unsigned count = glsl_count_attribute_slots(
2269 var->type,
2270 ctx->stage == MESA_SHADER_VERTEX);
2271 count -= chan / 4;
2272 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2273 &ctx->ac, ctx->abi->inputs + idx + chan, count,
2274 stride, false, true);
2275
2276 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2277 tmp_vec,
2278 indir_index, "");
2279 } else
2280 values[chan] = ctx->abi->inputs[idx + chan + const_index * stride];
2281 }
2282 break;
2283 case nir_var_function_temp:
2284 for (unsigned chan = 0; chan < ve; chan++) {
2285 if (indir_index) {
2286 unsigned count = glsl_count_attribute_slots(
2287 var->type, false);
2288 count -= chan / 4;
2289 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2290 &ctx->ac, ctx->locals + idx + chan, count,
2291 stride, true, true);
2292
2293 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2294 tmp_vec,
2295 indir_index, "");
2296 } else {
2297 values[chan] = LLVMBuildLoad(ctx->ac.builder, ctx->locals[idx + chan + const_index * stride], "");
2298 }
2299 }
2300 break;
2301 case nir_var_shader_out:
2302 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
2303 return load_tess_varyings(ctx, instr, false);
2304 }
2305
2306 if (ctx->stage == MESA_SHADER_FRAGMENT &&
2307 var->data.fb_fetch_output &&
2308 ctx->abi->emit_fbfetch)
2309 return ctx->abi->emit_fbfetch(ctx->abi);
2310
2311 for (unsigned chan = comp; chan < ve + comp; chan++) {
2312 if (indir_index) {
2313 unsigned count = glsl_count_attribute_slots(
2314 var->type, false);
2315 count -= chan / 4;
2316 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2317 &ctx->ac, ctx->abi->outputs + idx + chan, count,
2318 stride, true, true);
2319
2320 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2321 tmp_vec,
2322 indir_index, "");
2323 } else {
2324 values[chan] = LLVMBuildLoad(ctx->ac.builder,
2325 ctx->abi->outputs[idx + chan + const_index * stride],
2326 "");
2327 }
2328 }
2329 break;
2330 case nir_var_mem_global: {
2331 LLVMValueRef address = get_src(ctx, instr->src[0]);
2332 unsigned explicit_stride = glsl_get_explicit_stride(deref->type);
2333 unsigned natural_stride = type_scalar_size_bytes(deref->type);
2334 unsigned stride = explicit_stride ? explicit_stride : natural_stride;
2335
2336 LLVMTypeRef result_type = get_def_type(ctx, &instr->dest.ssa);
2337 if (stride != natural_stride) {
2338 LLVMTypeRef ptr_type = LLVMPointerType(LLVMGetElementType(result_type),
2339 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2340 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2341
2342 for (unsigned i = 0; i < instr->dest.ssa.num_components; ++i) {
2343 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, i * stride / natural_stride, 0);
2344 values[i] = LLVMBuildLoad(ctx->ac.builder,
2345 ac_build_gep_ptr(&ctx->ac, address, offset), "");
2346 }
2347 return ac_build_gather_values(&ctx->ac, values, instr->dest.ssa.num_components);
2348 } else {
2349 LLVMTypeRef ptr_type = LLVMPointerType(result_type,
2350 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2351 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2352 LLVMValueRef val = LLVMBuildLoad(ctx->ac.builder, address, "");
2353 return val;
2354 }
2355 }
2356 default:
2357 unreachable("unhandle variable mode");
2358 }
2359 ret = ac_build_varying_gather_values(&ctx->ac, values, ve, comp);
2360 return LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->dest.ssa), "");
2361 }
2362
2363 static void
2364 visit_store_var(struct ac_nir_context *ctx,
2365 nir_intrinsic_instr *instr)
2366 {
2367 if (ctx->ac.postponed_kill) {
2368 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
2369 ctx->ac.postponed_kill, "");
2370 ac_build_ifcc(&ctx->ac, cond, 7002);
2371 }
2372
2373 nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2374 nir_variable *var = nir_deref_instr_get_variable(deref);
2375
2376 LLVMValueRef temp_ptr, value;
2377 int idx = 0;
2378 unsigned comp = 0;
2379 LLVMValueRef src = ac_to_float(&ctx->ac, get_src(ctx, instr->src[1]));
2380 int writemask = instr->const_index[0];
2381 LLVMValueRef indir_index;
2382 unsigned const_index;
2383
2384 if (var) {
2385 get_deref_offset(ctx, deref, false,
2386 NULL, NULL, &const_index, &indir_index);
2387 idx = var->data.driver_location;
2388 comp = var->data.location_frac;
2389
2390 if (var->data.compact) {
2391 const_index += comp;
2392 comp = 0;
2393 }
2394 }
2395
2396 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src)) == 64 &&
2397 (deref->mode == nir_var_shader_out ||
2398 deref->mode == nir_var_function_temp)) {
2399
2400 src = LLVMBuildBitCast(ctx->ac.builder, src,
2401 LLVMVectorType(ctx->ac.f32, ac_get_llvm_num_components(src) * 2),
2402 "");
2403
2404 writemask = widen_mask(writemask, 2);
2405 }
2406
2407 writemask = writemask << comp;
2408
2409 switch (deref->mode) {
2410 case nir_var_shader_out:
2411
2412 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
2413 LLVMValueRef vertex_index = NULL;
2414 LLVMValueRef indir_index = NULL;
2415 unsigned const_index = 0;
2416 const bool is_patch = var->data.patch ||
2417 var->data.location == VARYING_SLOT_TESS_LEVEL_INNER ||
2418 var->data.location == VARYING_SLOT_TESS_LEVEL_OUTER;
2419
2420 get_deref_offset(ctx, deref, false, NULL,
2421 is_patch ? NULL : &vertex_index,
2422 &const_index, &indir_index);
2423
2424 ctx->abi->store_tcs_outputs(ctx->abi, var,
2425 vertex_index, indir_index,
2426 const_index, src, writemask);
2427 break;
2428 }
2429
2430 for (unsigned chan = 0; chan < 8; chan++) {
2431 int stride = 4;
2432 if (!(writemask & (1 << chan)))
2433 continue;
2434
2435 value = ac_llvm_extract_elem(&ctx->ac, src, chan - comp);
2436
2437 if (var->data.compact)
2438 stride = 1;
2439 if (indir_index) {
2440 unsigned count = glsl_count_attribute_slots(
2441 var->type, false);
2442 count -= chan / 4;
2443 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2444 &ctx->ac, ctx->abi->outputs + idx + chan, count,
2445 stride, true, true);
2446
2447 tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
2448 value, indir_index, "");
2449 build_store_values_extended(&ctx->ac, ctx->abi->outputs + idx + chan,
2450 count, stride, tmp_vec);
2451
2452 } else {
2453 temp_ptr = ctx->abi->outputs[idx + chan + const_index * stride];
2454
2455 LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
2456 }
2457 }
2458 break;
2459 case nir_var_function_temp:
2460 for (unsigned chan = 0; chan < 8; chan++) {
2461 if (!(writemask & (1 << chan)))
2462 continue;
2463
2464 value = ac_llvm_extract_elem(&ctx->ac, src, chan);
2465 if (indir_index) {
2466 unsigned count = glsl_count_attribute_slots(
2467 var->type, false);
2468 count -= chan / 4;
2469 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2470 &ctx->ac, ctx->locals + idx + chan, count,
2471 4, true, true);
2472
2473 tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
2474 value, indir_index, "");
2475 build_store_values_extended(&ctx->ac, ctx->locals + idx + chan,
2476 count, 4, tmp_vec);
2477 } else {
2478 temp_ptr = ctx->locals[idx + chan + const_index * 4];
2479
2480 LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
2481 }
2482 }
2483 break;
2484
2485 case nir_var_mem_global: {
2486 int writemask = instr->const_index[0];
2487 LLVMValueRef address = get_src(ctx, instr->src[0]);
2488 LLVMValueRef val = get_src(ctx, instr->src[1]);
2489
2490 unsigned explicit_stride = glsl_get_explicit_stride(deref->type);
2491 unsigned natural_stride = type_scalar_size_bytes(deref->type);
2492 unsigned stride = explicit_stride ? explicit_stride : natural_stride;
2493
2494 LLVMTypeRef ptr_type = LLVMPointerType(LLVMTypeOf(val),
2495 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2496 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2497
2498 if (writemask == (1u << ac_get_llvm_num_components(val)) - 1 &&
2499 stride == natural_stride) {
2500 LLVMTypeRef ptr_type = LLVMPointerType(LLVMTypeOf(val),
2501 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2502 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2503
2504 val = LLVMBuildBitCast(ctx->ac.builder, val,
2505 LLVMGetElementType(LLVMTypeOf(address)), "");
2506 LLVMBuildStore(ctx->ac.builder, val, address);
2507 } else {
2508 LLVMTypeRef ptr_type = LLVMPointerType(LLVMGetElementType(LLVMTypeOf(val)),
2509 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2510 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2511 for (unsigned chan = 0; chan < 4; chan++) {
2512 if (!(writemask & (1 << chan)))
2513 continue;
2514
2515 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, chan * stride / natural_stride, 0);
2516
2517 LLVMValueRef ptr = ac_build_gep_ptr(&ctx->ac, address, offset);
2518 LLVMValueRef src = ac_llvm_extract_elem(&ctx->ac, val,
2519 chan);
2520 src = LLVMBuildBitCast(ctx->ac.builder, src,
2521 LLVMGetElementType(LLVMTypeOf(ptr)), "");
2522 LLVMBuildStore(ctx->ac.builder, src, ptr);
2523 }
2524 }
2525 break;
2526 }
2527 default:
2528 abort();
2529 break;
2530 }
2531
2532 if (ctx->ac.postponed_kill)
2533 ac_build_endif(&ctx->ac, 7002);
2534 }
2535
2536 static int image_type_to_components_count(enum glsl_sampler_dim dim, bool array)
2537 {
2538 switch (dim) {
2539 case GLSL_SAMPLER_DIM_BUF:
2540 return 1;
2541 case GLSL_SAMPLER_DIM_1D:
2542 return array ? 2 : 1;
2543 case GLSL_SAMPLER_DIM_2D:
2544 return array ? 3 : 2;
2545 case GLSL_SAMPLER_DIM_MS:
2546 return array ? 4 : 3;
2547 case GLSL_SAMPLER_DIM_3D:
2548 case GLSL_SAMPLER_DIM_CUBE:
2549 return 3;
2550 case GLSL_SAMPLER_DIM_RECT:
2551 case GLSL_SAMPLER_DIM_SUBPASS:
2552 return 2;
2553 case GLSL_SAMPLER_DIM_SUBPASS_MS:
2554 return 3;
2555 default:
2556 break;
2557 }
2558 return 0;
2559 }
2560
2561 static LLVMValueRef adjust_sample_index_using_fmask(struct ac_llvm_context *ctx,
2562 LLVMValueRef coord_x, LLVMValueRef coord_y,
2563 LLVMValueRef coord_z,
2564 LLVMValueRef sample_index,
2565 LLVMValueRef fmask_desc_ptr)
2566 {
2567 unsigned sample_chan = coord_z ? 3 : 2;
2568 LLVMValueRef addr[4] = {coord_x, coord_y, coord_z};
2569 addr[sample_chan] = sample_index;
2570
2571 ac_apply_fmask_to_sample(ctx, fmask_desc_ptr, addr, coord_z != NULL);
2572 return addr[sample_chan];
2573 }
2574
2575 static nir_deref_instr *get_image_deref(const nir_intrinsic_instr *instr)
2576 {
2577 assert(instr->src[0].is_ssa);
2578 return nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2579 }
2580
2581 static LLVMValueRef get_image_descriptor(struct ac_nir_context *ctx,
2582 const nir_intrinsic_instr *instr,
2583 LLVMValueRef dynamic_index,
2584 enum ac_descriptor_type desc_type,
2585 bool write)
2586 {
2587 nir_deref_instr *deref_instr =
2588 instr->src[0].ssa->parent_instr->type == nir_instr_type_deref ?
2589 nir_instr_as_deref(instr->src[0].ssa->parent_instr) : NULL;
2590
2591 return get_sampler_desc(ctx, deref_instr, desc_type, &instr->instr, dynamic_index, true, write);
2592 }
2593
2594 static void get_image_coords(struct ac_nir_context *ctx,
2595 const nir_intrinsic_instr *instr,
2596 LLVMValueRef dynamic_desc_index,
2597 struct ac_image_args *args,
2598 enum glsl_sampler_dim dim,
2599 bool is_array)
2600 {
2601 LLVMValueRef src0 = get_src(ctx, instr->src[1]);
2602 LLVMValueRef masks[] = {
2603 LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
2604 LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
2605 };
2606 LLVMValueRef sample_index = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
2607
2608 int count;
2609 ASSERTED bool add_frag_pos = (dim == GLSL_SAMPLER_DIM_SUBPASS ||
2610 dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2611 bool is_ms = (dim == GLSL_SAMPLER_DIM_MS ||
2612 dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2613 bool gfx9_1d = ctx->ac.chip_class == GFX9 && dim == GLSL_SAMPLER_DIM_1D;
2614 assert(!add_frag_pos && "Input attachments should be lowered by this point.");
2615 count = image_type_to_components_count(dim, is_array);
2616
2617 if (is_ms && (instr->intrinsic == nir_intrinsic_image_deref_load ||
2618 instr->intrinsic == nir_intrinsic_bindless_image_load)) {
2619 LLVMValueRef fmask_load_address[3];
2620
2621 fmask_load_address[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
2622 fmask_load_address[1] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[1], "");
2623 if (is_array)
2624 fmask_load_address[2] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[2], "");
2625 else
2626 fmask_load_address[2] = NULL;
2627
2628 sample_index = adjust_sample_index_using_fmask(&ctx->ac,
2629 fmask_load_address[0],
2630 fmask_load_address[1],
2631 fmask_load_address[2],
2632 sample_index,
2633 get_sampler_desc(ctx, nir_instr_as_deref(instr->src[0].ssa->parent_instr),
2634 AC_DESC_FMASK, &instr->instr, dynamic_desc_index, true, false));
2635 }
2636 if (count == 1 && !gfx9_1d) {
2637 if (instr->src[1].ssa->num_components)
2638 args->coords[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
2639 else
2640 args->coords[0] = src0;
2641 } else {
2642 int chan;
2643 if (is_ms)
2644 count--;
2645 for (chan = 0; chan < count; ++chan) {
2646 args->coords[chan] = ac_llvm_extract_elem(&ctx->ac, src0, chan);
2647 }
2648
2649 if (gfx9_1d) {
2650 if (is_array) {
2651 args->coords[2] = args->coords[1];
2652 args->coords[1] = ctx->ac.i32_0;
2653 } else
2654 args->coords[1] = ctx->ac.i32_0;
2655 count++;
2656 }
2657 if (ctx->ac.chip_class == GFX9 &&
2658 dim == GLSL_SAMPLER_DIM_2D &&
2659 !is_array) {
2660 /* The hw can't bind a slice of a 3D image as a 2D
2661 * image, because it ignores BASE_ARRAY if the target
2662 * is 3D. The workaround is to read BASE_ARRAY and set
2663 * it as the 3rd address operand for all 2D images.
2664 */
2665 LLVMValueRef first_layer, const5, mask;
2666
2667 const5 = LLVMConstInt(ctx->ac.i32, 5, 0);
2668 mask = LLVMConstInt(ctx->ac.i32, S_008F24_BASE_ARRAY(~0), 0);
2669 first_layer = LLVMBuildExtractElement(ctx->ac.builder, args->resource, const5, "");
2670 first_layer = LLVMBuildAnd(ctx->ac.builder, first_layer, mask, "");
2671
2672 args->coords[count] = first_layer;
2673 count++;
2674 }
2675
2676
2677 if (is_ms) {
2678 args->coords[count] = sample_index;
2679 count++;
2680 }
2681 }
2682 }
2683
2684 static LLVMValueRef get_image_buffer_descriptor(struct ac_nir_context *ctx,
2685 const nir_intrinsic_instr *instr,
2686 LLVMValueRef dynamic_index,
2687 bool write, bool atomic)
2688 {
2689 LLVMValueRef rsrc = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_BUFFER, write);
2690 if (ctx->ac.chip_class == GFX9 && LLVM_VERSION_MAJOR < 9 && atomic) {
2691 LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 2, 0), "");
2692 LLVMValueRef stride = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 1, 0), "");
2693 stride = LLVMBuildLShr(ctx->ac.builder, stride, LLVMConstInt(ctx->ac.i32, 16, 0), "");
2694
2695 LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->ac.builder,
2696 LLVMBuildICmp(ctx->ac.builder, LLVMIntUGT, elem_count, stride, ""),
2697 elem_count, stride, "");
2698
2699 rsrc = LLVMBuildInsertElement(ctx->ac.builder, rsrc, new_elem_count,
2700 LLVMConstInt(ctx->ac.i32, 2, 0), "");
2701 }
2702 return rsrc;
2703 }
2704
2705 static LLVMValueRef enter_waterfall_image(struct ac_nir_context *ctx,
2706 struct waterfall_context *wctx,
2707 const nir_intrinsic_instr *instr)
2708 {
2709 nir_deref_instr *deref_instr = NULL;
2710
2711 if (instr->src[0].ssa->parent_instr->type == nir_instr_type_deref)
2712 deref_instr = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2713
2714 LLVMValueRef value = get_sampler_desc_index(ctx, deref_instr, &instr->instr, true);
2715 return enter_waterfall(ctx, wctx, value, nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
2716 }
2717
2718 static LLVMValueRef visit_image_load(struct ac_nir_context *ctx,
2719 const nir_intrinsic_instr *instr,
2720 bool bindless)
2721 {
2722 LLVMValueRef res;
2723
2724 enum glsl_sampler_dim dim;
2725 enum gl_access_qualifier access;
2726 bool is_array;
2727 if (bindless) {
2728 dim = nir_intrinsic_image_dim(instr);
2729 access = nir_intrinsic_access(instr);
2730 is_array = nir_intrinsic_image_array(instr);
2731 } else {
2732 const nir_deref_instr *image_deref = get_image_deref(instr);
2733 const struct glsl_type *type = image_deref->type;
2734 const nir_variable *var = nir_deref_instr_get_variable(image_deref);
2735 dim = glsl_get_sampler_dim(type);
2736 access = var->data.access;
2737 is_array = glsl_sampler_type_is_array(type);
2738 }
2739
2740 struct waterfall_context wctx;
2741 LLVMValueRef dynamic_index = enter_waterfall_image(ctx, &wctx, instr);
2742
2743 struct ac_image_args args = {};
2744
2745 args.cache_policy = get_cache_policy(ctx, access, false, false);
2746
2747 if (dim == GLSL_SAMPLER_DIM_BUF) {
2748 unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
2749 unsigned num_channels = util_last_bit(mask);
2750 LLVMValueRef rsrc, vindex;
2751
2752 rsrc = get_image_buffer_descriptor(ctx, instr, dynamic_index, false, false);
2753 vindex = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
2754 ctx->ac.i32_0, "");
2755
2756 bool can_speculate = access & ACCESS_CAN_REORDER;
2757 res = ac_build_buffer_load_format(&ctx->ac, rsrc, vindex,
2758 ctx->ac.i32_0, num_channels,
2759 args.cache_policy,
2760 can_speculate);
2761 res = ac_build_expand_to_vec4(&ctx->ac, res, num_channels);
2762
2763 res = ac_trim_vector(&ctx->ac, res, instr->dest.ssa.num_components);
2764 res = ac_to_integer(&ctx->ac, res);
2765 }