projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
ac/llvm: add option to clamp division by zero
[mesa.git] / src / amd / llvm / ac_nir_to_llvm.c
1 /*
2 * Copyright © 2016 Bas Nieuwenhuizen
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include <llvm/Config/llvm-config.h>
25
26 #include "ac_nir_to_llvm.h"
27 #include "ac_llvm_build.h"
28 #include "ac_llvm_util.h"
29 #include "ac_binary.h"
30 #include "sid.h"
31 #include "nir/nir.h"
32 #include "nir/nir_deref.h"
33 #include "util/bitscan.h"
34 #include "util/u_math.h"
35 #include "ac_shader_abi.h"
36 #include "ac_shader_util.h"
37
38 struct ac_nir_context {
39 struct ac_llvm_context ac;
40 struct ac_shader_abi *abi;
41 const struct ac_shader_args *args;
42
43 gl_shader_stage stage;
44 shader_info *info;
45
46 LLVMValueRef *ssa_defs;
47
48 LLVMValueRef scratch;
49 LLVMValueRef constant_data;
50
51 struct hash_table *defs;
52 struct hash_table *phis;
53 struct hash_table *vars;
54 struct hash_table *verified_interp;
55
56 LLVMValueRef main_function;
57 LLVMBasicBlockRef continue_block;
58 LLVMBasicBlockRef break_block;
59
60 int num_locals;
61 LLVMValueRef *locals;
62 };
63
64 static LLVMValueRef get_sampler_desc_index(struct ac_nir_context *ctx,
65 nir_deref_instr *deref_instr,
66 const nir_instr *instr,
67 bool image);
68
69 static LLVMValueRef get_sampler_desc(struct ac_nir_context *ctx,
70 nir_deref_instr *deref_instr,
71 enum ac_descriptor_type desc_type,
72 const nir_instr *instr,
73 LLVMValueRef index,
74 bool image, bool write);
75
76 static void
77 build_store_values_extended(struct ac_llvm_context *ac,
78 LLVMValueRef *values,
79 unsigned value_count,
80 unsigned value_stride,
81 LLVMValueRef vec)
82 {
83 LLVMBuilderRef builder = ac->builder;
84 unsigned i;
85
86 for (i = 0; i < value_count; i++) {
87 LLVMValueRef ptr = values[i * value_stride];
88 LLVMValueRef index = LLVMConstInt(ac->i32, i, false);
89 LLVMValueRef value = LLVMBuildExtractElement(builder, vec, index, "");
90 LLVMBuildStore(builder, value, ptr);
91 }
92 }
93
94 static LLVMTypeRef get_def_type(struct ac_nir_context *ctx,
95 const nir_ssa_def *def)
96 {
97 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, def->bit_size);
98 if (def->num_components > 1) {
99 type = LLVMVectorType(type, def->num_components);
100 }
101 return type;
102 }
103
104 static LLVMValueRef get_src(struct ac_nir_context *nir, nir_src src)
105 {
106 assert(src.is_ssa);
107 return nir->ssa_defs[src.ssa->index];
108 }
109
110 static LLVMValueRef
111 get_memory_ptr(struct ac_nir_context *ctx, nir_src src, unsigned bit_size)
112 {
113 LLVMValueRef ptr = get_src(ctx, src);
114 ptr = LLVMBuildGEP(ctx->ac.builder, ctx->ac.lds, &ptr, 1, "");
115 int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
116
117 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, bit_size);
118
119 return LLVMBuildBitCast(ctx->ac.builder, ptr,
120 LLVMPointerType(type, addr_space), "");
121 }
122
123 static LLVMBasicBlockRef get_block(struct ac_nir_context *nir,
124 const struct nir_block *b)
125 {
126 struct hash_entry *entry = _mesa_hash_table_search(nir->defs, b);
127 return (LLVMBasicBlockRef)entry->data;
128 }
129
130 static LLVMValueRef get_alu_src(struct ac_nir_context *ctx,
131 nir_alu_src src,
132 unsigned num_components)
133 {
134 LLVMValueRef value = get_src(ctx, src.src);
135 bool need_swizzle = false;
136
137 assert(value);
138 unsigned src_components = ac_get_llvm_num_components(value);
139 for (unsigned i = 0; i < num_components; ++i) {
140 assert(src.swizzle[i] < src_components);
141 if (src.swizzle[i] != i)
142 need_swizzle = true;
143 }
144
145 if (need_swizzle || num_components != src_components) {
146 LLVMValueRef masks[] = {
147 LLVMConstInt(ctx->ac.i32, src.swizzle[0], false),
148 LLVMConstInt(ctx->ac.i32, src.swizzle[1], false),
149 LLVMConstInt(ctx->ac.i32, src.swizzle[2], false),
150 LLVMConstInt(ctx->ac.i32, src.swizzle[3], false)};
151
152 if (src_components > 1 && num_components == 1) {
153 value = LLVMBuildExtractElement(ctx->ac.builder, value,
154 masks[0], "");
155 } else if (src_components == 1 && num_components > 1) {
156 LLVMValueRef values[] = {value, value, value, value};
157 value = ac_build_gather_values(&ctx->ac, values, num_components);
158 } else {
159 LLVMValueRef swizzle = LLVMConstVector(masks, num_components);
160 value = LLVMBuildShuffleVector(ctx->ac.builder, value, value,
161 swizzle, "");
162 }
163 }
164 assert(!src.negate);
165 assert(!src.abs);
166 return value;
167 }
168
169 static LLVMValueRef emit_int_cmp(struct ac_llvm_context *ctx,
170 LLVMIntPredicate pred, LLVMValueRef src0,
171 LLVMValueRef src1)
172 {
173 LLVMTypeRef src0_type = LLVMTypeOf(src0);
174 LLVMTypeRef src1_type = LLVMTypeOf(src1);
175
176 if (LLVMGetTypeKind(src0_type) == LLVMPointerTypeKind &&
177 LLVMGetTypeKind(src1_type) != LLVMPointerTypeKind) {
178 src1 = LLVMBuildIntToPtr(ctx->builder, src1, src0_type, "");
179 } else if (LLVMGetTypeKind(src1_type) == LLVMPointerTypeKind &&
180 LLVMGetTypeKind(src0_type) != LLVMPointerTypeKind) {
181 src0 = LLVMBuildIntToPtr(ctx->builder, src0, src1_type, "");
182 }
183
184 LLVMValueRef result = LLVMBuildICmp(ctx->builder, pred, src0, src1, "");
185 return LLVMBuildSelect(ctx->builder, result,
186 LLVMConstInt(ctx->i32, 0xFFFFFFFF, false),
187 ctx->i32_0, "");
188 }
189
190 static LLVMValueRef emit_float_cmp(struct ac_llvm_context *ctx,
191 LLVMRealPredicate pred, LLVMValueRef src0,
192 LLVMValueRef src1)
193 {
194 LLVMValueRef result;
195 src0 = ac_to_float(ctx, src0);
196 src1 = ac_to_float(ctx, src1);
197 result = LLVMBuildFCmp(ctx->builder, pred, src0, src1, "");
198 return LLVMBuildSelect(ctx->builder, result,
199 LLVMConstInt(ctx->i32, 0xFFFFFFFF, false),
200 ctx->i32_0, "");
201 }
202
203 static LLVMValueRef emit_intrin_1f_param(struct ac_llvm_context *ctx,
204 const char *intrin,
205 LLVMTypeRef result_type,
206 LLVMValueRef src0)
207 {
208 char name[64], type[64];
209 LLVMValueRef params[] = {
210 ac_to_float(ctx, src0),
211 };
212
213 ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
214 ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
215 assert(length < sizeof(name));
216 return ac_build_intrinsic(ctx, name, result_type, params, 1, AC_FUNC_ATTR_READNONE);
217 }
218
219 static LLVMValueRef emit_intrin_2f_param(struct ac_llvm_context *ctx,
220 const char *intrin,
221 LLVMTypeRef result_type,
222 LLVMValueRef src0, LLVMValueRef src1)
223 {
224 char name[64], type[64];
225 LLVMValueRef params[] = {
226 ac_to_float(ctx, src0),
227 ac_to_float(ctx, src1),
228 };
229
230 ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
231 ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
232 assert(length < sizeof(name));
233 return ac_build_intrinsic(ctx, name, result_type, params, 2, AC_FUNC_ATTR_READNONE);
234 }
235
236 static LLVMValueRef emit_intrin_3f_param(struct ac_llvm_context *ctx,
237 const char *intrin,
238 LLVMTypeRef result_type,
239 LLVMValueRef src0, LLVMValueRef src1, LLVMValueRef src2)
240 {
241 char name[64], type[64];
242 LLVMValueRef params[] = {
243 ac_to_float(ctx, src0),
244 ac_to_float(ctx, src1),
245 ac_to_float(ctx, src2),
246 };
247
248 ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
249 ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
250 assert(length < sizeof(name));
251 return ac_build_intrinsic(ctx, name, result_type, params, 3, AC_FUNC_ATTR_READNONE);
252 }
253
254 static LLVMValueRef emit_bcsel(struct ac_llvm_context *ctx,
255 LLVMValueRef src0, LLVMValueRef src1, LLVMValueRef src2)
256 {
257 LLVMTypeRef src1_type = LLVMTypeOf(src1);
258 LLVMTypeRef src2_type = LLVMTypeOf(src2);
259
260 assert(LLVMGetTypeKind(LLVMTypeOf(src0)) != LLVMVectorTypeKind);
261
262 if (LLVMGetTypeKind(src1_type) == LLVMPointerTypeKind &&
263 LLVMGetTypeKind(src2_type) != LLVMPointerTypeKind) {
264 src2 = LLVMBuildIntToPtr(ctx->builder, src2, src1_type, "");
265 } else if (LLVMGetTypeKind(src2_type) == LLVMPointerTypeKind &&
266 LLVMGetTypeKind(src1_type) != LLVMPointerTypeKind) {
267 src1 = LLVMBuildIntToPtr(ctx->builder, src1, src2_type, "");
268 }
269
270 LLVMValueRef v = LLVMBuildICmp(ctx->builder, LLVMIntNE, src0,
271 ctx->i32_0, "");
272 return LLVMBuildSelect(ctx->builder, v,
273 ac_to_integer_or_pointer(ctx, src1),
274 ac_to_integer_or_pointer(ctx, src2), "");
275 }
276
277 static LLVMValueRef emit_iabs(struct ac_llvm_context *ctx,
278 LLVMValueRef src0)
279 {
280 return ac_build_imax(ctx, src0, LLVMBuildNeg(ctx->builder, src0, ""));
281 }
282
283 static LLVMValueRef emit_uint_carry(struct ac_llvm_context *ctx,
284 const char *intrin,
285 LLVMValueRef src0, LLVMValueRef src1)
286 {
287 LLVMTypeRef ret_type;
288 LLVMTypeRef types[] = { ctx->i32, ctx->i1 };
289 LLVMValueRef res;
290 LLVMValueRef params[] = { src0, src1 };
291 ret_type = LLVMStructTypeInContext(ctx->context, types,
292 2, true);
293
294 res = ac_build_intrinsic(ctx, intrin, ret_type,
295 params, 2, AC_FUNC_ATTR_READNONE);
296
297 res = LLVMBuildExtractValue(ctx->builder, res, 1, "");
298 res = LLVMBuildZExt(ctx->builder, res, ctx->i32, "");
299 return res;
300 }
301
302 static LLVMValueRef emit_b2f(struct ac_llvm_context *ctx,
303 LLVMValueRef src0,
304 unsigned bitsize)
305 {
306 LLVMValueRef result = LLVMBuildAnd(ctx->builder, src0,
307 LLVMBuildBitCast(ctx->builder, LLVMConstReal(ctx->f32, 1.0), ctx->i32, ""),
308 "");
309 result = LLVMBuildBitCast(ctx->builder, result, ctx->f32, "");
310
311 switch (bitsize) {
312 case 16:
313 return LLVMBuildFPTrunc(ctx->builder, result, ctx->f16, "");
314 case 32:
315 return result;
316 case 64:
317 return LLVMBuildFPExt(ctx->builder, result, ctx->f64, "");
318 default:
319 unreachable("Unsupported bit size.");
320 }
321 }
322
323 static LLVMValueRef emit_f2b(struct ac_llvm_context *ctx,
324 LLVMValueRef src0)
325 {
326 src0 = ac_to_float(ctx, src0);
327 LLVMValueRef zero = LLVMConstNull(LLVMTypeOf(src0));
328 return LLVMBuildSExt(ctx->builder,
329 LLVMBuildFCmp(ctx->builder, LLVMRealUNE, src0, zero, ""),
330 ctx->i32, "");
331 }
332
333 static LLVMValueRef emit_b2i(struct ac_llvm_context *ctx,
334 LLVMValueRef src0,
335 unsigned bitsize)
336 {
337 LLVMValueRef result = LLVMBuildAnd(ctx->builder, src0, ctx->i32_1, "");
338
339 switch (bitsize) {
340 case 8:
341 return LLVMBuildTrunc(ctx->builder, result, ctx->i8, "");
342 case 16:
343 return LLVMBuildTrunc(ctx->builder, result, ctx->i16, "");
344 case 32:
345 return result;
346 case 64:
347 return LLVMBuildZExt(ctx->builder, result, ctx->i64, "");
348 default:
349 unreachable("Unsupported bit size.");
350 }
351 }
352
353 static LLVMValueRef emit_i2b(struct ac_llvm_context *ctx,
354 LLVMValueRef src0)
355 {
356 LLVMValueRef zero = LLVMConstNull(LLVMTypeOf(src0));
357 return LLVMBuildSExt(ctx->builder,
358 LLVMBuildICmp(ctx->builder, LLVMIntNE, src0, zero, ""),
359 ctx->i32, "");
360 }
361
362 static LLVMValueRef emit_f2f16(struct ac_llvm_context *ctx,
363 LLVMValueRef src0)
364 {
365 LLVMValueRef result;
366 LLVMValueRef cond = NULL;
367
368 src0 = ac_to_float(ctx, src0);
369 result = LLVMBuildFPTrunc(ctx->builder, src0, ctx->f16, "");
370
371 if (ctx->chip_class >= GFX8) {
372 LLVMValueRef args[2];
373 /* Check if the result is a denormal - and flush to 0 if so. */
374 args[0] = result;
375 args[1] = LLVMConstInt(ctx->i32, N_SUBNORMAL | P_SUBNORMAL, false);
376 cond = ac_build_intrinsic(ctx, "llvm.amdgcn.class.f16", ctx->i1, args, 2, AC_FUNC_ATTR_READNONE);
377 }
378
379 /* need to convert back up to f32 */
380 result = LLVMBuildFPExt(ctx->builder, result, ctx->f32, "");
381
382 if (ctx->chip_class >= GFX8)
383 result = LLVMBuildSelect(ctx->builder, cond, ctx->f32_0, result, "");
384 else {
385 /* for GFX6-GFX7 */
386 /* 0x38800000 is smallest half float value (2^-14) in 32-bit float,
387 * so compare the result and flush to 0 if it's smaller.
388 */
389 LLVMValueRef temp, cond2;
390 temp = emit_intrin_1f_param(ctx, "llvm.fabs", ctx->f32, result);
391 cond = LLVMBuildFCmp(ctx->builder, LLVMRealOGT,
392 LLVMBuildBitCast(ctx->builder, LLVMConstInt(ctx->i32, 0x38800000, false), ctx->f32, ""),
393 temp, "");
394 cond2 = LLVMBuildFCmp(ctx->builder, LLVMRealONE,
395 temp, ctx->f32_0, "");
396 cond = LLVMBuildAnd(ctx->builder, cond, cond2, "");
397 result = LLVMBuildSelect(ctx->builder, cond, ctx->f32_0, result, "");
398 }
399 return result;
400 }
401
402 static LLVMValueRef emit_umul_high(struct ac_llvm_context *ctx,
403 LLVMValueRef src0, LLVMValueRef src1)
404 {
405 LLVMValueRef dst64, result;
406 src0 = LLVMBuildZExt(ctx->builder, src0, ctx->i64, "");
407 src1 = LLVMBuildZExt(ctx->builder, src1, ctx->i64, "");
408
409 dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
410 dst64 = LLVMBuildLShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
411 result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
412 return result;
413 }
414
415 static LLVMValueRef emit_imul_high(struct ac_llvm_context *ctx,
416 LLVMValueRef src0, LLVMValueRef src1)
417 {
418 LLVMValueRef dst64, result;
419 src0 = LLVMBuildSExt(ctx->builder, src0, ctx->i64, "");
420 src1 = LLVMBuildSExt(ctx->builder, src1, ctx->i64, "");
421
422 dst64 = LLVMBuildMul(ctx->builder, src0, src1, "");
423 dst64 = LLVMBuildAShr(ctx->builder, dst64, LLVMConstInt(ctx->i64, 32, false), "");
424 result = LLVMBuildTrunc(ctx->builder, dst64, ctx->i32, "");
425 return result;
426 }
427
428 static LLVMValueRef emit_bfm(struct ac_llvm_context *ctx,
429 LLVMValueRef bits, LLVMValueRef offset)
430 {
431 /* mask = ((1 << bits) - 1) << offset */
432 return LLVMBuildShl(ctx->builder,
433 LLVMBuildSub(ctx->builder,
434 LLVMBuildShl(ctx->builder,
435 ctx->i32_1,
436 bits, ""),
437 ctx->i32_1, ""),
438 offset, "");
439 }
440
441 static LLVMValueRef emit_bitfield_select(struct ac_llvm_context *ctx,
442 LLVMValueRef mask, LLVMValueRef insert,
443 LLVMValueRef base)
444 {
445 /* Calculate:
446 * (mask & insert) | (~mask & base) = base ^ (mask & (insert ^ base))
447 * Use the right-hand side, which the LLVM backend can convert to V_BFI.
448 */
449 return LLVMBuildXor(ctx->builder, base,
450 LLVMBuildAnd(ctx->builder, mask,
451 LLVMBuildXor(ctx->builder, insert, base, ""), ""), "");
452 }
453
454 static LLVMValueRef emit_pack_2x16(struct ac_llvm_context *ctx,
455 LLVMValueRef src0,
456 LLVMValueRef (*pack)(struct ac_llvm_context *ctx,
457 LLVMValueRef args[2]))
458 {
459 LLVMValueRef comp[2];
460
461 src0 = ac_to_float(ctx, src0);
462 comp[0] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_0, "");
463 comp[1] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_1, "");
464
465 return LLVMBuildBitCast(ctx->builder, pack(ctx, comp), ctx->i32, "");
466 }
467
468 static LLVMValueRef emit_unpack_half_2x16(struct ac_llvm_context *ctx,
469 LLVMValueRef src0)
470 {
471 LLVMValueRef const16 = LLVMConstInt(ctx->i32, 16, false);
472 LLVMValueRef temps[2], val;
473 int i;
474
475 for (i = 0; i < 2; i++) {
476 val = i == 1 ? LLVMBuildLShr(ctx->builder, src0, const16, "") : src0;
477 val = LLVMBuildTrunc(ctx->builder, val, ctx->i16, "");
478 val = LLVMBuildBitCast(ctx->builder, val, ctx->f16, "");
479 temps[i] = LLVMBuildFPExt(ctx->builder, val, ctx->f32, "");
480 }
481 return ac_build_gather_values(ctx, temps, 2);
482 }
483
484 static LLVMValueRef emit_ddxy(struct ac_nir_context *ctx,
485 nir_op op,
486 LLVMValueRef src0)
487 {
488 unsigned mask;
489 int idx;
490 LLVMValueRef result;
491
492 if (op == nir_op_fddx_fine)
493 mask = AC_TID_MASK_LEFT;
494 else if (op == nir_op_fddy_fine)
495 mask = AC_TID_MASK_TOP;
496 else
497 mask = AC_TID_MASK_TOP_LEFT;
498
499 /* for DDX we want to next X pixel, DDY next Y pixel. */
500 if (op == nir_op_fddx_fine ||
501 op == nir_op_fddx_coarse ||
502 op == nir_op_fddx)
503 idx = 1;
504 else
505 idx = 2;
506
507 result = ac_build_ddxy(&ctx->ac, mask, idx, src0);
508 return result;
509 }
510
511 struct waterfall_context {
512 LLVMBasicBlockRef phi_bb[2];
513 bool use_waterfall;
514 };
515
516 /* To deal with divergent descriptors we can create a loop that handles all
517 * lanes with the same descriptor on a given iteration (henceforth a
518 * waterfall loop).
519 *
520 * These helper create the begin and end of the loop leaving the caller
521 * to implement the body.
522 *
523 * params:
524 * - ctx is the usal nir context
525 * - wctx is a temporary struct containing some loop info. Can be left uninitialized.
526 * - value is the possibly divergent value for which we built the loop
527 * - divergent is whether value is actually divergent. If false we just pass
528 * things through.
529 */
530 static LLVMValueRef enter_waterfall(struct ac_nir_context *ctx,
531 struct waterfall_context *wctx,
532 LLVMValueRef value, bool divergent)
533 {
534 /* If the app claims the value is divergent but it is constant we can
535 * end up with a dynamic index of NULL. */
536 if (!value)
537 divergent = false;
538
539 wctx->use_waterfall = divergent;
540 if (!divergent)
541 return value;
542
543 ac_build_bgnloop(&ctx->ac, 6000);
544
545 LLVMValueRef scalar_value = ac_build_readlane(&ctx->ac, value, NULL);
546
547 LLVMValueRef active = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, value,
548 scalar_value, "uniform_active");
549
550 wctx->phi_bb[0] = LLVMGetInsertBlock(ctx->ac.builder);
551 ac_build_ifcc(&ctx->ac, active, 6001);
552
553 return scalar_value;
554 }
555
556 static LLVMValueRef exit_waterfall(struct ac_nir_context *ctx,
557 struct waterfall_context *wctx,
558 LLVMValueRef value)
559 {
560 LLVMValueRef ret = NULL;
561 LLVMValueRef phi_src[2];
562 LLVMValueRef cc_phi_src[2] = {
563 LLVMConstInt(ctx->ac.i32, 0, false),
564 LLVMConstInt(ctx->ac.i32, 0xffffffff, false),
565 };
566
567 if (!wctx->use_waterfall)
568 return value;
569
570 wctx->phi_bb[1] = LLVMGetInsertBlock(ctx->ac.builder);
571
572 ac_build_endif(&ctx->ac, 6001);
573
574 if (value) {
575 phi_src[0] = LLVMGetUndef(LLVMTypeOf(value));
576 phi_src[1] = value;
577
578 ret = ac_build_phi(&ctx->ac, LLVMTypeOf(value), 2, phi_src, wctx->phi_bb);
579 }
580
581 /*
582 * By using the optimization barrier on the exit decision, we decouple
583 * the operations from the break, and hence avoid LLVM hoisting the
584 * opteration into the break block.
585 */
586 LLVMValueRef cc = ac_build_phi(&ctx->ac, ctx->ac.i32, 2, cc_phi_src, wctx->phi_bb);
587 ac_build_optimization_barrier(&ctx->ac, &cc);
588
589 LLVMValueRef active = LLVMBuildICmp(ctx->ac.builder, LLVMIntNE, cc, ctx->ac.i32_0, "uniform_active2");
590 ac_build_ifcc(&ctx->ac, active, 6002);
591 ac_build_break(&ctx->ac);
592 ac_build_endif(&ctx->ac, 6002);
593
594 ac_build_endloop(&ctx->ac, 6000);
595 return ret;
596 }
597
598 static void visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
599 {
600 LLVMValueRef src[4], result = NULL;
601 unsigned num_components = instr->dest.dest.ssa.num_components;
602 unsigned src_components;
603 LLVMTypeRef def_type = get_def_type(ctx, &instr->dest.dest.ssa);
604 bool saved_inexact = false;
605
606 if (instr->exact)
607 saved_inexact = ac_disable_inexact_math(ctx->ac.builder);
608
609 assert(nir_op_infos[instr->op].num_inputs <= ARRAY_SIZE(src));
610 switch (instr->op) {
611 case nir_op_vec2:
612 case nir_op_vec3:
613 case nir_op_vec4:
614 src_components = 1;
615 break;
616 case nir_op_pack_half_2x16:
617 case nir_op_pack_snorm_2x16:
618 case nir_op_pack_unorm_2x16:
619 src_components = 2;
620 break;
621 case nir_op_unpack_half_2x16:
622 src_components = 1;
623 break;
624 case nir_op_cube_face_coord:
625 case nir_op_cube_face_index:
626 src_components = 3;
627 break;
628 default:
629 src_components = num_components;
630 break;
631 }
632 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
633 src[i] = get_alu_src(ctx, instr->src[i], src_components);
634
635 switch (instr->op) {
636 case nir_op_mov:
637 result = src[0];
638 break;
639 case nir_op_fneg:
640 src[0] = ac_to_float(&ctx->ac, src[0]);
641 result = LLVMBuildFNeg(ctx->ac.builder, src[0], "");
642 if (ctx->ac.float_mode == AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO) {
643 /* fneg will be optimized by backend compiler with sign
644 * bit removed via XOR. This is probably a LLVM bug.
645 */
646 result = ac_build_canonicalize(&ctx->ac, result,
647 instr->dest.dest.ssa.bit_size);
648 }
649 break;
650 case nir_op_ineg:
651 result = LLVMBuildNeg(ctx->ac.builder, src[0], "");
652 break;
653 case nir_op_inot:
654 result = LLVMBuildNot(ctx->ac.builder, src[0], "");
655 break;
656 case nir_op_iadd:
657 result = LLVMBuildAdd(ctx->ac.builder, src[0], src[1], "");
658 break;
659 case nir_op_fadd:
660 src[0] = ac_to_float(&ctx->ac, src[0]);
661 src[1] = ac_to_float(&ctx->ac, src[1]);
662 result = LLVMBuildFAdd(ctx->ac.builder, src[0], src[1], "");
663 break;
664 case nir_op_fsub:
665 src[0] = ac_to_float(&ctx->ac, src[0]);
666 src[1] = ac_to_float(&ctx->ac, src[1]);
667 result = LLVMBuildFSub(ctx->ac.builder, src[0], src[1], "");
668 break;
669 case nir_op_isub:
670 result = LLVMBuildSub(ctx->ac.builder, src[0], src[1], "");
671 break;
672 case nir_op_imul:
673 result = LLVMBuildMul(ctx->ac.builder, src[0], src[1], "");
674 break;
675 case nir_op_imod:
676 result = LLVMBuildSRem(ctx->ac.builder, src[0], src[1], "");
677 break;
678 case nir_op_umod:
679 result = LLVMBuildURem(ctx->ac.builder, src[0], src[1], "");
680 break;
681 case nir_op_fmod:
682 /* lower_fmod only lower 16-bit and 32-bit fmod */
683 assert(instr->dest.dest.ssa.bit_size == 64);
684 src[0] = ac_to_float(&ctx->ac, src[0]);
685 src[1] = ac_to_float(&ctx->ac, src[1]);
686 result = ac_build_fdiv(&ctx->ac, src[0], src[1]);
687 result = emit_intrin_1f_param(&ctx->ac, "llvm.floor",
688 ac_to_float_type(&ctx->ac, def_type), result);
689 result = LLVMBuildFMul(ctx->ac.builder, src[1] , result, "");
690 result = LLVMBuildFSub(ctx->ac.builder, src[0], result, "");
691 break;
692 case nir_op_irem:
693 result = LLVMBuildSRem(ctx->ac.builder, src[0], src[1], "");
694 break;
695 case nir_op_idiv:
696 result = LLVMBuildSDiv(ctx->ac.builder, src[0], src[1], "");
697 break;
698 case nir_op_udiv:
699 result = LLVMBuildUDiv(ctx->ac.builder, src[0], src[1], "");
700 break;
701 case nir_op_fmul:
702 src[0] = ac_to_float(&ctx->ac, src[0]);
703 src[1] = ac_to_float(&ctx->ac, src[1]);
704 result = LLVMBuildFMul(ctx->ac.builder, src[0], src[1], "");
705 break;
706 case nir_op_frcp:
707 /* For doubles, we need precise division to pass GLCTS. */
708 if (ctx->ac.float_mode == AC_FLOAT_MODE_DEFAULT_OPENGL &&
709 ac_get_type_size(def_type) == 8) {
710 result = LLVMBuildFDiv(ctx->ac.builder, ctx->ac.f64_1,
711 ac_to_float(&ctx->ac, src[0]), "");
712 } else {
713 result = emit_intrin_1f_param(&ctx->ac, "llvm.amdgcn.rcp",
714 ac_to_float_type(&ctx->ac, def_type), src[0]);
715 }
716 if (ctx->abi->clamp_div_by_zero)
717 result = ac_build_fmin(&ctx->ac, result,
718 LLVMConstReal(ac_to_float_type(&ctx->ac, def_type), FLT_MAX));
719 break;
720 case nir_op_iand:
721 result = LLVMBuildAnd(ctx->ac.builder, src[0], src[1], "");
722 break;
723 case nir_op_ior:
724 result = LLVMBuildOr(ctx->ac.builder, src[0], src[1], "");
725 break;
726 case nir_op_ixor:
727 result = LLVMBuildXor(ctx->ac.builder, src[0], src[1], "");
728 break;
729 case nir_op_ishl:
730 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) < ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
731 src[1] = LLVMBuildZExt(ctx->ac.builder, src[1],
732 LLVMTypeOf(src[0]), "");
733 else if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) > ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
734 src[1] = LLVMBuildTrunc(ctx->ac.builder, src[1],
735 LLVMTypeOf(src[0]), "");
736 result = LLVMBuildShl(ctx->ac.builder, src[0], src[1], "");
737 break;
738 case nir_op_ishr:
739 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) < ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
740 src[1] = LLVMBuildZExt(ctx->ac.builder, src[1],
741 LLVMTypeOf(src[0]), "");
742 else if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) > ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
743 src[1] = LLVMBuildTrunc(ctx->ac.builder, src[1],
744 LLVMTypeOf(src[0]), "");
745 result = LLVMBuildAShr(ctx->ac.builder, src[0], src[1], "");
746 break;
747 case nir_op_ushr:
748 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) < ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
749 src[1] = LLVMBuildZExt(ctx->ac.builder, src[1],
750 LLVMTypeOf(src[0]), "");
751 else if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[1])) > ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])))
752 src[1] = LLVMBuildTrunc(ctx->ac.builder, src[1],
753 LLVMTypeOf(src[0]), "");
754 result = LLVMBuildLShr(ctx->ac.builder, src[0], src[1], "");
755 break;
756 case nir_op_ilt32:
757 result = emit_int_cmp(&ctx->ac, LLVMIntSLT, src[0], src[1]);
758 break;
759 case nir_op_ine32:
760 result = emit_int_cmp(&ctx->ac, LLVMIntNE, src[0], src[1]);
761 break;
762 case nir_op_ieq32:
763 result = emit_int_cmp(&ctx->ac, LLVMIntEQ, src[0], src[1]);
764 break;
765 case nir_op_ige32:
766 result = emit_int_cmp(&ctx->ac, LLVMIntSGE, src[0], src[1]);
767 break;
768 case nir_op_ult32:
769 result = emit_int_cmp(&ctx->ac, LLVMIntULT, src[0], src[1]);
770 break;
771 case nir_op_uge32:
772 result = emit_int_cmp(&ctx->ac, LLVMIntUGE, src[0], src[1]);
773 break;
774 case nir_op_feq32:
775 result = emit_float_cmp(&ctx->ac, LLVMRealOEQ, src[0], src[1]);
776 break;
777 case nir_op_fneu32:
778 result = emit_float_cmp(&ctx->ac, LLVMRealUNE, src[0], src[1]);
779 break;
780 case nir_op_flt32:
781 result = emit_float_cmp(&ctx->ac, LLVMRealOLT, src[0], src[1]);
782 break;
783 case nir_op_fge32:
784 result = emit_float_cmp(&ctx->ac, LLVMRealOGE, src[0], src[1]);
785 break;
786 case nir_op_fabs:
787 result = emit_intrin_1f_param(&ctx->ac, "llvm.fabs",
788 ac_to_float_type(&ctx->ac, def_type), src[0]);
789 if (ctx->ac.float_mode == AC_FLOAT_MODE_DENORM_FLUSH_TO_ZERO) {
790 /* fabs will be optimized by backend compiler with sign
791 * bit removed via AND.
792 */
793 result = ac_build_canonicalize(&ctx->ac, result,
794 instr->dest.dest.ssa.bit_size);
795 }
796 break;
797 case nir_op_iabs:
798 result = emit_iabs(&ctx->ac, src[0]);
799 break;
800 case nir_op_imax:
801 result = ac_build_imax(&ctx->ac, src[0], src[1]);
802 break;
803 case nir_op_imin:
804 result = ac_build_imin(&ctx->ac, src[0], src[1]);
805 break;
806 case nir_op_umax:
807 result = ac_build_umax(&ctx->ac, src[0], src[1]);
808 break;
809 case nir_op_umin:
810 result = ac_build_umin(&ctx->ac, src[0], src[1]);
811 break;
812 case nir_op_isign:
813 result = ac_build_isign(&ctx->ac, src[0],
814 instr->dest.dest.ssa.bit_size);
815 break;
816 case nir_op_fsign:
817 src[0] = ac_to_float(&ctx->ac, src[0]);
818 result = ac_build_fsign(&ctx->ac, src[0],
819 instr->dest.dest.ssa.bit_size);
820 break;
821 case nir_op_ffloor:
822 result = emit_intrin_1f_param(&ctx->ac, "llvm.floor",
823 ac_to_float_type(&ctx->ac, def_type), src[0]);
824 break;
825 case nir_op_ftrunc:
826 result = emit_intrin_1f_param(&ctx->ac, "llvm.trunc",
827 ac_to_float_type(&ctx->ac, def_type), src[0]);
828 break;
829 case nir_op_fceil:
830 result = emit_intrin_1f_param(&ctx->ac, "llvm.ceil",
831 ac_to_float_type(&ctx->ac, def_type), src[0]);
832 break;
833 case nir_op_fround_even:
834 result = emit_intrin_1f_param(&ctx->ac, "llvm.rint",
835 ac_to_float_type(&ctx->ac, def_type),src[0]);
836 break;
837 case nir_op_ffract:
838 src[0] = ac_to_float(&ctx->ac, src[0]);
839 result = ac_build_fract(&ctx->ac, src[0],
840 instr->dest.dest.ssa.bit_size);
841 break;
842 case nir_op_fsin:
843 result = emit_intrin_1f_param(&ctx->ac, "llvm.sin",
844 ac_to_float_type(&ctx->ac, def_type), src[0]);
845 break;
846 case nir_op_fcos:
847 result = emit_intrin_1f_param(&ctx->ac, "llvm.cos",
848 ac_to_float_type(&ctx->ac, def_type), src[0]);
849 break;
850 case nir_op_fsqrt:
851 result = emit_intrin_1f_param(&ctx->ac, "llvm.sqrt",
852 ac_to_float_type(&ctx->ac, def_type), src[0]);
853 break;
854 case nir_op_fexp2:
855 result = emit_intrin_1f_param(&ctx->ac, "llvm.exp2",
856 ac_to_float_type(&ctx->ac, def_type), src[0]);
857 break;
858 case nir_op_flog2:
859 result = emit_intrin_1f_param(&ctx->ac, "llvm.log2",
860 ac_to_float_type(&ctx->ac, def_type), src[0]);
861 break;
862 case nir_op_frsq:
863 result = emit_intrin_1f_param(&ctx->ac, "llvm.amdgcn.rsq",
864 ac_to_float_type(&ctx->ac, def_type), src[0]);
865 if (ctx->abi->clamp_div_by_zero)
866 result = ac_build_fmin(&ctx->ac, result,
867 LLVMConstReal(ac_to_float_type(&ctx->ac, def_type), FLT_MAX));
868 break;
869 case nir_op_frexp_exp:
870 src[0] = ac_to_float(&ctx->ac, src[0]);
871 result = ac_build_frexp_exp(&ctx->ac, src[0],
872 ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])));
873 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) == 16)
874 result = LLVMBuildSExt(ctx->ac.builder, result,
875 ctx->ac.i32, "");
876 break;
877 case nir_op_frexp_sig:
878 src[0] = ac_to_float(&ctx->ac, src[0]);
879 result = ac_build_frexp_mant(&ctx->ac, src[0],
880 instr->dest.dest.ssa.bit_size);
881 break;
882 case nir_op_fpow:
883 result = emit_intrin_2f_param(&ctx->ac, "llvm.pow",
884 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
885 break;
886 case nir_op_fmax:
887 result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
888 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
889 if (ctx->ac.chip_class < GFX9 &&
890 instr->dest.dest.ssa.bit_size == 32) {
891 /* Only pre-GFX9 chips do not flush denorms. */
892 result = ac_build_canonicalize(&ctx->ac, result,
893 instr->dest.dest.ssa.bit_size);
894 }
895 break;
896 case nir_op_fmin:
897 result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
898 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
899 if (ctx->ac.chip_class < GFX9 &&
900 instr->dest.dest.ssa.bit_size == 32) {
901 /* Only pre-GFX9 chips do not flush denorms. */
902 result = ac_build_canonicalize(&ctx->ac, result,
903 instr->dest.dest.ssa.bit_size);
904 }
905 break;
906 case nir_op_ffma:
907 /* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
908 result = emit_intrin_3f_param(&ctx->ac, ctx->ac.chip_class >= GFX10 ? "llvm.fma" : "llvm.fmuladd",
909 ac_to_float_type(&ctx->ac, def_type), src[0], src[1], src[2]);
910 break;
911 case nir_op_ldexp:
912 src[0] = ac_to_float(&ctx->ac, src[0]);
913 if (ac_get_elem_bits(&ctx->ac, def_type) == 32)
914 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f32", ctx->ac.f32, src, 2, AC_FUNC_ATTR_READNONE);
915 else if (ac_get_elem_bits(&ctx->ac, def_type) == 16)
916 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f16", ctx->ac.f16, src, 2, AC_FUNC_ATTR_READNONE);
917 else
918 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f64", ctx->ac.f64, src, 2, AC_FUNC_ATTR_READNONE);
919 break;
920 case nir_op_bfm:
921 result = emit_bfm(&ctx->ac, src[0], src[1]);
922 break;
923 case nir_op_bitfield_select:
924 result = emit_bitfield_select(&ctx->ac, src[0], src[1], src[2]);
925 break;
926 case nir_op_ubfe:
927 result = ac_build_bfe(&ctx->ac, src[0], src[1], src[2], false);
928 break;
929 case nir_op_ibfe:
930 result = ac_build_bfe(&ctx->ac, src[0], src[1], src[2], true);
931 break;
932 case nir_op_bitfield_reverse:
933 result = ac_build_bitfield_reverse(&ctx->ac, src[0]);
934 break;
935 case nir_op_bit_count:
936 result = ac_build_bit_count(&ctx->ac, src[0]);
937 break;
938 case nir_op_vec2:
939 case nir_op_vec3:
940 case nir_op_vec4:
941 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
942 src[i] = ac_to_integer(&ctx->ac, src[i]);
943 result = ac_build_gather_values(&ctx->ac, src, num_components);
944 break;
945 case nir_op_f2i8:
946 case nir_op_f2i16:
947 case nir_op_f2i32:
948 case nir_op_f2i64:
949 src[0] = ac_to_float(&ctx->ac, src[0]);
950 result = LLVMBuildFPToSI(ctx->ac.builder, src[0], def_type, "");
951 break;
952 case nir_op_f2u8:
953 case nir_op_f2u16:
954 case nir_op_f2u32:
955 case nir_op_f2u64:
956 src[0] = ac_to_float(&ctx->ac, src[0]);
957 result = LLVMBuildFPToUI(ctx->ac.builder, src[0], def_type, "");
958 break;
959 case nir_op_i2f16:
960 case nir_op_i2f32:
961 case nir_op_i2f64:
962 result = LLVMBuildSIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
963 break;
964 case nir_op_u2f16:
965 case nir_op_u2f32:
966 case nir_op_u2f64:
967 result = LLVMBuildUIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
968 break;
969 case nir_op_f2f16_rtz:
970 case nir_op_f2f16:
971 case nir_op_f2fmp:
972 src[0] = ac_to_float(&ctx->ac, src[0]);
973
974 /* For OpenGL, we want fast packing with v_cvt_pkrtz_f16, but if we use it,
975 * all f32->f16 conversions have to round towards zero, because both scalar
976 * and vec2 down-conversions have to round equally.
977 */
978 if (ctx->ac.float_mode == AC_FLOAT_MODE_DEFAULT_OPENGL ||
979 instr->op == nir_op_f2f16_rtz) {
980 src[0] = ac_to_float(&ctx->ac, src[0]);
981
982 if (LLVMTypeOf(src[0]) == ctx->ac.f64)
983 src[0] = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ctx->ac.f32, "");
984
985 /* Fast path conversion. This only works if NIR is vectorized
986 * to vec2 16.
987 */
988 if (LLVMTypeOf(src[0]) == ctx->ac.v2f32) {
989 LLVMValueRef args[] = {
990 ac_llvm_extract_elem(&ctx->ac, src[0], 0),
991 ac_llvm_extract_elem(&ctx->ac, src[0], 1),
992 };
993 result = ac_build_cvt_pkrtz_f16(&ctx->ac, args);
994 break;
995 }
996
997 assert(ac_get_llvm_num_components(src[0]) == 1);
998 LLVMValueRef param[2] = { src[0], LLVMGetUndef(ctx->ac.f32) };
999 result = ac_build_cvt_pkrtz_f16(&ctx->ac, param);
1000 result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
1001 } else {
1002 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
1003 result = LLVMBuildFPExt(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
1004 else
1005 result = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
1006 }
1007 break;
1008 case nir_op_f2f16_rtne:
1009 case nir_op_f2f32:
1010 case nir_op_f2f64:
1011 src[0] = ac_to_float(&ctx->ac, src[0]);
1012 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
1013 result = LLVMBuildFPExt(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
1014 else
1015 result = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
1016 break;
1017 case nir_op_u2u8:
1018 case nir_op_u2u16:
1019 case nir_op_u2ump:
1020 case nir_op_u2u32:
1021 case nir_op_u2u64:
1022 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
1023 result = LLVMBuildZExt(ctx->ac.builder, src[0], def_type, "");
1024 else
1025 result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
1026 break;
1027 case nir_op_i2i8:
1028 case nir_op_i2i16:
1029 case nir_op_i2imp:
1030 case nir_op_i2i32:
1031 case nir_op_i2i64:
1032 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
1033 result = LLVMBuildSExt(ctx->ac.builder, src[0], def_type, "");
1034 else
1035 result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
1036 break;
1037 case nir_op_b32csel:
1038 result = emit_bcsel(&ctx->ac, src[0], src[1], src[2]);
1039 break;
1040 case nir_op_find_lsb:
1041 result = ac_find_lsb(&ctx->ac, ctx->ac.i32, src[0]);
1042 break;
1043 case nir_op_ufind_msb:
1044 result = ac_build_umsb(&ctx->ac, src[0], ctx->ac.i32);
1045 break;
1046 case nir_op_ifind_msb:
1047 result = ac_build_imsb(&ctx->ac, src[0], ctx->ac.i32);
1048 break;
1049 case nir_op_uadd_carry:
1050 result = emit_uint_carry(&ctx->ac, "llvm.uadd.with.overflow.i32", src[0], src[1]);
1051 break;
1052 case nir_op_usub_borrow:
1053 result = emit_uint_carry(&ctx->ac, "llvm.usub.with.overflow.i32", src[0], src[1]);
1054 break;
1055 case nir_op_b2f16:
1056 case nir_op_b2f32:
1057 case nir_op_b2f64:
1058 result = emit_b2f(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
1059 break;
1060 case nir_op_f2b32:
1061 result = emit_f2b(&ctx->ac, src[0]);
1062 break;
1063 case nir_op_b2i8:
1064 case nir_op_b2i16:
1065 case nir_op_b2i32:
1066 case nir_op_b2i64:
1067 result = emit_b2i(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
1068 break;
1069 case nir_op_i2b32:
1070 result = emit_i2b(&ctx->ac, src[0]);
1071 break;
1072 case nir_op_fquantize2f16:
1073 result = emit_f2f16(&ctx->ac, src[0]);
1074 break;
1075 case nir_op_umul_high:
1076 result = emit_umul_high(&ctx->ac, src[0], src[1]);
1077 break;
1078 case nir_op_imul_high:
1079 result = emit_imul_high(&ctx->ac, src[0], src[1]);
1080 break;
1081 case nir_op_pack_half_2x16:
1082 result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pkrtz_f16);
1083 break;
1084 case nir_op_pack_snorm_2x16:
1085 result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pknorm_i16);
1086 break;
1087 case nir_op_pack_unorm_2x16:
1088 result = emit_pack_2x16(&ctx->ac, src[0], ac_build_cvt_pknorm_u16);
1089 break;
1090 case nir_op_unpack_half_2x16:
1091 result = emit_unpack_half_2x16(&ctx->ac, src[0]);
1092 break;
1093 case nir_op_fddx:
1094 case nir_op_fddy:
1095 case nir_op_fddx_fine:
1096 case nir_op_fddy_fine:
1097 case nir_op_fddx_coarse:
1098 case nir_op_fddy_coarse:
1099 result = emit_ddxy(ctx, instr->op, src[0]);
1100 break;
1101
1102 case nir_op_unpack_64_2x32_split_x: {
1103 assert(ac_get_llvm_num_components(src[0]) == 1);
1104 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1105 ctx->ac.v2i32,
1106 "");
1107 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1108 ctx->ac.i32_0, "");
1109 break;
1110 }
1111
1112 case nir_op_unpack_64_2x32_split_y: {
1113 assert(ac_get_llvm_num_components(src[0]) == 1);
1114 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1115 ctx->ac.v2i32,
1116 "");
1117 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1118 ctx->ac.i32_1, "");
1119 break;
1120 }
1121
1122 case nir_op_pack_64_2x32_split: {
1123 LLVMValueRef tmp = ac_build_gather_values(&ctx->ac, src, 2);
1124 result = LLVMBuildBitCast(ctx->ac.builder, tmp, ctx->ac.i64, "");
1125 break;
1126 }
1127
1128 case nir_op_pack_32_2x16_split: {
1129 LLVMValueRef tmp = ac_build_gather_values(&ctx->ac, src, 2);
1130 result = LLVMBuildBitCast(ctx->ac.builder, tmp, ctx->ac.i32, "");
1131 break;
1132 }
1133
1134 case nir_op_unpack_32_2x16_split_x: {
1135 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1136 ctx->ac.v2i16,
1137 "");
1138 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1139 ctx->ac.i32_0, "");
1140 break;
1141 }
1142
1143 case nir_op_unpack_32_2x16_split_y: {
1144 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1145 ctx->ac.v2i16,
1146 "");
1147 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1148 ctx->ac.i32_1, "");
1149 break;
1150 }
1151
1152 case nir_op_cube_face_coord: {
1153 src[0] = ac_to_float(&ctx->ac, src[0]);
1154 LLVMValueRef results[2];
1155 LLVMValueRef in[3];
1156 for (unsigned chan = 0; chan < 3; chan++)
1157 in[chan] = ac_llvm_extract_elem(&ctx->ac, src[0], chan);
1158 results[0] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubesc",
1159 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1160 results[1] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubetc",
1161 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1162 LLVMValueRef ma = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubema",
1163 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1164 results[0] = ac_build_fdiv(&ctx->ac, results[0], ma);
1165 results[1] = ac_build_fdiv(&ctx->ac, results[1], ma);
1166 LLVMValueRef offset = LLVMConstReal(ctx->ac.f32, 0.5);
1167 results[0] = LLVMBuildFAdd(ctx->ac.builder, results[0], offset, "");
1168 results[1] = LLVMBuildFAdd(ctx->ac.builder, results[1], offset, "");
1169 result = ac_build_gather_values(&ctx->ac, results, 2);
1170 break;
1171 }
1172
1173 case nir_op_cube_face_index: {
1174 src[0] = ac_to_float(&ctx->ac, src[0]);
1175 LLVMValueRef in[3];
1176 for (unsigned chan = 0; chan < 3; chan++)
1177 in[chan] = ac_llvm_extract_elem(&ctx->ac, src[0], chan);
1178 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubeid",
1179 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1180 break;
1181 }
1182
1183 default:
1184 fprintf(stderr, "Unknown NIR alu instr: ");
1185 nir_print_instr(&instr->instr, stderr);
1186 fprintf(stderr, "\n");
1187 abort();
1188 }
1189
1190 if (result) {
1191 assert(instr->dest.dest.is_ssa);
1192 result = ac_to_integer_or_pointer(&ctx->ac, result);
1193 ctx->ssa_defs[instr->dest.dest.ssa.index] = result;
1194 }
1195
1196 if (instr->exact)
1197 ac_restore_inexact_math(ctx->ac.builder, saved_inexact);
1198 }
1199
1200 static void visit_load_const(struct ac_nir_context *ctx,
1201 const nir_load_const_instr *instr)
1202 {
1203 LLVMValueRef values[4], value = NULL;
1204 LLVMTypeRef element_type =
1205 LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
1206
1207 for (unsigned i = 0; i < instr->def.num_components; ++i) {
1208 switch (instr->def.bit_size) {
1209 case 8:
1210 values[i] = LLVMConstInt(element_type,
1211 instr->value[i].u8, false);
1212 break;
1213 case 16:
1214 values[i] = LLVMConstInt(element_type,
1215 instr->value[i].u16, false);
1216 break;
1217 case 32:
1218 values[i] = LLVMConstInt(element_type,
1219 instr->value[i].u32, false);
1220 break;
1221 case 64:
1222 values[i] = LLVMConstInt(element_type,
1223 instr->value[i].u64, false);
1224 break;
1225 default:
1226 fprintf(stderr,
1227 "unsupported nir load_const bit_size: %d\n",
1228 instr->def.bit_size);
1229 abort();
1230 }
1231 }
1232 if (instr->def.num_components > 1) {
1233 value = LLVMConstVector(values, instr->def.num_components);
1234 } else
1235 value = values[0];
1236
1237 ctx->ssa_defs[instr->def.index] = value;
1238 }
1239
1240 static LLVMValueRef
1241 get_buffer_size(struct ac_nir_context *ctx, LLVMValueRef descriptor, bool in_elements)
1242 {
1243 LLVMValueRef size =
1244 LLVMBuildExtractElement(ctx->ac.builder, descriptor,
1245 LLVMConstInt(ctx->ac.i32, 2, false), "");
1246
1247 /* GFX8 only */
1248 if (ctx->ac.chip_class == GFX8 && in_elements) {
1249 /* On GFX8, the descriptor contains the size in bytes,
1250 * but TXQ must return the size in elements.
1251 * The stride is always non-zero for resources using TXQ.
1252 */
1253 LLVMValueRef stride =
1254 LLVMBuildExtractElement(ctx->ac.builder, descriptor,
1255 ctx->ac.i32_1, "");
1256 stride = LLVMBuildLShr(ctx->ac.builder, stride,
1257 LLVMConstInt(ctx->ac.i32, 16, false), "");
1258 stride = LLVMBuildAnd(ctx->ac.builder, stride,
1259 LLVMConstInt(ctx->ac.i32, 0x3fff, false), "");
1260
1261 size = LLVMBuildUDiv(ctx->ac.builder, size, stride, "");
1262 }
1263 return size;
1264 }
1265
1266 /* Gather4 should follow the same rules as bilinear filtering, but the hardware
1267 * incorrectly forces nearest filtering if the texture format is integer.
1268 * The only effect it has on Gather4, which always returns 4 texels for
1269 * bilinear filtering, is that the final coordinates are off by 0.5 of
1270 * the texel size.
1271 *
1272 * The workaround is to subtract 0.5 from the unnormalized coordinates,
1273 * or (0.5 / size) from the normalized coordinates.
1274 *
1275 * However, cube textures with 8_8_8_8 data formats require a different
1276 * workaround of overriding the num format to USCALED/SSCALED. This would lose
1277 * precision in 32-bit data formats, so it needs to be applied dynamically at
1278 * runtime. In this case, return an i1 value that indicates whether the
1279 * descriptor was overridden (and hence a fixup of the sampler result is needed).
1280 */
1281 static LLVMValueRef lower_gather4_integer(struct ac_llvm_context *ctx,
1282 nir_variable *var,
1283 struct ac_image_args *args,
1284 const nir_tex_instr *instr)
1285 {
1286 const struct glsl_type *type = glsl_without_array(var->type);
1287 enum glsl_base_type stype = glsl_get_sampler_result_type(type);
1288 LLVMValueRef wa_8888 = NULL;
1289 LLVMValueRef half_texel[2];
1290 LLVMValueRef result;
1291
1292 assert(stype == GLSL_TYPE_INT || stype == GLSL_TYPE_UINT);
1293
1294 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
1295 LLVMValueRef formats;
1296 LLVMValueRef data_format;
1297 LLVMValueRef wa_formats;
1298
1299 formats = LLVMBuildExtractElement(ctx->builder, args->resource, ctx->i32_1, "");
1300
1301 data_format = LLVMBuildLShr(ctx->builder, formats,
1302 LLVMConstInt(ctx->i32, 20, false), "");
1303 data_format = LLVMBuildAnd(ctx->builder, data_format,
1304 LLVMConstInt(ctx->i32, (1u << 6) - 1, false), "");
1305 wa_8888 = LLVMBuildICmp(
1306 ctx->builder, LLVMIntEQ, data_format,
1307 LLVMConstInt(ctx->i32, V_008F14_IMG_DATA_FORMAT_8_8_8_8, false),
1308 "");
1309
1310 uint32_t wa_num_format =
1311 stype == GLSL_TYPE_UINT ?
1312 S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_USCALED) :
1313 S_008F14_NUM_FORMAT(V_008F14_IMG_NUM_FORMAT_SSCALED);
1314 wa_formats = LLVMBuildAnd(ctx->builder, formats,
1315 LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT, false),
1316 "");
1317 wa_formats = LLVMBuildOr(ctx->builder, wa_formats,
1318 LLVMConstInt(ctx->i32, wa_num_format, false), "");
1319
1320 formats = LLVMBuildSelect(ctx->builder, wa_8888, wa_formats, formats, "");
1321 args->resource = LLVMBuildInsertElement(
1322 ctx->builder, args->resource, formats, ctx->i32_1, "");
1323 }
1324
1325 if (instr->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
1326 assert(!wa_8888);
1327 half_texel[0] = half_texel[1] = LLVMConstReal(ctx->f32, -0.5);
1328 } else {
1329 struct ac_image_args resinfo = {};
1330 LLVMBasicBlockRef bbs[2];
1331
1332 LLVMValueRef unnorm = NULL;
1333 LLVMValueRef default_offset = ctx->f32_0;
1334 if (instr->sampler_dim == GLSL_SAMPLER_DIM_2D &&
1335 !instr->is_array) {
1336 /* In vulkan, whether the sampler uses unnormalized
1337 * coordinates or not is a dynamic property of the
1338 * sampler. Hence, to figure out whether or not we
1339 * need to divide by the texture size, we need to test
1340 * the sampler at runtime. This tests the bit set by
1341 * radv_init_sampler().
1342 */
1343 LLVMValueRef sampler0 =
1344 LLVMBuildExtractElement(ctx->builder, args->sampler, ctx->i32_0, "");
1345 sampler0 = LLVMBuildLShr(ctx->builder, sampler0,
1346 LLVMConstInt(ctx->i32, 15, false), "");
1347 sampler0 = LLVMBuildAnd(ctx->builder, sampler0, ctx->i32_1, "");
1348 unnorm = LLVMBuildICmp(ctx->builder, LLVMIntEQ, sampler0, ctx->i32_1, "");
1349 default_offset = LLVMConstReal(ctx->f32, -0.5);
1350 }
1351
1352 bbs[0] = LLVMGetInsertBlock(ctx->builder);
1353 if (wa_8888 || unnorm) {
1354 assert(!(wa_8888 && unnorm));
1355 LLVMValueRef not_needed = wa_8888 ? wa_8888 : unnorm;
1356 /* Skip the texture size query entirely if we don't need it. */
1357 ac_build_ifcc(ctx, LLVMBuildNot(ctx->builder, not_needed, ""), 2000);
1358 bbs[1] = LLVMGetInsertBlock(ctx->builder);
1359 }
1360
1361 /* Query the texture size. */
1362 resinfo.dim = ac_get_sampler_dim(ctx->chip_class, instr->sampler_dim, instr->is_array);
1363 resinfo.opcode = ac_image_get_resinfo;
1364 resinfo.dmask = 0xf;
1365 resinfo.lod = ctx->i32_0;
1366 resinfo.resource = args->resource;
1367 resinfo.attributes = AC_FUNC_ATTR_READNONE;
1368 LLVMValueRef size = ac_build_image_opcode(ctx, &resinfo);
1369
1370 /* Compute -0.5 / size. */
1371 for (unsigned c = 0; c < 2; c++) {
1372 half_texel[c] =
1373 LLVMBuildExtractElement(ctx->builder, size,
1374 LLVMConstInt(ctx->i32, c, 0), "");
1375 half_texel[c] = LLVMBuildUIToFP(ctx->builder, half_texel[c], ctx->f32, "");
1376 half_texel[c] = ac_build_fdiv(ctx, ctx->f32_1, half_texel[c]);
1377 half_texel[c] = LLVMBuildFMul(ctx->builder, half_texel[c],
1378 LLVMConstReal(ctx->f32, -0.5), "");
1379 }
1380
1381 if (wa_8888 || unnorm) {
1382 ac_build_endif(ctx, 2000);
1383
1384 for (unsigned c = 0; c < 2; c++) {
1385 LLVMValueRef values[2] = { default_offset, half_texel[c] };
1386 half_texel[c] = ac_build_phi(ctx, ctx->f32, 2,
1387 values, bbs);
1388 }
1389 }
1390 }
1391
1392 for (unsigned c = 0; c < 2; c++) {
1393 LLVMValueRef tmp;
1394 tmp = LLVMBuildBitCast(ctx->builder, args->coords[c], ctx->f32, "");
1395 args->coords[c] = LLVMBuildFAdd(ctx->builder, tmp, half_texel[c], "");
1396 }
1397
1398 args->attributes = AC_FUNC_ATTR_READNONE;
1399 result = ac_build_image_opcode(ctx, args);
1400
1401 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
1402 LLVMValueRef tmp, tmp2;
1403
1404 /* if the cube workaround is in place, f2i the result. */
1405 for (unsigned c = 0; c < 4; c++) {
1406 tmp = LLVMBuildExtractElement(ctx->builder, result, LLVMConstInt(ctx->i32, c, false), "");
1407 if (stype == GLSL_TYPE_UINT)
1408 tmp2 = LLVMBuildFPToUI(ctx->builder, tmp, ctx->i32, "");
1409 else
1410 tmp2 = LLVMBuildFPToSI(ctx->builder, tmp, ctx->i32, "");
1411 tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->i32, "");
1412 tmp2 = LLVMBuildBitCast(ctx->builder, tmp2, ctx->i32, "");
1413 tmp = LLVMBuildSelect(ctx->builder, wa_8888, tmp2, tmp, "");
1414 tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->f32, "");
1415 result = LLVMBuildInsertElement(ctx->builder, result, tmp, LLVMConstInt(ctx->i32, c, false), "");
1416 }
1417 }
1418 return result;
1419 }
1420
1421 static nir_deref_instr *get_tex_texture_deref(const nir_tex_instr *instr)
1422 {
1423 nir_deref_instr *texture_deref_instr = NULL;
1424
1425 for (unsigned i = 0; i < instr->num_srcs; i++) {
1426 switch (instr->src[i].src_type) {
1427 case nir_tex_src_texture_deref:
1428 texture_deref_instr = nir_src_as_deref(instr->src[i].src);
1429 break;
1430 default:
1431 break;
1432 }
1433 }
1434 return texture_deref_instr;
1435 }
1436
1437 static LLVMValueRef build_tex_intrinsic(struct ac_nir_context *ctx,
1438 const nir_tex_instr *instr,
1439 struct ac_image_args *args)
1440 {
1441 if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
1442 unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
1443
1444 assert(instr->dest.is_ssa);
1445 return ac_build_buffer_load_format(&ctx->ac,
1446 args->resource,
1447 args->coords[0],
1448 ctx->ac.i32_0,
1449 util_last_bit(mask),
1450 0, true,
1451 instr->dest.ssa.bit_size == 16);
1452 }
1453
1454 args->opcode = ac_image_sample;
1455
1456 switch (instr->op) {
1457 case nir_texop_txf:
1458 case nir_texop_txf_ms:
1459 case nir_texop_samples_identical:
1460 args->opcode = args->level_zero ||
1461 instr->sampler_dim == GLSL_SAMPLER_DIM_MS ?
1462 ac_image_load : ac_image_load_mip;
1463 args->level_zero = false;
1464 break;
1465 case nir_texop_txs:
1466 case nir_texop_query_levels:
1467 args->opcode = ac_image_get_resinfo;
1468 if (!args->lod)
1469 args->lod = ctx->ac.i32_0;
1470 args->level_zero = false;
1471 break;
1472 case nir_texop_tex:
1473 if (ctx->stage != MESA_SHADER_FRAGMENT) {
1474 assert(!args->lod);
1475 args->level_zero = true;
1476 }
1477 break;
1478 case nir_texop_tg4:
1479 args->opcode = ac_image_gather4;
1480 if (!args->lod && !args->bias)
1481 args->level_zero = true;
1482 break;
1483 case nir_texop_lod:
1484 args->opcode = ac_image_get_lod;
1485 break;
1486 case nir_texop_fragment_fetch:
1487 case nir_texop_fragment_mask_fetch:
1488 args->opcode = ac_image_load;
1489 args->level_zero = false;
1490 break;
1491 default:
1492 break;
1493 }
1494
1495 if (instr->op == nir_texop_tg4 && ctx->ac.chip_class <= GFX8) {
1496 nir_deref_instr *texture_deref_instr = get_tex_texture_deref(instr);
1497 nir_variable *var = nir_deref_instr_get_variable(texture_deref_instr);
1498 const struct glsl_type *type = glsl_without_array(var->type);
1499 enum glsl_base_type stype = glsl_get_sampler_result_type(type);
1500 if (stype == GLSL_TYPE_UINT || stype == GLSL_TYPE_INT) {
1501 return lower_gather4_integer(&ctx->ac, var, args, instr);
1502 }
1503 }
1504
1505 /* Fixup for GFX9 which allocates 1D textures as 2D. */
1506 if (instr->op == nir_texop_lod && ctx->ac.chip_class == GFX9) {
1507 if ((args->dim == ac_image_2darray ||
1508 args->dim == ac_image_2d) && !args->coords[1]) {
1509 args->coords[1] = ctx->ac.i32_0;
1510 }
1511 }
1512
1513 args->attributes = AC_FUNC_ATTR_READNONE;
1514 bool cs_derivs = ctx->stage == MESA_SHADER_COMPUTE &&
1515 ctx->info->cs.derivative_group != DERIVATIVE_GROUP_NONE;
1516 if (ctx->stage == MESA_SHADER_FRAGMENT || cs_derivs) {
1517 /* Prevent texture instructions with implicit derivatives from being
1518 * sinked into branches. */
1519 switch (instr->op) {
1520 case nir_texop_tex:
1521 case nir_texop_txb:
1522 case nir_texop_lod:
1523 args->attributes |= AC_FUNC_ATTR_CONVERGENT;
1524 break;
1525 default:
1526 break;
1527 }
1528 }
1529
1530 return ac_build_image_opcode(&ctx->ac, args);
1531 }
1532
1533 static LLVMValueRef visit_vulkan_resource_reindex(struct ac_nir_context *ctx,
1534 nir_intrinsic_instr *instr)
1535 {
1536 LLVMValueRef ptr = get_src(ctx, instr->src[0]);
1537 LLVMValueRef index = get_src(ctx, instr->src[1]);
1538
1539 LLVMValueRef result = LLVMBuildGEP(ctx->ac.builder, ptr, &index, 1, "");
1540 LLVMSetMetadata(result, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
1541 return result;
1542 }
1543
1544 static LLVMValueRef visit_load_push_constant(struct ac_nir_context *ctx,
1545 nir_intrinsic_instr *instr)
1546 {
1547 LLVMValueRef ptr, addr;
1548 LLVMValueRef src0 = get_src(ctx, instr->src[0]);
1549 unsigned index = nir_intrinsic_base(instr);
1550
1551 addr = LLVMConstInt(ctx->ac.i32, index, 0);
1552 addr = LLVMBuildAdd(ctx->ac.builder, addr, src0, "");
1553
1554 /* Load constant values from user SGPRS when possible, otherwise
1555 * fallback to the default path that loads directly from memory.
1556 */
1557 if (LLVMIsConstant(src0) &&
1558 instr->dest.ssa.bit_size == 32) {
1559 unsigned count = instr->dest.ssa.num_components;
1560 unsigned offset = index;
1561
1562 offset += LLVMConstIntGetZExtValue(src0);
1563 offset /= 4;
1564
1565 offset -= ctx->args->base_inline_push_consts;
1566
1567 unsigned num_inline_push_consts = ctx->args->num_inline_push_consts;
1568 if (offset + count <= num_inline_push_consts) {
1569 LLVMValueRef push_constants[num_inline_push_consts];
1570 for (unsigned i = 0; i < num_inline_push_consts; i++)
1571 push_constants[i] = ac_get_arg(&ctx->ac,
1572 ctx->args->inline_push_consts[i]);
1573 return ac_build_gather_values(&ctx->ac,
1574 push_constants + offset,
1575 count);
1576 }
1577 }
1578
1579 ptr = LLVMBuildGEP(ctx->ac.builder,
1580 ac_get_arg(&ctx->ac, ctx->args->push_constants), &addr, 1, "");
1581
1582 if (instr->dest.ssa.bit_size == 8) {
1583 unsigned load_dwords = instr->dest.ssa.num_components > 1 ? 2 : 1;
1584 LLVMTypeRef vec_type = LLVMVectorType(ctx->ac.i8, 4 * load_dwords);
1585 ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
1586 LLVMValueRef res = LLVMBuildLoad(ctx->ac.builder, ptr, "");
1587
1588 LLVMValueRef params[3];
1589 if (load_dwords > 1) {
1590 LLVMValueRef res_vec = LLVMBuildBitCast(ctx->ac.builder, res, ctx->ac.v2i32, "");
1591 params[0] = LLVMBuildExtractElement(ctx->ac.builder, res_vec, LLVMConstInt(ctx->ac.i32, 1, false), "");
1592 params[1] = LLVMBuildExtractElement(ctx->ac.builder, res_vec, LLVMConstInt(ctx->ac.i32, 0, false), "");
1593 } else {
1594 res = LLVMBuildBitCast(ctx->ac.builder, res, ctx->ac.i32, "");
1595 params[0] = ctx->ac.i32_0;
1596 params[1] = res;
1597 }
1598 params[2] = addr;
1599 res = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.alignbyte", ctx->ac.i32, params, 3, 0);
1600
1601 res = LLVMBuildTrunc(ctx->ac.builder, res, LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.num_components * 8), "");
1602 if (instr->dest.ssa.num_components > 1)
1603 res = LLVMBuildBitCast(ctx->ac.builder, res, LLVMVectorType(ctx->ac.i8, instr->dest.ssa.num_components), "");
1604 return res;
1605 } else if (instr->dest.ssa.bit_size == 16) {
1606 unsigned load_dwords = instr->dest.ssa.num_components / 2 + 1;
1607 LLVMTypeRef vec_type = LLVMVectorType(ctx->ac.i16, 2 * load_dwords);
1608 ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
1609 LLVMValueRef res = LLVMBuildLoad(ctx->ac.builder, ptr, "");
1610 res = LLVMBuildBitCast(ctx->ac.builder, res, vec_type, "");
1611 LLVMValueRef cond = LLVMBuildLShr(ctx->ac.builder, addr, ctx->ac.i32_1, "");
1612 cond = LLVMBuildTrunc(ctx->ac.builder, cond, ctx->ac.i1, "");
1613 LLVMValueRef mask[] = { LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
1614 LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
1615 LLVMConstInt(ctx->ac.i32, 4, false)};
1616 LLVMValueRef swizzle_aligned = LLVMConstVector(&mask[0], instr->dest.ssa.num_components);
1617 LLVMValueRef swizzle_unaligned = LLVMConstVector(&mask[1], instr->dest.ssa.num_components);
1618 LLVMValueRef shuffle_aligned = LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_aligned, "");
1619 LLVMValueRef shuffle_unaligned = LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_unaligned, "");
1620 res = LLVMBuildSelect(ctx->ac.builder, cond, shuffle_unaligned, shuffle_aligned, "");
1621 return LLVMBuildBitCast(ctx->ac.builder, res, get_def_type(ctx, &instr->dest.ssa), "");
1622 }
1623
1624 ptr = ac_cast_ptr(&ctx->ac, ptr, get_def_type(ctx, &instr->dest.ssa));
1625
1626 return LLVMBuildLoad(ctx->ac.builder, ptr, "");
1627 }
1628
1629 static LLVMValueRef visit_get_buffer_size(struct ac_nir_context *ctx,
1630 const nir_intrinsic_instr *instr)
1631 {
1632 LLVMValueRef index = get_src(ctx, instr->src[0]);
1633
1634 return get_buffer_size(ctx, ctx->abi->load_ssbo(ctx->abi, index, false), false);
1635 }
1636
1637 static uint32_t widen_mask(uint32_t mask, unsigned multiplier)
1638 {
1639 uint32_t new_mask = 0;
1640 for(unsigned i = 0; i < 32 && (1u << i) <= mask; ++i)
1641 if (mask & (1u << i))
1642 new_mask |= ((1u << multiplier) - 1u) << (i * multiplier);
1643 return new_mask;
1644 }
1645
1646 static LLVMValueRef extract_vector_range(struct ac_llvm_context *ctx, LLVMValueRef src,
1647 unsigned start, unsigned count)
1648 {
1649 LLVMValueRef mask[] = {
1650 ctx->i32_0, ctx->i32_1,
1651 LLVMConstInt(ctx->i32, 2, false), LLVMConstInt(ctx->i32, 3, false) };
1652
1653 unsigned src_elements = ac_get_llvm_num_components(src);
1654
1655 if (count == src_elements) {
1656 assert(start == 0);
1657 return src;
1658 } else if (count == 1) {
1659 assert(start < src_elements);
1660 return LLVMBuildExtractElement(ctx->builder, src, mask[start], "");
1661 } else {
1662 assert(start + count <= src_elements);
1663 assert(count <= 4);
1664 LLVMValueRef swizzle = LLVMConstVector(&mask[start], count);
1665 return LLVMBuildShuffleVector(ctx->builder, src, src, swizzle, "");
1666 }
1667 }
1668
1669 static unsigned get_cache_policy(struct ac_nir_context *ctx,
1670 enum gl_access_qualifier access,
1671 bool may_store_unaligned,
1672 bool writeonly_memory)
1673 {
1674 unsigned cache_policy = 0;
1675
1676 /* GFX6 has a TC L1 bug causing corruption of 8bit/16bit stores. All
1677 * store opcodes not aligned to a dword are affected. The only way to
1678 * get unaligned stores is through shader images.
1679 */
1680 if (((may_store_unaligned && ctx->ac.chip_class == GFX6) ||
1681 /* If this is write-only, don't keep data in L1 to prevent
1682 * evicting L1 cache lines that may be needed by other
1683 * instructions.
1684 */
1685 writeonly_memory ||
1686 access & (ACCESS_COHERENT | ACCESS_VOLATILE))) {
1687 cache_policy |= ac_glc;
1688 }
1689
1690 if (access & ACCESS_STREAM_CACHE_POLICY)
1691 cache_policy |= ac_slc | ac_glc;
1692
1693 return cache_policy;
1694 }
1695
1696 static LLVMValueRef enter_waterfall_ssbo(struct ac_nir_context *ctx,
1697 struct waterfall_context *wctx,
1698 const nir_intrinsic_instr *instr,
1699 nir_src src)
1700 {
1701 return enter_waterfall(ctx, wctx, get_src(ctx, src),
1702 nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
1703 }
1704
1705 static void visit_store_ssbo(struct ac_nir_context *ctx,
1706 nir_intrinsic_instr *instr)
1707 {
1708 if (ctx->ac.postponed_kill) {
1709 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
1710 ctx->ac.postponed_kill, "");
1711 ac_build_ifcc(&ctx->ac, cond, 7000);
1712 }
1713
1714 LLVMValueRef src_data = get_src(ctx, instr->src[0]);
1715 int elem_size_bytes = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src_data)) / 8;
1716 unsigned writemask = nir_intrinsic_write_mask(instr);
1717 enum gl_access_qualifier access = nir_intrinsic_access(instr);
1718 bool writeonly_memory = access & ACCESS_NON_READABLE;
1719 unsigned cache_policy = get_cache_policy(ctx, access, false, writeonly_memory);
1720
1721 struct waterfall_context wctx;
1722 LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[1]);
1723
1724 LLVMValueRef rsrc = ctx->abi->load_ssbo(ctx->abi, rsrc_base, true);
1725 LLVMValueRef base_data = src_data;
1726 base_data = ac_trim_vector(&ctx->ac, base_data, instr->num_components);
1727 LLVMValueRef base_offset = get_src(ctx, instr->src[2]);
1728
1729 while (writemask) {
1730 int start, count;
1731 LLVMValueRef data, offset;
1732 LLVMTypeRef data_type;
1733
1734 u_bit_scan_consecutive_range(&writemask, &start, &count);
1735
1736 /* Due to an LLVM limitation with LLVM < 9, split 3-element
1737 * writes into a 2-element and a 1-element write. */
1738 if (count == 3 &&
1739 (elem_size_bytes != 4 || !ac_has_vec3_support(ctx->ac.chip_class, false))) {
1740 writemask |= 1 << (start + 2);
1741 count = 2;
1742 }
1743 int num_bytes = count * elem_size_bytes; /* count in bytes */
1744
1745 /* we can only store 4 DWords at the same time.
1746 * can only happen for 64 Bit vectors. */
1747 if (num_bytes > 16) {
1748 writemask |= ((1u << (count - 2)) - 1u) << (start + 2);
1749 count = 2;
1750 num_bytes = 16;
1751 }
1752
1753 /* check alignment of 16 Bit stores */
1754 if (elem_size_bytes == 2 && num_bytes > 2 && (start % 2) == 1) {
1755 writemask |= ((1u << (count - 1)) - 1u) << (start + 1);
1756 count = 1;
1757 num_bytes = 2;
1758 }
1759
1760 /* Due to alignment issues, split stores of 8-bit/16-bit
1761 * vectors.
1762 */
1763 if (ctx->ac.chip_class == GFX6 && count > 1 && elem_size_bytes < 4) {
1764 writemask |= ((1u << (count - 1)) - 1u) << (start + 1);
1765 count = 1;
1766 num_bytes = elem_size_bytes;
1767 }
1768
1769 data = extract_vector_range(&ctx->ac, base_data, start, count);
1770
1771 offset = LLVMBuildAdd(ctx->ac.builder, base_offset,
1772 LLVMConstInt(ctx->ac.i32, start * elem_size_bytes, false), "");
1773
1774 if (num_bytes == 1) {
1775 ac_build_tbuffer_store_byte(&ctx->ac, rsrc, data,
1776 offset, ctx->ac.i32_0,
1777 cache_policy);
1778 } else if (num_bytes == 2) {
1779 ac_build_tbuffer_store_short(&ctx->ac, rsrc, data,
1780 offset, ctx->ac.i32_0,
1781 cache_policy);
1782 } else {
1783 int num_channels = num_bytes / 4;
1784
1785 switch (num_bytes) {
1786 case 16: /* v4f32 */
1787 data_type = ctx->ac.v4f32;
1788 break;
1789 case 12: /* v3f32 */
1790 data_type = ctx->ac.v3f32;
1791 break;
1792 case 8: /* v2f32 */
1793 data_type = ctx->ac.v2f32;
1794 break;
1795 case 4: /* f32 */
1796 data_type = ctx->ac.f32;
1797 break;
1798 default:
1799 unreachable("Malformed vector store.");
1800 }
1801 data = LLVMBuildBitCast(ctx->ac.builder, data, data_type, "");
1802
1803 ac_build_buffer_store_dword(&ctx->ac, rsrc, data,
1804 num_channels, offset,
1805 ctx->ac.i32_0, 0,
1806 cache_policy);
1807 }
1808 }
1809
1810 exit_waterfall(ctx, &wctx, NULL);
1811
1812 if (ctx->ac.postponed_kill)
1813 ac_build_endif(&ctx->ac, 7000);
1814 }
1815
1816 static LLVMValueRef emit_ssbo_comp_swap_64(struct ac_nir_context *ctx,
1817 LLVMValueRef descriptor,
1818 LLVMValueRef offset,
1819 LLVMValueRef compare,
1820 LLVMValueRef exchange)
1821 {
1822 LLVMBasicBlockRef start_block = NULL, then_block = NULL;
1823 if (ctx->abi->robust_buffer_access) {
1824 LLVMValueRef size = ac_llvm_extract_elem(&ctx->ac, descriptor, 2);
1825
1826 LLVMValueRef cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntULT, offset, size, "");
1827 start_block = LLVMGetInsertBlock(ctx->ac.builder);
1828
1829 ac_build_ifcc(&ctx->ac, cond, -1);
1830
1831 then_block = LLVMGetInsertBlock(ctx->ac.builder);
1832 }
1833
1834 LLVMValueRef ptr_parts[2] = {
1835 ac_llvm_extract_elem(&ctx->ac, descriptor, 0),
1836 LLVMBuildAnd(ctx->ac.builder,
1837 ac_llvm_extract_elem(&ctx->ac, descriptor, 1),
1838 LLVMConstInt(ctx->ac.i32, 65535, 0), "")
1839 };
1840
1841 ptr_parts[1] = LLVMBuildTrunc(ctx->ac.builder, ptr_parts[1], ctx->ac.i16, "");
1842 ptr_parts[1] = LLVMBuildSExt(ctx->ac.builder, ptr_parts[1], ctx->ac.i32, "");
1843
1844 offset = LLVMBuildZExt(ctx->ac.builder, offset, ctx->ac.i64, "");
1845
1846 LLVMValueRef ptr = ac_build_gather_values(&ctx->ac, ptr_parts, 2);
1847 ptr = LLVMBuildBitCast(ctx->ac.builder, ptr, ctx->ac.i64, "");
1848 ptr = LLVMBuildAdd(ctx->ac.builder, ptr, offset, "");
1849 ptr = LLVMBuildIntToPtr(ctx->ac.builder, ptr, LLVMPointerType(ctx->ac.i64, AC_ADDR_SPACE_GLOBAL), "");
1850
1851 LLVMValueRef result = ac_build_atomic_cmp_xchg(&ctx->ac, ptr, compare, exchange, "singlethread-one-as");
1852 result = LLVMBuildExtractValue(ctx->ac.builder, result, 0, "");
1853
1854 if (ctx->abi->robust_buffer_access) {
1855 ac_build_endif(&ctx->ac, -1);
1856
1857 LLVMBasicBlockRef incoming_blocks[2] = {
1858 start_block,
1859 then_block,
1860 };
1861
1862 LLVMValueRef incoming_values[2] = {
1863 LLVMConstInt(ctx->ac.i64, 0, 0),
1864 result,
1865 };
1866 LLVMValueRef ret = LLVMBuildPhi(ctx->ac.builder, ctx->ac.i64, "");
1867 LLVMAddIncoming(ret, incoming_values, incoming_blocks, 2);
1868 return ret;
1869 } else {
1870 return result;
1871 }
1872 }
1873
1874 static LLVMValueRef visit_atomic_ssbo(struct ac_nir_context *ctx,
1875 nir_intrinsic_instr *instr)
1876 {
1877 if (ctx->ac.postponed_kill) {
1878 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
1879 ctx->ac.postponed_kill, "");
1880 ac_build_ifcc(&ctx->ac, cond, 7001);
1881 }
1882
1883 LLVMTypeRef return_type = LLVMTypeOf(get_src(ctx, instr->src[2]));
1884 const char *op;
1885 char name[64], type[8];
1886 LLVMValueRef params[6], descriptor;
1887 LLVMValueRef result;
1888 int arg_count = 0;
1889
1890 struct waterfall_context wctx;
1891 LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[0]);
1892
1893 switch (instr->intrinsic) {
1894 case nir_intrinsic_ssbo_atomic_add:
1895 op = "add";
1896 break;
1897 case nir_intrinsic_ssbo_atomic_imin:
1898 op = "smin";
1899 break;
1900 case nir_intrinsic_ssbo_atomic_umin:
1901 op = "umin";
1902 break;
1903 case nir_intrinsic_ssbo_atomic_imax:
1904 op = "smax";
1905 break;
1906 case nir_intrinsic_ssbo_atomic_umax:
1907 op = "umax";
1908 break;
1909 case nir_intrinsic_ssbo_atomic_and:
1910 op = "and";
1911 break;
1912 case nir_intrinsic_ssbo_atomic_or:
1913 op = "or";
1914 break;
1915 case nir_intrinsic_ssbo_atomic_xor:
1916 op = "xor";
1917 break;
1918 case nir_intrinsic_ssbo_atomic_exchange:
1919 op = "swap";
1920 break;
1921 case nir_intrinsic_ssbo_atomic_comp_swap:
1922 op = "cmpswap";
1923 break;
1924 default:
1925 abort();
1926 }
1927
1928 descriptor = ctx->abi->load_ssbo(ctx->abi,
1929 rsrc_base,
1930 true);
1931
1932 if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap &&
1933 return_type == ctx->ac.i64) {
1934 result = emit_ssbo_comp_swap_64(ctx, descriptor,
1935 get_src(ctx, instr->src[1]),
1936 get_src(ctx, instr->src[2]),
1937 get_src(ctx, instr->src[3]));
1938 } else {
1939 if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap) {
1940 params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[3]), 0);
1941 }
1942 params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
1943 params[arg_count++] = descriptor;
1944
1945 if (LLVM_VERSION_MAJOR >= 9) {
1946 /* XXX: The new raw/struct atomic intrinsics are buggy with
1947 * LLVM 8, see r358579.
1948 */
1949 params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
1950 params[arg_count++] = ctx->ac.i32_0; /* soffset */
1951 params[arg_count++] = ctx->ac.i32_0; /* slc */
1952
1953 ac_build_type_name_for_intr(return_type, type, sizeof(type));
1954 snprintf(name, sizeof(name),
1955 "llvm.amdgcn.raw.buffer.atomic.%s.%s", op, type);
1956 } else {
1957 params[arg_count++] = ctx->ac.i32_0; /* vindex */
1958 params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
1959 params[arg_count++] = ctx->ac.i1false; /* slc */
1960
1961 assert(return_type == ctx->ac.i32);
1962 snprintf(name, sizeof(name),
1963 "llvm.amdgcn.buffer.atomic.%s", op);
1964 }
1965
1966 result = ac_build_intrinsic(&ctx->ac, name, return_type, params,
1967 arg_count, 0);
1968 }
1969
1970 result = exit_waterfall(ctx, &wctx, result);
1971 if (ctx->ac.postponed_kill)
1972 ac_build_endif(&ctx->ac, 7001);
1973 return result;
1974 }
1975
1976 static LLVMValueRef visit_load_buffer(struct ac_nir_context *ctx,
1977 nir_intrinsic_instr *instr)
1978 {
1979 struct waterfall_context wctx;
1980 LLVMValueRef rsrc_base = enter_waterfall_ssbo(ctx, &wctx, instr, instr->src[0]);
1981
1982 int elem_size_bytes = instr->dest.ssa.bit_size / 8;
1983 int num_components = instr->num_components;
1984 enum gl_access_qualifier access = nir_intrinsic_access(instr);
1985 unsigned cache_policy = get_cache_policy(ctx, access, false, false);
1986
1987 LLVMValueRef offset = get_src(ctx, instr->src[1]);
1988 LLVMValueRef rsrc = ctx->abi->load_ssbo(ctx->abi, rsrc_base, false);
1989 LLVMValueRef vindex = ctx->ac.i32_0;
1990
1991 LLVMTypeRef def_type = get_def_type(ctx, &instr->dest.ssa);
1992 LLVMTypeRef def_elem_type = num_components > 1 ? LLVMGetElementType(def_type) : def_type;
1993
1994 LLVMValueRef results[4];
1995 for (int i = 0; i < num_components;) {
1996 int num_elems = num_components - i;
1997 if (elem_size_bytes < 4 && nir_intrinsic_align(instr) % 4 != 0)
1998 num_elems = 1;
1999 if (num_elems * elem_size_bytes > 16)
2000 num_elems = 16 / elem_size_bytes;
2001 int load_bytes = num_elems * elem_size_bytes;
2002
2003 LLVMValueRef immoffset = LLVMConstInt(ctx->ac.i32, i * elem_size_bytes, false);
2004
2005 LLVMValueRef ret;
2006
2007 if (load_bytes == 1) {
2008 ret = ac_build_tbuffer_load_byte(&ctx->ac,
2009 rsrc,
2010 offset,
2011 ctx->ac.i32_0,
2012 immoffset,
2013 cache_policy);
2014 } else if (load_bytes == 2) {
2015 ret = ac_build_tbuffer_load_short(&ctx->ac,
2016 rsrc,
2017 offset,
2018 ctx->ac.i32_0,
2019 immoffset,
2020 cache_policy);
2021 } else {
2022 int num_channels = util_next_power_of_two(load_bytes) / 4;
2023 bool can_speculate = access & ACCESS_CAN_REORDER;
2024
2025 ret = ac_build_buffer_load(&ctx->ac, rsrc, num_channels,
2026 vindex, offset, immoffset, 0,
2027 cache_policy, can_speculate, false);
2028 }
2029
2030 LLVMTypeRef byte_vec = LLVMVectorType(ctx->ac.i8, ac_get_type_size(LLVMTypeOf(ret)));
2031 ret = LLVMBuildBitCast(ctx->ac.builder, ret, byte_vec, "");
2032 ret = ac_trim_vector(&ctx->ac, ret, load_bytes);
2033
2034 LLVMTypeRef ret_type = LLVMVectorType(def_elem_type, num_elems);
2035 ret = LLVMBuildBitCast(ctx->ac.builder, ret, ret_type, "");
2036
2037 for (unsigned j = 0; j < num_elems; j++) {
2038 results[i + j] = LLVMBuildExtractElement(ctx->ac.builder, ret, LLVMConstInt(ctx->ac.i32, j, false), "");
2039 }
2040 i += num_elems;
2041 }
2042
2043 LLVMValueRef ret = ac_build_gather_values(&ctx->ac, results, num_components);
2044 return exit_waterfall(ctx, &wctx, ret);
2045 }
2046
2047 static LLVMValueRef enter_waterfall_ubo(struct ac_nir_context *ctx,
2048 struct waterfall_context *wctx,
2049 const nir_intrinsic_instr *instr)
2050 {
2051 return enter_waterfall(ctx, wctx, get_src(ctx, instr->src[0]),
2052 nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
2053 }
2054
2055 static LLVMValueRef visit_load_ubo_buffer(struct ac_nir_context *ctx,
2056 nir_intrinsic_instr *instr)
2057 {
2058 struct waterfall_context wctx;
2059 LLVMValueRef rsrc_base = enter_waterfall_ubo(ctx, &wctx, instr);
2060
2061 LLVMValueRef ret;
2062 LLVMValueRef rsrc = rsrc_base;
2063 LLVMValueRef offset = get_src(ctx, instr->src[1]);
2064 int num_components = instr->num_components;
2065
2066 if (ctx->abi->load_ubo)
2067 rsrc = ctx->abi->load_ubo(ctx->abi, rsrc);
2068
2069 if (instr->dest.ssa.bit_size == 64)
2070 num_components *= 2;
2071
2072 if (instr->dest.ssa.bit_size == 16 || instr->dest.ssa.bit_size == 8) {
2073 unsigned load_bytes = instr->dest.ssa.bit_size / 8;
2074 LLVMValueRef results[num_components];
2075 for (unsigned i = 0; i < num_components; ++i) {
2076 LLVMValueRef immoffset = LLVMConstInt(ctx->ac.i32,
2077 load_bytes * i, 0);
2078
2079 if (load_bytes == 1) {
2080 results[i] = ac_build_tbuffer_load_byte(&ctx->ac,
2081 rsrc,
2082 offset,
2083 ctx->ac.i32_0,
2084 immoffset,
2085 0);
2086 } else {
2087 assert(load_bytes == 2);
2088 results[i] = ac_build_tbuffer_load_short(&ctx->ac,
2089 rsrc,
2090 offset,
2091 ctx->ac.i32_0,
2092 immoffset,
2093 0);
2094 }
2095 }
2096 ret = ac_build_gather_values(&ctx->ac, results, num_components);
2097 } else {
2098 ret = ac_build_buffer_load(&ctx->ac, rsrc, num_components, NULL, offset,
2099 NULL, 0, 0, true, true);
2100
2101 ret = ac_trim_vector(&ctx->ac, ret, num_components);
2102 }
2103
2104 ret = LLVMBuildBitCast(ctx->ac.builder, ret,
2105 get_def_type(ctx, &instr->dest.ssa), "");
2106
2107 return exit_waterfall(ctx, &wctx, ret);
2108 }
2109
2110 static void
2111 get_deref_offset(struct ac_nir_context *ctx, nir_deref_instr *instr,
2112 bool vs_in, unsigned *vertex_index_out,
2113 LLVMValueRef *vertex_index_ref,
2114 unsigned *const_out, LLVMValueRef *indir_out)
2115 {
2116 nir_variable *var = nir_deref_instr_get_variable(instr);
2117 nir_deref_path path;
2118 unsigned idx_lvl = 1;
2119
2120 nir_deref_path_init(&path, instr, NULL);
2121
2122 if (vertex_index_out != NULL || vertex_index_ref != NULL) {
2123 if (vertex_index_ref) {
2124 *vertex_index_ref = get_src(ctx, path.path[idx_lvl]->arr.index);
2125 if (vertex_index_out)
2126 *vertex_index_out = 0;
2127 } else {
2128 *vertex_index_out = nir_src_as_uint(path.path[idx_lvl]->arr.index);
2129 }
2130 ++idx_lvl;
2131 }
2132
2133 uint32_t const_offset = 0;
2134 LLVMValueRef offset = NULL;
2135
2136 if (var->data.compact) {
2137 assert(instr->deref_type == nir_deref_type_array);
2138 const_offset = nir_src_as_uint(instr->arr.index);
2139 goto out;
2140 }
2141
2142 for (; path.path[idx_lvl]; ++idx_lvl) {
2143 const struct glsl_type *parent_type = path.path[idx_lvl - 1]->type;
2144 if (path.path[idx_lvl]->deref_type == nir_deref_type_struct) {
2145 unsigned index = path.path[idx_lvl]->strct.index;
2146
2147 for (unsigned i = 0; i < index; i++) {
2148 const struct glsl_type *ft = glsl_get_struct_field(parent_type, i);
2149 const_offset += glsl_count_attribute_slots(ft, vs_in);
2150 }
2151 } else if(path.path[idx_lvl]->deref_type == nir_deref_type_array) {
2152 unsigned size = glsl_count_attribute_slots(path.path[idx_lvl]->type, vs_in);
2153 if (nir_src_is_const(path.path[idx_lvl]->arr.index)) {
2154 const_offset += size *
2155 nir_src_as_uint(path.path[idx_lvl]->arr.index);
2156 } else {
2157 LLVMValueRef array_off = LLVMBuildMul(ctx->ac.builder, LLVMConstInt(ctx->ac.i32, size, 0),
2158 get_src(ctx, path.path[idx_lvl]->arr.index), "");
2159 if (offset)
2160 offset = LLVMBuildAdd(ctx->ac.builder, offset, array_off, "");
2161 else
2162 offset = array_off;
2163 }
2164 } else
2165 unreachable("Uhandled deref type in get_deref_instr_offset");
2166 }
2167
2168 out:
2169 nir_deref_path_finish(&path);
2170
2171 if (const_offset && offset)
2172 offset = LLVMBuildAdd(ctx->ac.builder, offset,
2173 LLVMConstInt(ctx->ac.i32, const_offset, 0),
2174 "");
2175
2176 *const_out = const_offset;
2177 *indir_out = offset;
2178 }
2179
2180 static LLVMValueRef load_tess_varyings(struct ac_nir_context *ctx,
2181 nir_intrinsic_instr *instr,
2182 bool load_inputs)
2183 {
2184 LLVMValueRef result;
2185 LLVMValueRef vertex_index = NULL;
2186 LLVMValueRef indir_index = NULL;
2187 unsigned const_index = 0;
2188
2189 nir_variable *var = nir_deref_instr_get_variable(nir_instr_as_deref(instr->src[0].ssa->parent_instr));
2190
2191 unsigned location = var->data.location;
2192 unsigned driver_location = var->data.driver_location;
2193 const bool is_patch = var->data.patch ||
2194 var->data.location == VARYING_SLOT_TESS_LEVEL_INNER ||
2195 var->data.location == VARYING_SLOT_TESS_LEVEL_OUTER;
2196 const bool is_compact = var->data.compact;
2197
2198 get_deref_offset(ctx, nir_instr_as_deref(instr->src[0].ssa->parent_instr),
2199 false, NULL, is_patch ? NULL : &vertex_index,
2200 &const_index, &indir_index);
2201
2202 LLVMTypeRef dest_type = get_def_type(ctx, &instr->dest.ssa);
2203
2204 LLVMTypeRef src_component_type;
2205 if (LLVMGetTypeKind(dest_type) == LLVMVectorTypeKind)
2206 src_component_type = LLVMGetElementType(dest_type);
2207 else
2208 src_component_type = dest_type;
2209
2210 result = ctx->abi->load_tess_varyings(ctx->abi, src_component_type,
2211 vertex_index, indir_index,
2212 const_index, location, driver_location,
2213 var->data.location_frac,
2214 instr->num_components,
2215 is_patch, is_compact, load_inputs);
2216 if (instr->dest.ssa.bit_size == 16) {
2217 result = ac_to_integer(&ctx->ac, result);
2218 result = LLVMBuildTrunc(ctx->ac.builder, result, dest_type, "");
2219 }
2220 return LLVMBuildBitCast(ctx->ac.builder, result, dest_type, "");
2221 }
2222
2223 static unsigned
2224 type_scalar_size_bytes(const struct glsl_type *type)
2225 {
2226 assert(glsl_type_is_vector_or_scalar(type) ||
2227 glsl_type_is_matrix(type));
2228 return glsl_type_is_boolean(type) ? 4 : glsl_get_bit_size(type) / 8;
2229 }
2230
2231 static LLVMValueRef visit_load_var(struct ac_nir_context *ctx,
2232 nir_intrinsic_instr *instr)
2233 {
2234 nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2235 nir_variable *var = nir_deref_instr_get_variable(deref);
2236
2237 LLVMValueRef values[8];
2238 int idx = 0;
2239 int ve = instr->dest.ssa.num_components;
2240 unsigned comp = 0;
2241 LLVMValueRef indir_index;
2242 LLVMValueRef ret;
2243 unsigned const_index;
2244 unsigned stride = 4;
2245 int mode = deref->mode;
2246
2247 if (var) {
2248 bool vs_in = ctx->stage == MESA_SHADER_VERTEX &&
2249 var->data.mode == nir_var_shader_in;
2250 idx = var->data.driver_location;
2251 comp = var->data.location_frac;
2252 mode = var->data.mode;
2253
2254 get_deref_offset(ctx, deref, vs_in, NULL, NULL,
2255 &const_index, &indir_index);
2256
2257 if (var->data.compact) {
2258 stride = 1;
2259 const_index += comp;
2260 comp = 0;
2261 }
2262 }
2263
2264 if (instr->dest.ssa.bit_size == 64 &&
2265 (deref->mode == nir_var_shader_in ||
2266 deref->mode == nir_var_shader_out ||
2267 deref->mode == nir_var_function_temp))
2268 ve *= 2;
2269
2270 switch (mode) {
2271 case nir_var_shader_in:
2272 if (ctx->stage == MESA_SHADER_TESS_CTRL ||
2273 ctx->stage == MESA_SHADER_TESS_EVAL) {
2274 return load_tess_varyings(ctx, instr, true);
2275 }
2276
2277 if (ctx->stage == MESA_SHADER_GEOMETRY) {
2278 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
2279 LLVMValueRef indir_index;
2280 unsigned const_index, vertex_index;
2281 get_deref_offset(ctx, deref, false, &vertex_index, NULL,
2282 &const_index, &indir_index);
2283 assert(indir_index == NULL);
2284
2285 return ctx->abi->load_inputs(ctx->abi, var->data.location,
2286 var->data.driver_location,
2287 var->data.location_frac,
2288 instr->num_components, vertex_index, const_index, type);
2289 }
2290
2291 for (unsigned chan = comp; chan < ve + comp; chan++) {
2292 if (indir_index) {
2293 unsigned count = glsl_count_attribute_slots(
2294 var->type,
2295 ctx->stage == MESA_SHADER_VERTEX);
2296 count -= chan / 4;
2297 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2298 &ctx->ac, ctx->abi->inputs + idx + chan, count,
2299 stride, false, true);
2300
2301 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2302 tmp_vec,
2303 indir_index, "");
2304 } else
2305 values[chan] = ctx->abi->inputs[idx + chan + const_index * stride];
2306 }
2307 break;
2308 case nir_var_function_temp:
2309 for (unsigned chan = 0; chan < ve; chan++) {
2310 if (indir_index) {
2311 unsigned count = glsl_count_attribute_slots(
2312 var->type, false);
2313 count -= chan / 4;
2314 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2315 &ctx->ac, ctx->locals + idx + chan, count,
2316 stride, true, true);
2317
2318 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2319 tmp_vec,
2320 indir_index, "");
2321 } else {
2322 values[chan] = LLVMBuildLoad(ctx->ac.builder, ctx->locals[idx + chan + const_index * stride], "");
2323 }
2324 }
2325 break;
2326 case nir_var_shader_out:
2327 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
2328 return load_tess_varyings(ctx, instr, false);
2329 }
2330
2331 if (ctx->stage == MESA_SHADER_FRAGMENT &&
2332 var->data.fb_fetch_output &&
2333 ctx->abi->emit_fbfetch)
2334 return ctx->abi->emit_fbfetch(ctx->abi);
2335
2336 for (unsigned chan = comp; chan < ve + comp; chan++) {
2337 if (indir_index) {
2338 unsigned count = glsl_count_attribute_slots(
2339 var->type, false);
2340 count -= chan / 4;
2341 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2342 &ctx->ac, ctx->abi->outputs + idx + chan, count,
2343 stride, true, true);
2344
2345 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2346 tmp_vec,
2347 indir_index, "");
2348 } else {
2349 values[chan] = LLVMBuildLoad(ctx->ac.builder,
2350 ctx->abi->outputs[idx + chan + const_index * stride],
2351 "");
2352 }
2353 }
2354 break;
2355 case nir_var_mem_global: {
2356 LLVMValueRef address = get_src(ctx, instr->src[0]);
2357 LLVMTypeRef result_type = get_def_type(ctx, &instr->dest.ssa);
2358 unsigned explicit_stride = glsl_get_explicit_stride(deref->type);
2359 unsigned natural_stride = type_scalar_size_bytes(deref->type);
2360 unsigned stride = explicit_stride ? explicit_stride : natural_stride;
2361 int elem_size_bytes = ac_get_elem_bits(&ctx->ac, result_type) / 8;
2362 bool split_loads = ctx->ac.chip_class == GFX6 && elem_size_bytes < 4;
2363
2364 if (stride != natural_stride || split_loads) {
2365 if (LLVMGetTypeKind(result_type) == LLVMVectorTypeKind)
2366 result_type = LLVMGetElementType(result_type);
2367
2368 LLVMTypeRef ptr_type = LLVMPointerType(result_type,
2369 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2370 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2371
2372 for (unsigned i = 0; i < instr->dest.ssa.num_components; ++i) {
2373 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, i * stride / natural_stride, 0);
2374 values[i] = LLVMBuildLoad(ctx->ac.builder,
2375 ac_build_gep_ptr(&ctx->ac, address, offset), "");
2376
2377 if (nir_intrinsic_access(instr) & (ACCESS_COHERENT | ACCESS_VOLATILE))
2378 LLVMSetOrdering(values[i], LLVMAtomicOrderingMonotonic);
2379 }
2380 return ac_build_gather_values(&ctx->ac, values, instr->dest.ssa.num_components);
2381 } else {
2382 LLVMTypeRef ptr_type = LLVMPointerType(result_type,
2383 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2384 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2385 LLVMValueRef val = LLVMBuildLoad(ctx->ac.builder, address, "");
2386
2387 if (nir_intrinsic_access(instr) & (ACCESS_COHERENT | ACCESS_VOLATILE))
2388 LLVMSetOrdering(val, LLVMAtomicOrderingMonotonic);
2389 return val;
2390 }
2391 }
2392 default:
2393 unreachable("unhandle variable mode");
2394 }
2395 ret = ac_build_varying_gather_values(&ctx->ac, values, ve, comp);
2396 return LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->dest.ssa), "");
2397 }
2398
2399 static void
2400 visit_store_var(struct ac_nir_context *ctx,
2401 nir_intrinsic_instr *instr)
2402 {
2403 if (ctx->ac.postponed_kill) {
2404 LLVMValueRef cond = LLVMBuildLoad(ctx->ac.builder,
2405 ctx->ac.postponed_kill, "");
2406 ac_build_ifcc(&ctx->ac, cond, 7002);
2407 }
2408
2409 nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2410 nir_variable *var = nir_deref_instr_get_variable(deref);
2411
2412 LLVMValueRef temp_ptr, value;
2413 int idx = 0;
2414 unsigned comp = 0;
2415 LLVMValueRef src = ac_to_float(&ctx->ac, get_src(ctx, instr->src[1]));
2416 int writemask = instr->const_index[0];
2417 LLVMValueRef indir_index;
2418 unsigned const_index;
2419
2420 if (var) {
2421 get_deref_offset(ctx, deref, false,
2422 NULL, NULL, &const_index, &indir_index);
2423 idx = var->data.driver_location;
2424 comp = var->data.location_frac;
2425
2426 if (var->data.compact) {
2427 const_index += comp;
2428 comp = 0;
2429 }
2430 }
2431
2432 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src)) == 64 &&
2433 (deref->mode == nir_var_shader_out ||
2434 deref->mode == nir_var_function_temp)) {
2435
2436 src = LLVMBuildBitCast(ctx->ac.builder, src,
2437 LLVMVectorType(ctx->ac.f32, ac_get_llvm_num_components(src) * 2),
2438 "");
2439
2440 writemask = widen_mask(writemask, 2);
2441 }
2442
2443 writemask = writemask << comp;
2444
2445 switch (deref->mode) {
2446 case nir_var_shader_out:
2447
2448 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
2449 LLVMValueRef vertex_index = NULL;
2450 LLVMValueRef indir_index = NULL;
2451 unsigned const_index = 0;
2452 const bool is_patch = var->data.patch ||
2453 var->data.location == VARYING_SLOT_TESS_LEVEL_INNER ||
2454 var->data.location == VARYING_SLOT_TESS_LEVEL_OUTER;
2455
2456 get_deref_offset(ctx, deref, false, NULL,
2457 is_patch ? NULL : &vertex_index,
2458 &const_index, &indir_index);
2459
2460 ctx->abi->store_tcs_outputs(ctx->abi, var,
2461 vertex_index, indir_index,
2462 const_index, src, writemask);
2463 break;
2464 }
2465
2466 for (unsigned chan = 0; chan < 8; chan++) {
2467 int stride = 4;
2468 if (!(writemask & (1 << chan)))
2469 continue;
2470
2471 value = ac_llvm_extract_elem(&ctx->ac, src, chan - comp);
2472
2473 if (var->data.compact)
2474 stride = 1;
2475 if (indir_index) {
2476 unsigned count = glsl_count_attribute_slots(
2477 var->type, false);
2478 count -= chan / 4;
2479 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2480 &ctx->ac, ctx->abi->outputs + idx + chan, count,
2481 stride, true, true);
2482
2483 tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
2484 value, indir_index, "");
2485 build_store_values_extended(&ctx->ac, ctx->abi->outputs + idx + chan,
2486 count, stride, tmp_vec);
2487
2488 } else {
2489 temp_ptr = ctx->abi->outputs[idx + chan + const_index * stride];
2490
2491 LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
2492 }
2493 }
2494 break;
2495 case nir_var_function_temp:
2496 for (unsigned chan = 0; chan < 8; chan++) {
2497 if (!(writemask & (1 << chan)))
2498 continue;
2499
2500 value = ac_llvm_extract_elem(&ctx->ac, src, chan);
2501 if (indir_index) {
2502 unsigned count = glsl_count_attribute_slots(
2503 var->type, false);
2504 count -= chan / 4;
2505 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2506 &ctx->ac, ctx->locals + idx + chan, count,
2507 4, true, true);
2508
2509 tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
2510 value, indir_index, "");
2511 build_store_values_extended(&ctx->ac, ctx->locals + idx + chan,
2512 count, 4, tmp_vec);
2513 } else {
2514 temp_ptr = ctx->locals[idx + chan + const_index * 4];
2515
2516 LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
2517 }
2518 }
2519 break;
2520
2521 case nir_var_mem_global: {
2522 int writemask = instr->const_index[0];
2523 LLVMValueRef address = get_src(ctx, instr->src[0]);
2524 LLVMValueRef val = get_src(ctx, instr->src[1]);
2525
2526 unsigned explicit_stride = glsl_get_explicit_stride(deref->type);
2527 unsigned natural_stride = type_scalar_size_bytes(deref->type);
2528 unsigned stride = explicit_stride ? explicit_stride : natural_stride;
2529 int elem_size_bytes = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(val)) / 8;
2530 bool split_stores = ctx->ac.chip_class == GFX6 && elem_size_bytes < 4;
2531
2532 LLVMTypeRef ptr_type = LLVMPointerType(LLVMTypeOf(val),
2533 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2534 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2535
2536 if (writemask == (1u << ac_get_llvm_num_components(val)) - 1 &&
2537 stride == natural_stride && !split_stores) {
2538 LLVMTypeRef ptr_type = LLVMPointerType(LLVMTypeOf(val),
2539 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2540 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2541
2542 val = LLVMBuildBitCast(ctx->ac.builder, val,
2543 LLVMGetElementType(LLVMTypeOf(address)), "");
2544 LLVMValueRef store = LLVMBuildStore(ctx->ac.builder, val, address);
2545
2546 if (nir_intrinsic_access(instr) & (ACCESS_COHERENT | ACCESS_VOLATILE))
2547 LLVMSetOrdering(store, LLVMAtomicOrderingMonotonic);
2548 } else {
2549 LLVMTypeRef val_type = LLVMTypeOf(val);
2550 if (LLVMGetTypeKind(LLVMTypeOf(val)) == LLVMVectorTypeKind)
2551 val_type = LLVMGetElementType(val_type);
2552
2553 LLVMTypeRef ptr_type = LLVMPointerType(val_type,
2554 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2555 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2556 for (unsigned chan = 0; chan < 4; chan++) {
2557 if (!(writemask & (1 << chan)))
2558 continue;
2559
2560 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, chan * stride / natural_stride, 0);
2561
2562 LLVMValueRef ptr = ac_build_gep_ptr(&ctx->ac, address, offset);
2563 LLVMValueRef src = ac_llvm_extract_elem(&ctx->ac, val,
2564 chan);
2565 src = LLVMBuildBitCast(ctx->ac.builder, src,
2566 LLVMGetElementType(LLVMTypeOf(ptr)), "");
2567 LLVMValueRef store = LLVMBuildStore(ctx->ac.builder, src, ptr);
2568
2569 if (nir_intrinsic_access(instr) & (ACCESS_COHERENT | ACCESS_VOLATILE))
2570 LLVMSetOrdering(store, LLVMAtomicOrderingMonotonic);
2571 }
2572 }
2573 break;
2574 }
2575 default:
2576 abort();
2577 break;
2578 }
2579
2580 if (ctx->ac.postponed_kill)
2581 ac_build_endif(&ctx->ac, 7002);
2582 }
2583
2584 static int image_type_to_components_count(enum glsl_sampler_dim dim, bool array)
2585 {
2586 switch (dim) {
2587 case GLSL_SAMPLER_DIM_BUF:
2588 return 1;
2589 case GLSL_SAMPLER_DIM_1D:
2590 return array ? 2 : 1;
2591 case GLSL_SAMPLER_DIM_2D:
2592 return array ? 3 : 2;
2593 case GLSL_SAMPLER_DIM_MS:
2594 return array ? 4 : 3;
2595 case GLSL_SAMPLER_DIM_3D:
2596 case GLSL_SAMPLER_DIM_CUBE:
2597 return 3;
2598 case GLSL_SAMPLER_DIM_RECT:
2599 case GLSL_SAMPLER_DIM_SUBPASS:
2600 return 2;
2601 case GLSL_SAMPLER_DIM_SUBPASS_MS:
2602 return 3;
2603 default:
2604 break;
2605 }
2606 return 0;
2607 }
2608
2609 static LLVMValueRef adjust_sample_index_using_fmask(struct ac_llvm_context *ctx,
2610 LLVMValueRef coord_x, LLVMValueRef coord_y,
2611 LLVMValueRef coord_z,
2612 LLVMValueRef sample_index,
2613 LLVMValueRef fmask_desc_ptr)
2614 {
2615 unsigned sample_chan = coord_z ? 3 : 2;
2616 LLVMValueRef addr[4] = {coord_x, coord_y, coord_z};
2617 addr[sample_chan] = sample_index;
2618
2619 ac_apply_fmask_to_sample(ctx, fmask_desc_ptr, addr, coord_z != NULL);
2620 return addr[sample_chan];
2621 }
2622
2623 static nir_deref_instr *get_image_deref(const nir_intrinsic_instr *instr)
2624 {
2625 assert(instr->src[0].is_ssa);
2626 return nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2627 }
2628
2629 static LLVMValueRef get_image_descriptor(struct ac_nir_context *ctx,
2630 const nir_intrinsic_instr *instr,
2631 LLVMValueRef dynamic_index,
2632 enum ac_descriptor_type desc_type,
2633 bool write)
2634 {
2635 nir_deref_instr *deref_instr =
2636 instr->src[0].ssa->parent_instr->type == nir_instr_type_deref ?
2637 nir_instr_as_deref(instr->src[0].ssa->parent_instr) : NULL;
2638
2639 return get_sampler_desc(ctx, deref_instr, desc_type, &instr->instr, dynamic_index, true, write);
2640 }
2641
2642 static void get_image_coords(struct ac_nir_context *ctx,
2643 const nir_intrinsic_instr *instr,
2644 LLVMValueRef dynamic_desc_index,
2645 struct ac_image_args *args,
2646 enum glsl_sampler_dim dim,
2647 bool is_array)
2648 {
2649 LLVMValueRef src0 = get_src(ctx, instr->src[1]);
2650 LLVMValueRef masks[] = {
2651 LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
2652 LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
2653 };
2654 LLVMValueRef sample_index = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
2655
2656 int count;
2657 ASSERTED bool add_frag_pos = (dim == GLSL_SAMPLER_DIM_SUBPASS ||
2658 dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2659 bool is_ms = (dim == GLSL_SAMPLER_DIM_MS ||
2660 dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2661 bool gfx9_1d = ctx->ac.chip_class == GFX9 && dim == GLSL_SAMPLER_DIM_1D;
2662 assert(!add_frag_pos && "Input attachments should be lowered by this point.");
2663 count = image_type_to_components_count(dim, is_array);
2664
2665 if (is_ms && (instr->intrinsic == nir_intrinsic_image_deref_load ||
2666 instr->intrinsic == nir_intrinsic_bindless_image_load)) {
2667 LLVMValueRef fmask_load_address[3];
2668
2669 fmask_load_address[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
2670 fmask_load_address[1] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[1], "");
2671 if (is_array)
2672 fmask_load_address[2] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[2], "");
2673 else
2674 fmask_load_address[2] = NULL;
2675
2676 sample_index = adjust_sample_index_using_fmask(&ctx->ac,
2677 fmask_load_address[0],
2678 fmask_load_address[1],
2679 fmask_load_address[2],
2680 sample_index,
2681 get_sampler_desc(ctx, nir_instr_as_deref(instr->src[0].ssa->parent_instr),
2682 AC_DESC_FMASK, &instr->instr, dynamic_desc_index, true, false));
2683 }
2684 if (count == 1 && !gfx9_1d) {
2685 if (instr->src[1].ssa->num_components)
2686 args->coords[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
2687 else
2688 args->coords[0] = src0;
2689 } else {
2690 int chan;
2691 if (is_ms)
2692 count--;
2693 for (chan = 0; chan < count; ++chan) {
2694 args->coords[chan] = ac_llvm_extract_elem(&ctx->ac, src0, chan);
2695 }
2696
2697 if (gfx9_1d) {
2698 if (is_array) {
2699 args->coords[2] = args->coords[1];
2700 args->coords[1] = ctx->ac.i32_0;
2701 } else
2702 args->coords[1] = ctx->ac.i32_0;
2703 count++;
2704 }
2705 if (ctx->ac.chip_class == GFX9 &&
2706 dim == GLSL_SAMPLER_DIM_2D &&
2707 !is_array) {
2708 /* The hw can't bind a slice of a 3D image as a 2D
2709 * image, because it ignores BASE_ARRAY if the target
2710 * is 3D. The workaround is to read BASE_ARRAY and set
2711 * it as the 3rd address operand for all 2D images.
2712 */
2713 LLVMValueRef first_layer, const5, mask;
2714
2715 const5 = LLVMConstInt(ctx->ac.i32, 5, 0);
2716 mask = LLVMConstInt(ctx->ac.i32, S_008F24_BASE_ARRAY(~0), 0);
2717 first_layer = LLVMBuildExtractElement(ctx->ac.builder, args->resource, const5, "");
2718 first_layer = LLVMBuildAnd(ctx->ac.builder, first_layer, mask, "");
2719
2720 args->coords[count] = first_layer;
2721 count++;
2722 }
2723
2724
2725 if (is_ms) {
2726 args->coords[count] = sample_index;
2727 count++;
2728 }
2729 }
2730 }
2731
2732 static LLVMValueRef get_image_buffer_descriptor(struct ac_nir_context *ctx,
2733 const nir_intrinsic_instr *instr,
2734 LLVMValueRef dynamic_index,
2735 bool write, bool atomic)
2736 {
2737 LLVMValueRef rsrc = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_BUFFER, write);
2738 if (ctx->ac.chip_class == GFX9 && LLVM_VERSION_MAJOR < 9 && atomic) {
2739 LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 2, 0), "");
2740 LLVMValueRef stride = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 1, 0), "");
2741 stride = LLVMBuildLShr(ctx->ac.builder, stride, LLVMConstInt(ctx->ac.i32, 16, 0), "");
2742
2743 LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->ac.builder,
2744 LLVMBuildICmp(ctx->ac.builder, LLVMIntUGT, elem_count, stride, ""),
2745 elem_count, stride, "");
2746
2747 rsrc = LLVMBuildInsertElement(ctx->ac.builder, rsrc, new_elem_count,
2748 LLVMConstInt(ctx->ac.i32, 2, 0), "");
2749 }
2750 return rsrc;
2751 }
2752
2753 static LLVMValueRef enter_waterfall_image(struct ac_nir_context *ctx,
2754 struct waterfall_context *wctx,
2755 const nir_intrinsic_instr *instr)
2756 {
2757 nir_deref_instr *deref_instr = NULL;
2758
2759 if (instr->src[0].ssa->parent_instr->type == nir_instr_type_deref)
2760 deref_instr = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2761
2762 LLVMValueRef value = get_sampler_desc_index(ctx, deref_instr, &instr->instr, true);
2763 return enter_waterfall(ctx, wctx, value, nir_intrinsic_access(instr) & ACCESS_NON_UNIFORM);
2764 }
2765
2766 static LLVMValueRef visit_image_load(struct ac_nir_context *ctx,