projects / mesa.git / blob
? search:


fd6dbcce530ce9cf73837a54deb4909d547c1392
[mesa.git] / src / amd / common / 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 "ac_nir_to_llvm.h"
25 #include "ac_llvm_build.h"
26 #include "ac_llvm_util.h"
27 #include "ac_binary.h"
28 #include "sid.h"
29 #include "nir/nir.h"
30 #include "nir/nir_deref.h"
31 #include "util/bitscan.h"
32 #include "util/u_math.h"
33 #include "ac_shader_abi.h"
34 #include "ac_shader_util.h"
35
36 struct ac_nir_context {
37 struct ac_llvm_context ac;
38 struct ac_shader_abi *abi;
39
40 gl_shader_stage stage;
41
42 LLVMValueRef *ssa_defs;
43
44 struct hash_table *defs;
45 struct hash_table *phis;
46 struct hash_table *vars;
47
48 LLVMValueRef main_function;
49 LLVMBasicBlockRef continue_block;
50 LLVMBasicBlockRef break_block;
51
52 int num_locals;
53 LLVMValueRef *locals;
54 };
55
56 static LLVMValueRef get_sampler_desc(struct ac_nir_context *ctx,
57 nir_deref_instr *deref_instr,
58 enum ac_descriptor_type desc_type,
59 const nir_tex_instr *instr,
60 bool image, bool write);
61
62 static void
63 build_store_values_extended(struct ac_llvm_context *ac,
64 LLVMValueRef *values,
65 unsigned value_count,
66 unsigned value_stride,
67 LLVMValueRef vec)
68 {
69 LLVMBuilderRef builder = ac->builder;
70 unsigned i;
71
72 for (i = 0; i < value_count; i++) {
73 LLVMValueRef ptr = values[i * value_stride];
74 LLVMValueRef index = LLVMConstInt(ac->i32, i, false);
75 LLVMValueRef value = LLVMBuildExtractElement(builder, vec, index, "");
76 LLVMBuildStore(builder, value, ptr);
77 }
78 }
79
80 static enum ac_image_dim
81 get_ac_sampler_dim(const struct ac_llvm_context *ctx, enum glsl_sampler_dim dim,
82 bool is_array)
83 {
84 switch (dim) {
85 case GLSL_SAMPLER_DIM_1D:
86 if (ctx->chip_class >= GFX9)
87 return is_array ? ac_image_2darray : ac_image_2d;
88 return is_array ? ac_image_1darray : ac_image_1d;
89 case GLSL_SAMPLER_DIM_2D:
90 case GLSL_SAMPLER_DIM_RECT:
91 case GLSL_SAMPLER_DIM_EXTERNAL:
92 return is_array ? ac_image_2darray : ac_image_2d;
93 case GLSL_SAMPLER_DIM_3D:
94 return ac_image_3d;
95 case GLSL_SAMPLER_DIM_CUBE:
96 return ac_image_cube;
97 case GLSL_SAMPLER_DIM_MS:
98 return is_array ? ac_image_2darraymsaa : ac_image_2dmsaa;
99 case GLSL_SAMPLER_DIM_SUBPASS:
100 return ac_image_2darray;
101 case GLSL_SAMPLER_DIM_SUBPASS_MS:
102 return ac_image_2darraymsaa;
103 default:
104 unreachable("bad sampler dim");
105 }
106 }
107
108 static enum ac_image_dim
109 get_ac_image_dim(const struct ac_llvm_context *ctx, enum glsl_sampler_dim sdim,
110 bool is_array)
111 {
112 enum ac_image_dim dim = get_ac_sampler_dim(ctx, sdim, is_array);
113
114 if (dim == ac_image_cube ||
115 (ctx->chip_class <= VI && dim == ac_image_3d))
116 dim = ac_image_2darray;
117
118 return dim;
119 }
120
121 static LLVMTypeRef get_def_type(struct ac_nir_context *ctx,
122 const nir_ssa_def *def)
123 {
124 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, def->bit_size);
125 if (def->num_components > 1) {
126 type = LLVMVectorType(type, def->num_components);
127 }
128 return type;
129 }
130
131 static LLVMValueRef get_src(struct ac_nir_context *nir, nir_src src)
132 {
133 assert(src.is_ssa);
134 return nir->ssa_defs[src.ssa->index];
135 }
136
137 static LLVMValueRef
138 get_memory_ptr(struct ac_nir_context *ctx, nir_src src)
139 {
140 LLVMValueRef ptr = get_src(ctx, src);
141 ptr = LLVMBuildGEP(ctx->ac.builder, ctx->ac.lds, &ptr, 1, "");
142 int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
143
144 return LLVMBuildBitCast(ctx->ac.builder, ptr,
145 LLVMPointerType(ctx->ac.i32, addr_space), "");
146 }
147
148 static LLVMBasicBlockRef get_block(struct ac_nir_context *nir,
149 const struct nir_block *b)
150 {
151 struct hash_entry *entry = _mesa_hash_table_search(nir->defs, b);
152 return (LLVMBasicBlockRef)entry->data;
153 }
154
155 static LLVMValueRef get_alu_src(struct ac_nir_context *ctx,
156 nir_alu_src src,
157 unsigned num_components)
158 {
159 LLVMValueRef value = get_src(ctx, src.src);
160 bool need_swizzle = false;
161
162 assert(value);
163 unsigned src_components = ac_get_llvm_num_components(value);
164 for (unsigned i = 0; i < num_components; ++i) {
165 assert(src.swizzle[i] < src_components);
166 if (src.swizzle[i] != i)
167 need_swizzle = true;
168 }
169
170 if (need_swizzle || num_components != src_components) {
171 LLVMValueRef masks[] = {
172 LLVMConstInt(ctx->ac.i32, src.swizzle[0], false),
173 LLVMConstInt(ctx->ac.i32, src.swizzle[1], false),
174 LLVMConstInt(ctx->ac.i32, src.swizzle[2], false),
175 LLVMConstInt(ctx->ac.i32, src.swizzle[3], false)};
176
177 if (src_components > 1 && num_components == 1) {
178 value = LLVMBuildExtractElement(ctx->ac.builder, value,
179 masks[0], "");
180 } else if (src_components == 1 && num_components > 1) {
181 LLVMValueRef values[] = {value, value, value, value};
182 value = ac_build_gather_values(&ctx->ac, values, num_components);
183 } else {
184 LLVMValueRef swizzle = LLVMConstVector(masks, num_components);
185 value = LLVMBuildShuffleVector(ctx->ac.builder, value, value,
186 swizzle, "");
187 }
188 }
189 assert(!src.negate);
190 assert(!src.abs);
191 return value;
192 }
193
194 static LLVMValueRef emit_int_cmp(struct ac_llvm_context *ctx,
195 LLVMIntPredicate pred, LLVMValueRef src0,
196 LLVMValueRef src1)
197 {
198 LLVMValueRef result = LLVMBuildICmp(ctx->builder, pred, src0, src1, "");
199 return LLVMBuildSelect(ctx->builder, result,
200 LLVMConstInt(ctx->i32, 0xFFFFFFFF, false),
201 ctx->i32_0, "");
202 }
203
204 static LLVMValueRef emit_float_cmp(struct ac_llvm_context *ctx,
205 LLVMRealPredicate pred, LLVMValueRef src0,
206 LLVMValueRef src1)
207 {
208 LLVMValueRef result;
209 src0 = ac_to_float(ctx, src0);
210 src1 = ac_to_float(ctx, src1);
211 result = LLVMBuildFCmp(ctx->builder, pred, src0, src1, "");
212 return LLVMBuildSelect(ctx->builder, result,
213 LLVMConstInt(ctx->i32, 0xFFFFFFFF, false),
214 ctx->i32_0, "");
215 }
216
217 static LLVMValueRef emit_intrin_1f_param(struct ac_llvm_context *ctx,
218 const char *intrin,
219 LLVMTypeRef result_type,
220 LLVMValueRef src0)
221 {
222 char name[64];
223 LLVMValueRef params[] = {
224 ac_to_float(ctx, src0),
225 };
226
227 MAYBE_UNUSED const int length = snprintf(name, sizeof(name), "%s.f%d", intrin,
228 ac_get_elem_bits(ctx, result_type));
229 assert(length < sizeof(name));
230 return ac_build_intrinsic(ctx, name, result_type, params, 1, AC_FUNC_ATTR_READNONE);
231 }
232
233 static LLVMValueRef emit_intrin_2f_param(struct ac_llvm_context *ctx,
234 const char *intrin,
235 LLVMTypeRef result_type,
236 LLVMValueRef src0, LLVMValueRef src1)
237 {
238 char name[64];
239 LLVMValueRef params[] = {
240 ac_to_float(ctx, src0),
241 ac_to_float(ctx, src1),
242 };
243
244 MAYBE_UNUSED const int length = snprintf(name, sizeof(name), "%s.f%d", intrin,
245 ac_get_elem_bits(ctx, result_type));
246 assert(length < sizeof(name));
247 return ac_build_intrinsic(ctx, name, result_type, params, 2, AC_FUNC_ATTR_READNONE);
248 }
249
250 static LLVMValueRef emit_intrin_3f_param(struct ac_llvm_context *ctx,
251 const char *intrin,
252 LLVMTypeRef result_type,
253 LLVMValueRef src0, LLVMValueRef src1, LLVMValueRef src2)
254 {
255 char name[64];
256 LLVMValueRef params[] = {
257 ac_to_float(ctx, src0),
258 ac_to_float(ctx, src1),
259 ac_to_float(ctx, src2),
260 };
261
262 MAYBE_UNUSED const int length = snprintf(name, sizeof(name), "%s.f%d", intrin,
263 ac_get_elem_bits(ctx, result_type));
264 assert(length < sizeof(name));
265 return ac_build_intrinsic(ctx, name, result_type, params, 3, AC_FUNC_ATTR_READNONE);
266 }
267
268 static LLVMValueRef emit_bcsel(struct ac_llvm_context *ctx,
269 LLVMValueRef src0, LLVMValueRef src1, LLVMValueRef src2)
270 {
271 assert(LLVMGetTypeKind(LLVMTypeOf(src0)) != LLVMVectorTypeKind);
272
273 LLVMValueRef v = LLVMBuildICmp(ctx->builder, LLVMIntNE, src0,
274 ctx->i32_0, "");
275 return LLVMBuildSelect(ctx->builder, v,
276 ac_to_integer_or_pointer(ctx, src1),
277 ac_to_integer_or_pointer(ctx, src2), "");
278 }
279
280 static LLVMValueRef emit_minmax_int(struct ac_llvm_context *ctx,
281 LLVMIntPredicate pred,
282 LLVMValueRef src0, LLVMValueRef src1)
283 {
284 return LLVMBuildSelect(ctx->builder,
285 LLVMBuildICmp(ctx->builder, pred, src0, src1, ""),
286 src0,
287 src1, "");
288
289 }
290 static LLVMValueRef emit_iabs(struct ac_llvm_context *ctx,
291 LLVMValueRef src0)
292 {
293 return emit_minmax_int(ctx, LLVMIntSGT, src0,
294 LLVMBuildNeg(ctx->builder, src0, ""));
295 }
296
297 static LLVMValueRef emit_uint_carry(struct ac_llvm_context *ctx,
298 const char *intrin,
299 LLVMValueRef src0, LLVMValueRef src1)
300 {
301 LLVMTypeRef ret_type;
302 LLVMTypeRef types[] = { ctx->i32, ctx->i1 };
303 LLVMValueRef res;
304 LLVMValueRef params[] = { src0, src1 };
305 ret_type = LLVMStructTypeInContext(ctx->context, types,
306 2, true);
307
308 res = ac_build_intrinsic(ctx, intrin, ret_type,
309 params, 2, AC_FUNC_ATTR_READNONE);
310
311 res = LLVMBuildExtractValue(ctx->builder, res, 1, "");
312 res = LLVMBuildZExt(ctx->builder, res, ctx->i32, "");
313 return res;
314 }
315
316 static LLVMValueRef emit_b2f(struct ac_llvm_context *ctx,
317 LLVMValueRef src0,
318 unsigned bitsize)
319 {
320 LLVMValueRef result = LLVMBuildAnd(ctx->builder, src0,
321 LLVMBuildBitCast(ctx->builder, LLVMConstReal(ctx->f32, 1.0), ctx->i32, ""),
322 "");
323 result = LLVMBuildBitCast(ctx->builder, result, ctx->f32, "");
324
325 if (bitsize == 32)
326 return result;
327
328 return LLVMBuildFPExt(ctx->builder, result, ctx->f64, "");
329 }
330
331 static LLVMValueRef emit_f2b(struct ac_llvm_context *ctx,
332 LLVMValueRef src0)
333 {
334 src0 = ac_to_float(ctx, src0);
335 LLVMValueRef zero = LLVMConstNull(LLVMTypeOf(src0));
336 return LLVMBuildSExt(ctx->builder,
337 LLVMBuildFCmp(ctx->builder, LLVMRealUNE, src0, zero, ""),
338 ctx->i32, "");
339 }
340
341 static LLVMValueRef emit_b2i(struct ac_llvm_context *ctx,
342 LLVMValueRef src0,
343 unsigned bitsize)
344 {
345 LLVMValueRef result = LLVMBuildAnd(ctx->builder, src0, ctx->i32_1, "");
346
347 if (bitsize == 32)
348 return result;
349
350 return LLVMBuildZExt(ctx->builder, result, ctx->i64, "");
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 >= VI) {
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 >= VI)
383 result = LLVMBuildSelect(ctx->builder, cond, ctx->f32_0, result, "");
384 else {
385 /* for SI/CIK */
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, LLVMRealUGT,
392 LLVMBuildBitCast(ctx->builder, LLVMConstInt(ctx->i32, 0x38800000, false), ctx->f32, ""),
393 temp, "");
394 cond2 = LLVMBuildFCmp(ctx->builder, LLVMRealUNE,
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_bitfield_extract(struct ac_llvm_context *ctx,
429 bool is_signed,
430 const LLVMValueRef srcs[3])
431 {
432 LLVMValueRef result;
433
434 if (HAVE_LLVM >= 0x0800) {
435 LLVMValueRef icond = LLVMBuildICmp(ctx->builder, LLVMIntEQ, srcs[2], LLVMConstInt(ctx->i32, 32, false), "");
436 result = ac_build_bfe(ctx, srcs[0], srcs[1], srcs[2], is_signed);
437 result = LLVMBuildSelect(ctx->builder, icond, srcs[0], result, "");
438 } else {
439 /* FIXME: LLVM 7+ returns incorrect result when count is 0.
440 * https://bugs.freedesktop.org/show_bug.cgi?id=107276
441 */
442 LLVMValueRef zero = ctx->i32_0;
443 LLVMValueRef icond1 = LLVMBuildICmp(ctx->builder, LLVMIntEQ, srcs[2], LLVMConstInt(ctx->i32, 32, false), "");
444 LLVMValueRef icond2 = LLVMBuildICmp(ctx->builder, LLVMIntEQ, srcs[2], zero, "");
445
446 result = ac_build_bfe(ctx, srcs[0], srcs[1], srcs[2], is_signed);
447 result = LLVMBuildSelect(ctx->builder, icond1, srcs[0], result, "");
448 result = LLVMBuildSelect(ctx->builder, icond2, zero, result, "");
449 }
450
451 return result;
452 }
453
454 static LLVMValueRef emit_bitfield_insert(struct ac_llvm_context *ctx,
455 LLVMValueRef src0, LLVMValueRef src1,
456 LLVMValueRef src2, LLVMValueRef src3)
457 {
458 LLVMValueRef bfi_args[3], result;
459
460 bfi_args[0] = LLVMBuildShl(ctx->builder,
461 LLVMBuildSub(ctx->builder,
462 LLVMBuildShl(ctx->builder,
463 ctx->i32_1,
464 src3, ""),
465 ctx->i32_1, ""),
466 src2, "");
467 bfi_args[1] = LLVMBuildShl(ctx->builder, src1, src2, "");
468 bfi_args[2] = src0;
469
470 LLVMValueRef icond = LLVMBuildICmp(ctx->builder, LLVMIntEQ, src3, LLVMConstInt(ctx->i32, 32, false), "");
471
472 /* Calculate:
473 * (arg0 & arg1) | (~arg0 & arg2) = arg2 ^ (arg0 & (arg1 ^ arg2)
474 * Use the right-hand side, which the LLVM backend can convert to V_BFI.
475 */
476 result = LLVMBuildXor(ctx->builder, bfi_args[2],
477 LLVMBuildAnd(ctx->builder, bfi_args[0],
478 LLVMBuildXor(ctx->builder, bfi_args[1], bfi_args[2], ""), ""), "");
479
480 result = LLVMBuildSelect(ctx->builder, icond, src1, result, "");
481 return result;
482 }
483
484 static LLVMValueRef emit_pack_half_2x16(struct ac_llvm_context *ctx,
485 LLVMValueRef src0)
486 {
487 LLVMValueRef comp[2];
488
489 src0 = ac_to_float(ctx, src0);
490 comp[0] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_0, "");
491 comp[1] = LLVMBuildExtractElement(ctx->builder, src0, ctx->i32_1, "");
492
493 return LLVMBuildBitCast(ctx->builder, ac_build_cvt_pkrtz_f16(ctx, comp),
494 ctx->i32, "");
495 }
496
497 static LLVMValueRef emit_unpack_half_2x16(struct ac_llvm_context *ctx,
498 LLVMValueRef src0)
499 {
500 LLVMValueRef const16 = LLVMConstInt(ctx->i32, 16, false);
501 LLVMValueRef temps[2], val;
502 int i;
503
504 for (i = 0; i < 2; i++) {
505 val = i == 1 ? LLVMBuildLShr(ctx->builder, src0, const16, "") : src0;
506 val = LLVMBuildTrunc(ctx->builder, val, ctx->i16, "");
507 val = LLVMBuildBitCast(ctx->builder, val, ctx->f16, "");
508 temps[i] = LLVMBuildFPExt(ctx->builder, val, ctx->f32, "");
509 }
510 return ac_build_gather_values(ctx, temps, 2);
511 }
512
513 static LLVMValueRef emit_ddxy(struct ac_nir_context *ctx,
514 nir_op op,
515 LLVMValueRef src0)
516 {
517 unsigned mask;
518 int idx;
519 LLVMValueRef result;
520
521 if (op == nir_op_fddx_fine)
522 mask = AC_TID_MASK_LEFT;
523 else if (op == nir_op_fddy_fine)
524 mask = AC_TID_MASK_TOP;
525 else
526 mask = AC_TID_MASK_TOP_LEFT;
527
528 /* for DDX we want to next X pixel, DDY next Y pixel. */
529 if (op == nir_op_fddx_fine ||
530 op == nir_op_fddx_coarse ||
531 op == nir_op_fddx)
532 idx = 1;
533 else
534 idx = 2;
535
536 result = ac_build_ddxy(&ctx->ac, mask, idx, src0);
537 return result;
538 }
539
540 /*
541 * this takes an I,J coordinate pair,
542 * and works out the X and Y derivatives.
543 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
544 */
545 static LLVMValueRef emit_ddxy_interp(
546 struct ac_nir_context *ctx,
547 LLVMValueRef interp_ij)
548 {
549 LLVMValueRef result[4], a;
550 unsigned i;
551
552 for (i = 0; i < 2; i++) {
553 a = LLVMBuildExtractElement(ctx->ac.builder, interp_ij,
554 LLVMConstInt(ctx->ac.i32, i, false), "");
555 result[i] = emit_ddxy(ctx, nir_op_fddx, a);
556 result[2+i] = emit_ddxy(ctx, nir_op_fddy, a);
557 }
558 return ac_build_gather_values(&ctx->ac, result, 4);
559 }
560
561 static void visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
562 {
563 LLVMValueRef src[4], result = NULL;
564 unsigned num_components = instr->dest.dest.ssa.num_components;
565 unsigned src_components;
566 LLVMTypeRef def_type = get_def_type(ctx, &instr->dest.dest.ssa);
567
568 assert(nir_op_infos[instr->op].num_inputs <= ARRAY_SIZE(src));
569 switch (instr->op) {
570 case nir_op_vec2:
571 case nir_op_vec3:
572 case nir_op_vec4:
573 src_components = 1;
574 break;
575 case nir_op_pack_half_2x16:
576 src_components = 2;
577 break;
578 case nir_op_unpack_half_2x16:
579 src_components = 1;
580 break;
581 case nir_op_cube_face_coord:
582 case nir_op_cube_face_index:
583 src_components = 3;
584 break;
585 default:
586 src_components = num_components;
587 break;
588 }
589 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
590 src[i] = get_alu_src(ctx, instr->src[i], src_components);
591
592 switch (instr->op) {
593 case nir_op_fmov:
594 case nir_op_imov:
595 result = src[0];
596 break;
597 case nir_op_fneg:
598 src[0] = ac_to_float(&ctx->ac, src[0]);
599 result = LLVMBuildFNeg(ctx->ac.builder, src[0], "");
600 break;
601 case nir_op_ineg:
602 result = LLVMBuildNeg(ctx->ac.builder, src[0], "");
603 break;
604 case nir_op_inot:
605 result = LLVMBuildNot(ctx->ac.builder, src[0], "");
606 break;
607 case nir_op_iadd:
608 result = LLVMBuildAdd(ctx->ac.builder, src[0], src[1], "");
609 break;
610 case nir_op_fadd:
611 src[0] = ac_to_float(&ctx->ac, src[0]);
612 src[1] = ac_to_float(&ctx->ac, src[1]);
613 result = LLVMBuildFAdd(ctx->ac.builder, src[0], src[1], "");
614 break;
615 case nir_op_fsub:
616 src[0] = ac_to_float(&ctx->ac, src[0]);
617 src[1] = ac_to_float(&ctx->ac, src[1]);
618 result = LLVMBuildFSub(ctx->ac.builder, src[0], src[1], "");
619 break;
620 case nir_op_isub:
621 result = LLVMBuildSub(ctx->ac.builder, src[0], src[1], "");
622 break;
623 case nir_op_imul:
624 result = LLVMBuildMul(ctx->ac.builder, src[0], src[1], "");
625 break;
626 case nir_op_imod:
627 result = LLVMBuildSRem(ctx->ac.builder, src[0], src[1], "");
628 break;
629 case nir_op_umod:
630 result = LLVMBuildURem(ctx->ac.builder, src[0], src[1], "");
631 break;
632 case nir_op_fmod:
633 src[0] = ac_to_float(&ctx->ac, src[0]);
634 src[1] = ac_to_float(&ctx->ac, src[1]);
635 result = ac_build_fdiv(&ctx->ac, src[0], src[1]);
636 result = emit_intrin_1f_param(&ctx->ac, "llvm.floor",
637 ac_to_float_type(&ctx->ac, def_type), result);
638 result = LLVMBuildFMul(ctx->ac.builder, src[1] , result, "");
639 result = LLVMBuildFSub(ctx->ac.builder, src[0], result, "");
640 break;
641 case nir_op_frem:
642 src[0] = ac_to_float(&ctx->ac, src[0]);
643 src[1] = ac_to_float(&ctx->ac, src[1]);
644 result = LLVMBuildFRem(ctx->ac.builder, src[0], src[1], "");
645 break;
646 case nir_op_irem:
647 result = LLVMBuildSRem(ctx->ac.builder, src[0], src[1], "");
648 break;
649 case nir_op_idiv:
650 result = LLVMBuildSDiv(ctx->ac.builder, src[0], src[1], "");
651 break;
652 case nir_op_udiv:
653 result = LLVMBuildUDiv(ctx->ac.builder, src[0], src[1], "");
654 break;
655 case nir_op_fmul:
656 src[0] = ac_to_float(&ctx->ac, src[0]);
657 src[1] = ac_to_float(&ctx->ac, src[1]);
658 result = LLVMBuildFMul(ctx->ac.builder, src[0], src[1], "");
659 break;
660 case nir_op_frcp:
661 src[0] = ac_to_float(&ctx->ac, src[0]);
662 result = ac_build_fdiv(&ctx->ac, LLVMConstReal(LLVMTypeOf(src[0]), 1.0), src[0]);
663 break;
664 case nir_op_iand:
665 result = LLVMBuildAnd(ctx->ac.builder, src[0], src[1], "");
666 break;
667 case nir_op_ior:
668 result = LLVMBuildOr(ctx->ac.builder, src[0], src[1], "");
669 break;
670 case nir_op_ixor:
671 result = LLVMBuildXor(ctx->ac.builder, src[0], src[1], "");
672 break;
673 case nir_op_ishl:
674 result = LLVMBuildShl(ctx->ac.builder, src[0],
675 LLVMBuildZExt(ctx->ac.builder, src[1],
676 LLVMTypeOf(src[0]), ""),
677 "");
678 break;
679 case nir_op_ishr:
680 result = LLVMBuildAShr(ctx->ac.builder, src[0],
681 LLVMBuildZExt(ctx->ac.builder, src[1],
682 LLVMTypeOf(src[0]), ""),
683 "");
684 break;
685 case nir_op_ushr:
686 result = LLVMBuildLShr(ctx->ac.builder, src[0],
687 LLVMBuildZExt(ctx->ac.builder, src[1],
688 LLVMTypeOf(src[0]), ""),
689 "");
690 break;
691 case nir_op_ilt32:
692 result = emit_int_cmp(&ctx->ac, LLVMIntSLT, src[0], src[1]);
693 break;
694 case nir_op_ine32:
695 result = emit_int_cmp(&ctx->ac, LLVMIntNE, src[0], src[1]);
696 break;
697 case nir_op_ieq32:
698 result = emit_int_cmp(&ctx->ac, LLVMIntEQ, src[0], src[1]);
699 break;
700 case nir_op_ige32:
701 result = emit_int_cmp(&ctx->ac, LLVMIntSGE, src[0], src[1]);
702 break;
703 case nir_op_ult32:
704 result = emit_int_cmp(&ctx->ac, LLVMIntULT, src[0], src[1]);
705 break;
706 case nir_op_uge32:
707 result = emit_int_cmp(&ctx->ac, LLVMIntUGE, src[0], src[1]);
708 break;
709 case nir_op_feq32:
710 result = emit_float_cmp(&ctx->ac, LLVMRealOEQ, src[0], src[1]);
711 break;
712 case nir_op_fne32:
713 result = emit_float_cmp(&ctx->ac, LLVMRealUNE, src[0], src[1]);
714 break;
715 case nir_op_flt32:
716 result = emit_float_cmp(&ctx->ac, LLVMRealOLT, src[0], src[1]);
717 break;
718 case nir_op_fge32:
719 result = emit_float_cmp(&ctx->ac, LLVMRealOGE, src[0], src[1]);
720 break;
721 case nir_op_fabs:
722 result = emit_intrin_1f_param(&ctx->ac, "llvm.fabs",
723 ac_to_float_type(&ctx->ac, def_type), src[0]);
724 break;
725 case nir_op_iabs:
726 result = emit_iabs(&ctx->ac, src[0]);
727 break;
728 case nir_op_imax:
729 result = emit_minmax_int(&ctx->ac, LLVMIntSGT, src[0], src[1]);
730 break;
731 case nir_op_imin:
732 result = emit_minmax_int(&ctx->ac, LLVMIntSLT, src[0], src[1]);
733 break;
734 case nir_op_umax:
735 result = emit_minmax_int(&ctx->ac, LLVMIntUGT, src[0], src[1]);
736 break;
737 case nir_op_umin:
738 result = emit_minmax_int(&ctx->ac, LLVMIntULT, src[0], src[1]);
739 break;
740 case nir_op_isign:
741 result = ac_build_isign(&ctx->ac, src[0],
742 instr->dest.dest.ssa.bit_size);
743 break;
744 case nir_op_fsign:
745 src[0] = ac_to_float(&ctx->ac, src[0]);
746 result = ac_build_fsign(&ctx->ac, src[0],
747 instr->dest.dest.ssa.bit_size);
748 break;
749 case nir_op_ffloor:
750 result = emit_intrin_1f_param(&ctx->ac, "llvm.floor",
751 ac_to_float_type(&ctx->ac, def_type), src[0]);
752 break;
753 case nir_op_ftrunc:
754 result = emit_intrin_1f_param(&ctx->ac, "llvm.trunc",
755 ac_to_float_type(&ctx->ac, def_type), src[0]);
756 break;
757 case nir_op_fceil:
758 result = emit_intrin_1f_param(&ctx->ac, "llvm.ceil",
759 ac_to_float_type(&ctx->ac, def_type), src[0]);
760 break;
761 case nir_op_fround_even:
762 result = emit_intrin_1f_param(&ctx->ac, "llvm.rint",
763 ac_to_float_type(&ctx->ac, def_type),src[0]);
764 break;
765 case nir_op_ffract:
766 src[0] = ac_to_float(&ctx->ac, src[0]);
767 result = ac_build_fract(&ctx->ac, src[0],
768 instr->dest.dest.ssa.bit_size);
769 break;
770 case nir_op_fsin:
771 result = emit_intrin_1f_param(&ctx->ac, "llvm.sin",
772 ac_to_float_type(&ctx->ac, def_type), src[0]);
773 break;
774 case nir_op_fcos:
775 result = emit_intrin_1f_param(&ctx->ac, "llvm.cos",
776 ac_to_float_type(&ctx->ac, def_type), src[0]);
777 break;
778 case nir_op_fsqrt:
779 result = emit_intrin_1f_param(&ctx->ac, "llvm.sqrt",
780 ac_to_float_type(&ctx->ac, def_type), src[0]);
781 break;
782 case nir_op_fexp2:
783 result = emit_intrin_1f_param(&ctx->ac, "llvm.exp2",
784 ac_to_float_type(&ctx->ac, def_type), src[0]);
785 break;
786 case nir_op_flog2:
787 result = emit_intrin_1f_param(&ctx->ac, "llvm.log2",
788 ac_to_float_type(&ctx->ac, def_type), src[0]);
789 break;
790 case nir_op_frsq:
791 result = emit_intrin_1f_param(&ctx->ac, "llvm.sqrt",
792 ac_to_float_type(&ctx->ac, def_type), src[0]);
793 result = ac_build_fdiv(&ctx->ac, LLVMConstReal(LLVMTypeOf(result), 1.0), result);
794 break;
795 case nir_op_frexp_exp:
796 src[0] = ac_to_float(&ctx->ac, src[0]);
797 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.frexp.exp.i32.f64",
798 ctx->ac.i32, src, 1, AC_FUNC_ATTR_READNONE);
799
800 break;
801 case nir_op_frexp_sig:
802 src[0] = ac_to_float(&ctx->ac, src[0]);
803 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.frexp.mant.f64",
804 ctx->ac.f64, src, 1, AC_FUNC_ATTR_READNONE);
805 break;
806 case nir_op_fpow:
807 result = emit_intrin_2f_param(&ctx->ac, "llvm.pow",
808 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
809 break;
810 case nir_op_fmax:
811 result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
812 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
813 if (ctx->ac.chip_class < GFX9 &&
814 instr->dest.dest.ssa.bit_size == 32) {
815 /* Only pre-GFX9 chips do not flush denorms. */
816 result = emit_intrin_1f_param(&ctx->ac, "llvm.canonicalize",
817 ac_to_float_type(&ctx->ac, def_type),
818 result);
819 }
820 break;
821 case nir_op_fmin:
822 result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
823 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
824 if (ctx->ac.chip_class < GFX9 &&
825 instr->dest.dest.ssa.bit_size == 32) {
826 /* Only pre-GFX9 chips do not flush denorms. */
827 result = emit_intrin_1f_param(&ctx->ac, "llvm.canonicalize",
828 ac_to_float_type(&ctx->ac, def_type),
829 result);
830 }
831 break;
832 case nir_op_ffma:
833 result = emit_intrin_3f_param(&ctx->ac, "llvm.fmuladd",
834 ac_to_float_type(&ctx->ac, def_type), src[0], src[1], src[2]);
835 break;
836 case nir_op_ldexp:
837 src[0] = ac_to_float(&ctx->ac, src[0]);
838 if (ac_get_elem_bits(&ctx->ac, def_type) == 32)
839 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f32", ctx->ac.f32, src, 2, AC_FUNC_ATTR_READNONE);
840 else if (ac_get_elem_bits(&ctx->ac, def_type) == 16)
841 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f16", ctx->ac.f16, src, 2, AC_FUNC_ATTR_READNONE);
842 else
843 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.ldexp.f64", ctx->ac.f64, src, 2, AC_FUNC_ATTR_READNONE);
844 break;
845 case nir_op_ibitfield_extract:
846 result = emit_bitfield_extract(&ctx->ac, true, src);
847 break;
848 case nir_op_ubitfield_extract:
849 result = emit_bitfield_extract(&ctx->ac, false, src);
850 break;
851 case nir_op_bitfield_insert:
852 result = emit_bitfield_insert(&ctx->ac, src[0], src[1], src[2], src[3]);
853 break;
854 case nir_op_bitfield_reverse:
855 result = ac_build_bitfield_reverse(&ctx->ac, src[0]);
856 break;
857 case nir_op_bit_count:
858 result = ac_build_bit_count(&ctx->ac, src[0]);
859 break;
860 case nir_op_vec2:
861 case nir_op_vec3:
862 case nir_op_vec4:
863 for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
864 src[i] = ac_to_integer(&ctx->ac, src[i]);
865 result = ac_build_gather_values(&ctx->ac, src, num_components);
866 break;
867 case nir_op_f2i16:
868 case nir_op_f2i32:
869 case nir_op_f2i64:
870 src[0] = ac_to_float(&ctx->ac, src[0]);
871 result = LLVMBuildFPToSI(ctx->ac.builder, src[0], def_type, "");
872 break;
873 case nir_op_f2u16:
874 case nir_op_f2u32:
875 case nir_op_f2u64:
876 src[0] = ac_to_float(&ctx->ac, src[0]);
877 result = LLVMBuildFPToUI(ctx->ac.builder, src[0], def_type, "");
878 break;
879 case nir_op_i2f16:
880 case nir_op_i2f32:
881 case nir_op_i2f64:
882 src[0] = ac_to_integer(&ctx->ac, src[0]);
883 result = LLVMBuildSIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
884 break;
885 case nir_op_u2f16:
886 case nir_op_u2f32:
887 case nir_op_u2f64:
888 src[0] = ac_to_integer(&ctx->ac, src[0]);
889 result = LLVMBuildUIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
890 break;
891 case nir_op_f2f16_rtz:
892 src[0] = ac_to_float(&ctx->ac, src[0]);
893 if (LLVMTypeOf(src[0]) == ctx->ac.f64)
894 src[0] = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ctx->ac.f32, "");
895 LLVMValueRef param[2] = { src[0], ctx->ac.f32_0 };
896 result = ac_build_cvt_pkrtz_f16(&ctx->ac, param);
897 result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
898 break;
899 case nir_op_f2f16_rtne:
900 case nir_op_f2f16:
901 case nir_op_f2f32:
902 case nir_op_f2f64:
903 src[0] = ac_to_float(&ctx->ac, src[0]);
904 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
905 result = LLVMBuildFPExt(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
906 else
907 result = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
908 break;
909 case nir_op_u2u16:
910 case nir_op_u2u32:
911 case nir_op_u2u64:
912 src[0] = ac_to_integer(&ctx->ac, src[0]);
913 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
914 result = LLVMBuildZExt(ctx->ac.builder, src[0], def_type, "");
915 else
916 result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
917 break;
918 case nir_op_i2i16:
919 case nir_op_i2i32:
920 case nir_op_i2i64:
921 src[0] = ac_to_integer(&ctx->ac, src[0]);
922 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
923 result = LLVMBuildSExt(ctx->ac.builder, src[0], def_type, "");
924 else
925 result = LLVMBuildTrunc(ctx->ac.builder, src[0], def_type, "");
926 break;
927 case nir_op_b32csel:
928 result = emit_bcsel(&ctx->ac, src[0], src[1], src[2]);
929 break;
930 case nir_op_find_lsb:
931 src[0] = ac_to_integer(&ctx->ac, src[0]);
932 result = ac_find_lsb(&ctx->ac, ctx->ac.i32, src[0]);
933 break;
934 case nir_op_ufind_msb:
935 src[0] = ac_to_integer(&ctx->ac, src[0]);
936 result = ac_build_umsb(&ctx->ac, src[0], ctx->ac.i32);
937 break;
938 case nir_op_ifind_msb:
939 src[0] = ac_to_integer(&ctx->ac, src[0]);
940 result = ac_build_imsb(&ctx->ac, src[0], ctx->ac.i32);
941 break;
942 case nir_op_uadd_carry:
943 src[0] = ac_to_integer(&ctx->ac, src[0]);
944 src[1] = ac_to_integer(&ctx->ac, src[1]);
945 result = emit_uint_carry(&ctx->ac, "llvm.uadd.with.overflow.i32", src[0], src[1]);
946 break;
947 case nir_op_usub_borrow:
948 src[0] = ac_to_integer(&ctx->ac, src[0]);
949 src[1] = ac_to_integer(&ctx->ac, src[1]);
950 result = emit_uint_carry(&ctx->ac, "llvm.usub.with.overflow.i32", src[0], src[1]);
951 break;
952 case nir_op_b2f16:
953 case nir_op_b2f32:
954 case nir_op_b2f64:
955 result = emit_b2f(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
956 break;
957 case nir_op_f2b32:
958 result = emit_f2b(&ctx->ac, src[0]);
959 break;
960 case nir_op_b2i16:
961 case nir_op_b2i32:
962 case nir_op_b2i64:
963 result = emit_b2i(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
964 break;
965 case nir_op_i2b32:
966 src[0] = ac_to_integer(&ctx->ac, src[0]);
967 result = emit_i2b(&ctx->ac, src[0]);
968 break;
969 case nir_op_fquantize2f16:
970 result = emit_f2f16(&ctx->ac, src[0]);
971 break;
972 case nir_op_umul_high:
973 src[0] = ac_to_integer(&ctx->ac, src[0]);
974 src[1] = ac_to_integer(&ctx->ac, src[1]);
975 result = emit_umul_high(&ctx->ac, src[0], src[1]);
976 break;
977 case nir_op_imul_high:
978 src[0] = ac_to_integer(&ctx->ac, src[0]);
979 src[1] = ac_to_integer(&ctx->ac, src[1]);
980 result = emit_imul_high(&ctx->ac, src[0], src[1]);
981 break;
982 case nir_op_pack_half_2x16:
983 result = emit_pack_half_2x16(&ctx->ac, src[0]);
984 break;
985 case nir_op_unpack_half_2x16:
986 result = emit_unpack_half_2x16(&ctx->ac, src[0]);
987 break;
988 case nir_op_fddx:
989 case nir_op_fddy:
990 case nir_op_fddx_fine:
991 case nir_op_fddy_fine:
992 case nir_op_fddx_coarse:
993 case nir_op_fddy_coarse:
994 result = emit_ddxy(ctx, instr->op, src[0]);
995 break;
996
997 case nir_op_unpack_64_2x32_split_x: {
998 assert(ac_get_llvm_num_components(src[0]) == 1);
999 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1000 ctx->ac.v2i32,
1001 "");
1002 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1003 ctx->ac.i32_0, "");
1004 break;
1005 }
1006
1007 case nir_op_unpack_64_2x32_split_y: {
1008 assert(ac_get_llvm_num_components(src[0]) == 1);
1009 LLVMValueRef tmp = LLVMBuildBitCast(ctx->ac.builder, src[0],
1010 ctx->ac.v2i32,
1011 "");
1012 result = LLVMBuildExtractElement(ctx->ac.builder, tmp,
1013 ctx->ac.i32_1, "");
1014 break;
1015 }
1016
1017 case nir_op_pack_64_2x32_split: {
1018 LLVMValueRef tmp = LLVMGetUndef(ctx->ac.v2i32);
1019 tmp = ac_build_gather_values(&ctx->ac, src, 2);
1020 result = LLVMBuildBitCast(ctx->ac.builder, tmp, ctx->ac.i64, "");
1021 break;
1022 }
1023
1024 case nir_op_cube_face_coord: {
1025 src[0] = ac_to_float(&ctx->ac, src[0]);
1026 LLVMValueRef results[2];
1027 LLVMValueRef in[3];
1028 for (unsigned chan = 0; chan < 3; chan++)
1029 in[chan] = ac_llvm_extract_elem(&ctx->ac, src[0], chan);
1030 results[0] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubetc",
1031 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1032 results[1] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubesc",
1033 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1034 result = ac_build_gather_values(&ctx->ac, results, 2);
1035 break;
1036 }
1037
1038 case nir_op_cube_face_index: {
1039 src[0] = ac_to_float(&ctx->ac, src[0]);
1040 LLVMValueRef in[3];
1041 for (unsigned chan = 0; chan < 3; chan++)
1042 in[chan] = ac_llvm_extract_elem(&ctx->ac, src[0], chan);
1043 result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubeid",
1044 ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
1045 break;
1046 }
1047
1048 case nir_op_fmin3:
1049 result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
1050 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
1051 result = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
1052 ac_to_float_type(&ctx->ac, def_type), result, src[2]);
1053 break;
1054 case nir_op_umin3:
1055 result = emit_minmax_int(&ctx->ac, LLVMIntULT, src[0], src[1]);
1056 result = emit_minmax_int(&ctx->ac, LLVMIntULT, result, src[2]);
1057 break;
1058 case nir_op_imin3:
1059 result = emit_minmax_int(&ctx->ac, LLVMIntSLT, src[0], src[1]);
1060 result = emit_minmax_int(&ctx->ac, LLVMIntSLT, result, src[2]);
1061 break;
1062 case nir_op_fmax3:
1063 result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
1064 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
1065 result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
1066 ac_to_float_type(&ctx->ac, def_type), result, src[2]);
1067 break;
1068 case nir_op_umax3:
1069 result = emit_minmax_int(&ctx->ac, LLVMIntUGT, src[0], src[1]);
1070 result = emit_minmax_int(&ctx->ac, LLVMIntUGT, result, src[2]);
1071 break;
1072 case nir_op_imax3:
1073 result = emit_minmax_int(&ctx->ac, LLVMIntSGT, src[0], src[1]);
1074 result = emit_minmax_int(&ctx->ac, LLVMIntSGT, result, src[2]);
1075 break;
1076 case nir_op_fmed3: {
1077 LLVMValueRef tmp1 = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
1078 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
1079 LLVMValueRef tmp2 = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
1080 ac_to_float_type(&ctx->ac, def_type), src[0], src[1]);
1081 tmp2 = emit_intrin_2f_param(&ctx->ac, "llvm.minnum",
1082 ac_to_float_type(&ctx->ac, def_type), tmp2, src[2]);
1083 result = emit_intrin_2f_param(&ctx->ac, "llvm.maxnum",
1084 ac_to_float_type(&ctx->ac, def_type), tmp1, tmp2);
1085 break;
1086 }
1087 case nir_op_imed3: {
1088 LLVMValueRef tmp1 = emit_minmax_int(&ctx->ac, LLVMIntSLT, src[0], src[1]);
1089 LLVMValueRef tmp2 = emit_minmax_int(&ctx->ac, LLVMIntSGT, src[0], src[1]);
1090 tmp2 = emit_minmax_int(&ctx->ac, LLVMIntSLT, tmp2, src[2]);
1091 result = emit_minmax_int(&ctx->ac, LLVMIntSGT, tmp1, tmp2);
1092 break;
1093 }
1094 case nir_op_umed3: {
1095 LLVMValueRef tmp1 = emit_minmax_int(&ctx->ac, LLVMIntULT, src[0], src[1]);
1096 LLVMValueRef tmp2 = emit_minmax_int(&ctx->ac, LLVMIntUGT, src[0], src[1]);
1097 tmp2 = emit_minmax_int(&ctx->ac, LLVMIntULT, tmp2, src[2]);
1098 result = emit_minmax_int(&ctx->ac, LLVMIntUGT, tmp1, tmp2);
1099 break;
1100 }
1101
1102 default:
1103 fprintf(stderr, "Unknown NIR alu instr: ");
1104 nir_print_instr(&instr->instr, stderr);
1105 fprintf(stderr, "\n");
1106 abort();
1107 }
1108
1109 if (result) {
1110 assert(instr->dest.dest.is_ssa);
1111 result = ac_to_integer_or_pointer(&ctx->ac, result);
1112 ctx->ssa_defs[instr->dest.dest.ssa.index] = result;
1113 }
1114 }
1115
1116 static void visit_load_const(struct ac_nir_context *ctx,
1117 const nir_load_const_instr *instr)
1118 {
1119 LLVMValueRef values[4], value = NULL;
1120 LLVMTypeRef element_type =
1121 LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
1122
1123 for (unsigned i = 0; i < instr->def.num_components; ++i) {
1124 switch (instr->def.bit_size) {
1125 case 8:
1126 values[i] = LLVMConstInt(element_type,
1127 instr->value.u8[i], false);
1128 break;
1129 case 16:
1130 values[i] = LLVMConstInt(element_type,
1131 instr->value.u16[i], false);
1132 break;
1133 case 32:
1134 values[i] = LLVMConstInt(element_type,
1135 instr->value.u32[i], false);
1136 break;
1137 case 64:
1138 values[i] = LLVMConstInt(element_type,
1139 instr->value.u64[i], false);
1140 break;
1141 default:
1142 fprintf(stderr,
1143 "unsupported nir load_const bit_size: %d\n",
1144 instr->def.bit_size);
1145 abort();
1146 }
1147 }
1148 if (instr->def.num_components > 1) {
1149 value = LLVMConstVector(values, instr->def.num_components);
1150 } else
1151 value = values[0];
1152
1153 ctx->ssa_defs[instr->def.index] = value;
1154 }
1155
1156 static LLVMValueRef
1157 get_buffer_size(struct ac_nir_context *ctx, LLVMValueRef descriptor, bool in_elements)
1158 {
1159 LLVMValueRef size =
1160 LLVMBuildExtractElement(ctx->ac.builder, descriptor,
1161 LLVMConstInt(ctx->ac.i32, 2, false), "");
1162
1163 /* VI only */
1164 if (ctx->ac.chip_class == VI && in_elements) {
1165 /* On VI, the descriptor contains the size in bytes,
1166 * but TXQ must return the size in elements.
1167 * The stride is always non-zero for resources using TXQ.
1168 */
1169 LLVMValueRef stride =
1170 LLVMBuildExtractElement(ctx->ac.builder, descriptor,
1171 ctx->ac.i32_1, "");
1172 stride = LLVMBuildLShr(ctx->ac.builder, stride,
1173 LLVMConstInt(ctx->ac.i32, 16, false), "");
1174 stride = LLVMBuildAnd(ctx->ac.builder, stride,
1175 LLVMConstInt(ctx->ac.i32, 0x3fff, false), "");
1176
1177 size = LLVMBuildUDiv(ctx->ac.builder, size, stride, "");
1178 }
1179 return size;
1180 }
1181
1182 static LLVMValueRef lower_gather4_integer(struct ac_llvm_context *ctx,
1183 nir_variable *var,
1184 struct ac_image_args *args,
1185 const nir_tex_instr *instr)
1186 {
1187 const struct glsl_type *type = glsl_without_array(var->type);
1188 enum glsl_base_type stype = glsl_get_sampler_result_type(type);
1189 LLVMValueRef half_texel[2];
1190 LLVMValueRef compare_cube_wa = NULL;
1191 LLVMValueRef result;
1192
1193 //TODO Rect
1194 {
1195 struct ac_image_args txq_args = { 0 };
1196
1197 txq_args.dim = get_ac_sampler_dim(ctx, instr->sampler_dim, instr->is_array);
1198 txq_args.opcode = ac_image_get_resinfo;
1199 txq_args.dmask = 0xf;
1200 txq_args.lod = ctx->i32_0;
1201 txq_args.resource = args->resource;
1202 txq_args.attributes = AC_FUNC_ATTR_READNONE;
1203 LLVMValueRef size = ac_build_image_opcode(ctx, &txq_args);
1204
1205 for (unsigned c = 0; c < 2; c++) {
1206 half_texel[c] = LLVMBuildExtractElement(ctx->builder, size,
1207 LLVMConstInt(ctx->i32, c, false), "");
1208 half_texel[c] = LLVMBuildUIToFP(ctx->builder, half_texel[c], ctx->f32, "");
1209 half_texel[c] = ac_build_fdiv(ctx, ctx->f32_1, half_texel[c]);
1210 half_texel[c] = LLVMBuildFMul(ctx->builder, half_texel[c],
1211 LLVMConstReal(ctx->f32, -0.5), "");
1212 }
1213 }
1214
1215 LLVMValueRef orig_coords[2] = { args->coords[0], args->coords[1] };
1216
1217 for (unsigned c = 0; c < 2; c++) {
1218 LLVMValueRef tmp;
1219 tmp = LLVMBuildBitCast(ctx->builder, args->coords[c], ctx->f32, "");
1220 args->coords[c] = LLVMBuildFAdd(ctx->builder, tmp, half_texel[c], "");
1221 }
1222
1223 /*
1224 * Apparantly cube has issue with integer types that the workaround doesn't solve,
1225 * so this tests if the format is 8_8_8_8 and an integer type do an alternate
1226 * workaround by sampling using a scaled type and converting.
1227 * This is taken from amdgpu-pro shaders.
1228 */
1229 /* NOTE this produces some ugly code compared to amdgpu-pro,
1230 * LLVM ends up dumping SGPRs into VGPRs to deal with the compare/select,
1231 * and then reads them back. -pro generates two selects,
1232 * one s_cmp for the descriptor rewriting
1233 * one v_cmp for the coordinate and result changes.
1234 */
1235 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
1236 LLVMValueRef tmp, tmp2;
1237
1238 /* workaround 8/8/8/8 uint/sint cube gather bug */
1239 /* first detect it then change to a scaled read and f2i */
1240 tmp = LLVMBuildExtractElement(ctx->builder, args->resource, ctx->i32_1, "");
1241 tmp2 = tmp;
1242
1243 /* extract the DATA_FORMAT */
1244 tmp = ac_build_bfe(ctx, tmp, LLVMConstInt(ctx->i32, 20, false),
1245 LLVMConstInt(ctx->i32, 6, false), false);
1246
1247 /* is the DATA_FORMAT == 8_8_8_8 */
1248 compare_cube_wa = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tmp, LLVMConstInt(ctx->i32, V_008F14_IMG_DATA_FORMAT_8_8_8_8, false), "");
1249
1250 if (stype == GLSL_TYPE_UINT)
1251 /* Create a NUM FORMAT - 0x2 or 0x4 - USCALED or UINT */
1252 tmp = LLVMBuildSelect(ctx->builder, compare_cube_wa, LLVMConstInt(ctx->i32, 0x8000000, false),
1253 LLVMConstInt(ctx->i32, 0x10000000, false), "");
1254 else
1255 /* Create a NUM FORMAT - 0x3 or 0x5 - SSCALED or SINT */
1256 tmp = LLVMBuildSelect(ctx->builder, compare_cube_wa, LLVMConstInt(ctx->i32, 0xc000000, false),
1257 LLVMConstInt(ctx->i32, 0x14000000, false), "");
1258
1259 /* replace the NUM FORMAT in the descriptor */
1260 tmp2 = LLVMBuildAnd(ctx->builder, tmp2, LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT_GFX6, false), "");
1261 tmp2 = LLVMBuildOr(ctx->builder, tmp2, tmp, "");
1262
1263 args->resource = LLVMBuildInsertElement(ctx->builder, args->resource, tmp2, ctx->i32_1, "");
1264
1265 /* don't modify the coordinates for this case */
1266 for (unsigned c = 0; c < 2; ++c)
1267 args->coords[c] = LLVMBuildSelect(
1268 ctx->builder, compare_cube_wa,
1269 orig_coords[c], args->coords[c], "");
1270 }
1271
1272 args->attributes = AC_FUNC_ATTR_READNONE;
1273 result = ac_build_image_opcode(ctx, args);
1274
1275 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
1276 LLVMValueRef tmp, tmp2;
1277
1278 /* if the cube workaround is in place, f2i the result. */
1279 for (unsigned c = 0; c < 4; c++) {
1280 tmp = LLVMBuildExtractElement(ctx->builder, result, LLVMConstInt(ctx->i32, c, false), "");
1281 if (stype == GLSL_TYPE_UINT)
1282 tmp2 = LLVMBuildFPToUI(ctx->builder, tmp, ctx->i32, "");
1283 else
1284 tmp2 = LLVMBuildFPToSI(ctx->builder, tmp, ctx->i32, "");
1285 tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->i32, "");
1286 tmp2 = LLVMBuildBitCast(ctx->builder, tmp2, ctx->i32, "");
1287 tmp = LLVMBuildSelect(ctx->builder, compare_cube_wa, tmp2, tmp, "");
1288 tmp = LLVMBuildBitCast(ctx->builder, tmp, ctx->f32, "");
1289 result = LLVMBuildInsertElement(ctx->builder, result, tmp, LLVMConstInt(ctx->i32, c, false), "");
1290 }
1291 }
1292 return result;
1293 }
1294
1295 static nir_deref_instr *get_tex_texture_deref(const nir_tex_instr *instr)
1296 {
1297 nir_deref_instr *texture_deref_instr = NULL;
1298
1299 for (unsigned i = 0; i < instr->num_srcs; i++) {
1300 switch (instr->src[i].src_type) {
1301 case nir_tex_src_texture_deref:
1302 texture_deref_instr = nir_src_as_deref(instr->src[i].src);
1303 break;
1304 default:
1305 break;
1306 }
1307 }
1308 return texture_deref_instr;
1309 }
1310
1311 static LLVMValueRef build_tex_intrinsic(struct ac_nir_context *ctx,
1312 const nir_tex_instr *instr,
1313 struct ac_image_args *args)
1314 {
1315 if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
1316 unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
1317
1318 if (ctx->abi->gfx9_stride_size_workaround) {
1319 return ac_build_buffer_load_format_gfx9_safe(&ctx->ac,
1320 args->resource,
1321 args->coords[0],
1322 ctx->ac.i32_0,
1323 util_last_bit(mask),
1324 false, true);
1325 } else {
1326 return ac_build_buffer_load_format(&ctx->ac,
1327 args->resource,
1328 args->coords[0],
1329 ctx->ac.i32_0,
1330 util_last_bit(mask),
1331 false, true);
1332 }
1333 }
1334
1335 args->opcode = ac_image_sample;
1336
1337 switch (instr->op) {
1338 case nir_texop_txf:
1339 case nir_texop_txf_ms:
1340 case nir_texop_samples_identical:
1341 args->opcode = args->level_zero ||
1342 instr->sampler_dim == GLSL_SAMPLER_DIM_MS ?
1343 ac_image_load : ac_image_load_mip;
1344 args->level_zero = false;
1345 break;
1346 case nir_texop_txs:
1347 case nir_texop_query_levels:
1348 args->opcode = ac_image_get_resinfo;
1349 if (!args->lod)
1350 args->lod = ctx->ac.i32_0;
1351 args->level_zero = false;
1352 break;
1353 case nir_texop_tex:
1354 if (ctx->stage != MESA_SHADER_FRAGMENT) {
1355 assert(!args->lod);
1356 args->level_zero = true;
1357 }
1358 break;
1359 case nir_texop_tg4:
1360 args->opcode = ac_image_gather4;
1361 args->level_zero = true;
1362 break;
1363 case nir_texop_lod:
1364 args->opcode = ac_image_get_lod;
1365 break;
1366 default:
1367 break;
1368 }
1369
1370 if (instr->op == nir_texop_tg4 && ctx->ac.chip_class <= VI) {
1371 nir_deref_instr *texture_deref_instr = get_tex_texture_deref(instr);
1372 nir_variable *var = nir_deref_instr_get_variable(texture_deref_instr);
1373 const struct glsl_type *type = glsl_without_array(var->type);
1374 enum glsl_base_type stype = glsl_get_sampler_result_type(type);
1375 if (stype == GLSL_TYPE_UINT || stype == GLSL_TYPE_INT) {
1376 return lower_gather4_integer(&ctx->ac, var, args, instr);
1377 }
1378 }
1379
1380 /* Fixup for GFX9 which allocates 1D textures as 2D. */
1381 if (instr->op == nir_texop_lod && ctx->ac.chip_class >= GFX9) {
1382 if ((args->dim == ac_image_2darray ||
1383 args->dim == ac_image_2d) && !args->coords[1]) {
1384 args->coords[1] = ctx->ac.i32_0;
1385 }
1386 }
1387
1388 args->attributes = AC_FUNC_ATTR_READNONE;
1389 return ac_build_image_opcode(&ctx->ac, args);
1390 }
1391
1392 static LLVMValueRef visit_vulkan_resource_reindex(struct ac_nir_context *ctx,
1393 nir_intrinsic_instr *instr)
1394 {
1395 LLVMValueRef ptr = get_src(ctx, instr->src[0]);
1396 LLVMValueRef index = get_src(ctx, instr->src[1]);
1397
1398 LLVMValueRef result = LLVMBuildGEP(ctx->ac.builder, ptr, &index, 1, "");
1399 LLVMSetMetadata(result, ctx->ac.uniform_md_kind, ctx->ac.empty_md);
1400 return result;
1401 }
1402
1403 static LLVMValueRef visit_load_push_constant(struct ac_nir_context *ctx,
1404 nir_intrinsic_instr *instr)
1405 {
1406 LLVMValueRef ptr, addr;
1407 LLVMValueRef src0 = get_src(ctx, instr->src[0]);
1408 unsigned index = nir_intrinsic_base(instr);
1409
1410 addr = LLVMConstInt(ctx->ac.i32, index, 0);
1411 addr = LLVMBuildAdd(ctx->ac.builder, addr, src0, "");
1412
1413 /* Load constant values from user SGPRS when possible, otherwise
1414 * fallback to the default path that loads directly from memory.
1415 */
1416 if (LLVMIsConstant(src0) &&
1417 instr->dest.ssa.bit_size == 32) {
1418 unsigned count = instr->dest.ssa.num_components;
1419 unsigned offset = index;
1420
1421 offset += LLVMConstIntGetZExtValue(src0);
1422 offset /= 4;
1423
1424 offset -= ctx->abi->base_inline_push_consts;
1425
1426 if (offset + count <= ctx->abi->num_inline_push_consts) {
1427 return ac_build_gather_values(&ctx->ac,
1428 ctx->abi->inline_push_consts + offset,
1429 count);
1430 }
1431 }
1432
1433 ptr = ac_build_gep0(&ctx->ac, ctx->abi->push_constants, addr);
1434
1435 if (instr->dest.ssa.bit_size == 16) {
1436 unsigned load_dwords = instr->dest.ssa.num_components / 2 + 1;
1437 LLVMTypeRef vec_type = LLVMVectorType(LLVMInt16TypeInContext(ctx->ac.context), 2 * load_dwords);
1438 ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
1439 LLVMValueRef res = LLVMBuildLoad(ctx->ac.builder, ptr, "");
1440 res = LLVMBuildBitCast(ctx->ac.builder, res, vec_type, "");
1441 LLVMValueRef cond = LLVMBuildLShr(ctx->ac.builder, addr, ctx->ac.i32_1, "");
1442 cond = LLVMBuildTrunc(ctx->ac.builder, cond, ctx->ac.i1, "");
1443 LLVMValueRef mask[] = { LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
1444 LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
1445 LLVMConstInt(ctx->ac.i32, 4, false)};
1446 LLVMValueRef swizzle_aligned = LLVMConstVector(&mask[0], instr->dest.ssa.num_components);
1447 LLVMValueRef swizzle_unaligned = LLVMConstVector(&mask[1], instr->dest.ssa.num_components);
1448 LLVMValueRef shuffle_aligned = LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_aligned, "");
1449 LLVMValueRef shuffle_unaligned = LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_unaligned, "");
1450 res = LLVMBuildSelect(ctx->ac.builder, cond, shuffle_unaligned, shuffle_aligned, "");
1451 return LLVMBuildBitCast(ctx->ac.builder, res, get_def_type(ctx, &instr->dest.ssa), "");
1452 }
1453
1454 ptr = ac_cast_ptr(&ctx->ac, ptr, get_def_type(ctx, &instr->dest.ssa));
1455
1456 return LLVMBuildLoad(ctx->ac.builder, ptr, "");
1457 }
1458
1459 static LLVMValueRef visit_get_buffer_size(struct ac_nir_context *ctx,
1460 const nir_intrinsic_instr *instr)
1461 {
1462 LLVMValueRef index = get_src(ctx, instr->src[0]);
1463
1464 return get_buffer_size(ctx, ctx->abi->load_ssbo(ctx->abi, index, false), false);
1465 }
1466
1467 static uint32_t widen_mask(uint32_t mask, unsigned multiplier)
1468 {
1469 uint32_t new_mask = 0;
1470 for(unsigned i = 0; i < 32 && (1u << i) <= mask; ++i)
1471 if (mask & (1u << i))
1472 new_mask |= ((1u << multiplier) - 1u) << (i * multiplier);
1473 return new_mask;
1474 }
1475
1476 static LLVMValueRef extract_vector_range(struct ac_llvm_context *ctx, LLVMValueRef src,
1477 unsigned start, unsigned count)
1478 {
1479 LLVMValueRef mask[] = {
1480 ctx->i32_0, ctx->i32_1,
1481 LLVMConstInt(ctx->i32, 2, false), LLVMConstInt(ctx->i32, 3, false) };
1482
1483 unsigned src_elements = ac_get_llvm_num_components(src);
1484
1485 if (count == src_elements) {
1486 assert(start == 0);
1487 return src;
1488 } else if (count == 1) {
1489 assert(start < src_elements);
1490 return LLVMBuildExtractElement(ctx->builder, src, mask[start], "");
1491 } else {
1492 assert(start + count <= src_elements);
1493 assert(count <= 4);
1494 LLVMValueRef swizzle = LLVMConstVector(&mask[start], count);
1495 return LLVMBuildShuffleVector(ctx->builder, src, src, swizzle, "");
1496 }
1497 }
1498
1499 static unsigned get_cache_policy(struct ac_nir_context *ctx,
1500 enum gl_access_qualifier access,
1501 bool may_store_unaligned,
1502 bool writeonly_memory)
1503 {
1504 unsigned cache_policy = 0;
1505
1506 /* SI has a TC L1 bug causing corruption of 8bit/16bit stores. All
1507 * store opcodes not aligned to a dword are affected. The only way to
1508 * get unaligned stores is through shader images.
1509 */
1510 if (((may_store_unaligned && ctx->ac.chip_class == SI) ||
1511 /* If this is write-only, don't keep data in L1 to prevent
1512 * evicting L1 cache lines that may be needed by other
1513 * instructions.
1514 */
1515 writeonly_memory ||
1516 access & (ACCESS_COHERENT | ACCESS_VOLATILE))) {
1517 cache_policy |= ac_glc;
1518 }
1519
1520 return cache_policy;
1521 }
1522
1523 static void visit_store_ssbo(struct ac_nir_context *ctx,
1524 nir_intrinsic_instr *instr)
1525 {
1526 LLVMValueRef src_data = get_src(ctx, instr->src[0]);
1527 int elem_size_bytes = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src_data)) / 8;
1528 unsigned writemask = nir_intrinsic_write_mask(instr);
1529 enum gl_access_qualifier access = nir_intrinsic_access(instr);
1530 bool writeonly_memory = access & ACCESS_NON_READABLE;
1531 unsigned cache_policy = get_cache_policy(ctx, access, false, writeonly_memory);
1532
1533 LLVMValueRef rsrc = ctx->abi->load_ssbo(ctx->abi,
1534 get_src(ctx, instr->src[1]), true);
1535 LLVMValueRef base_data = ac_to_float(&ctx->ac, src_data);
1536 base_data = ac_trim_vector(&ctx->ac, base_data, instr->num_components);
1537 LLVMValueRef base_offset = get_src(ctx, instr->src[2]);
1538
1539 while (writemask) {
1540 int start, count;
1541 LLVMValueRef data, offset;
1542 LLVMTypeRef data_type;
1543
1544 u_bit_scan_consecutive_range(&writemask, &start, &count);
1545
1546 /* Due to an LLVM limitation, split 3-element writes
1547 * into a 2-element and a 1-element write. */
1548 if (count == 3) {
1549 writemask |= 1 << (start + 2);
1550 count = 2;
1551 }
1552 int num_bytes = count * elem_size_bytes; /* count in bytes */
1553
1554 /* we can only store 4 DWords at the same time.
1555 * can only happen for 64 Bit vectors. */
1556 if (num_bytes > 16) {
1557 writemask |= ((1u << (count - 2)) - 1u) << (start + 2);
1558 count = 2;
1559 num_bytes = 16;
1560 }
1561
1562 /* check alignment of 16 Bit stores */
1563 if (elem_size_bytes == 2 && num_bytes > 2 && (start % 2) == 1) {
1564 writemask |= ((1u << (count - 1)) - 1u) << (start + 1);
1565 count = 1;
1566 num_bytes = 2;
1567 }
1568 data = extract_vector_range(&ctx->ac, base_data, start, count);
1569
1570 offset = LLVMBuildAdd(ctx->ac.builder, base_offset,
1571 LLVMConstInt(ctx->ac.i32, start * elem_size_bytes, false), "");
1572
1573 if (num_bytes == 2) {
1574 ac_build_tbuffer_store_short(&ctx->ac, rsrc, data,
1575 offset, ctx->ac.i32_0,
1576 cache_policy & ac_glc,
1577 writeonly_memory);
1578 } else {
1579 int num_channels = num_bytes / 4;
1580
1581 switch (num_bytes) {
1582 case 16: /* v4f32 */
1583 data_type = ctx->ac.v4f32;
1584 break;
1585 case 8: /* v2f32 */
1586 data_type = ctx->ac.v2f32;
1587 break;
1588 case 4: /* f32 */
1589 data_type = ctx->ac.f32;
1590 break;
1591 default:
1592 unreachable("Malformed vector store.");
1593 }
1594 data = LLVMBuildBitCast(ctx->ac.builder, data, data_type, "");
1595
1596 ac_build_buffer_store_dword(&ctx->ac, rsrc, data,
1597 num_channels, offset,
1598 ctx->ac.i32_0, 0,
1599 cache_policy & ac_glc,
1600 false, writeonly_memory,
1601 false);
1602 }
1603 }
1604 }
1605
1606 static LLVMValueRef visit_atomic_ssbo(struct ac_nir_context *ctx,
1607 const nir_intrinsic_instr *instr)
1608 {
1609 const char *atomic_name;
1610 char intrinsic_name[64];
1611 LLVMValueRef params[7];
1612 int arg_count = 0;
1613 int length;
1614
1615 switch (instr->intrinsic) {
1616 case nir_intrinsic_ssbo_atomic_add:
1617 atomic_name = "add";
1618 break;
1619 case nir_intrinsic_ssbo_atomic_imin:
1620 atomic_name = "smin";
1621 break;
1622 case nir_intrinsic_ssbo_atomic_umin:
1623 atomic_name = "umin";
1624 break;
1625 case nir_intrinsic_ssbo_atomic_imax:
1626 atomic_name = "smax";
1627 break;
1628 case nir_intrinsic_ssbo_atomic_umax:
1629 atomic_name = "umax";
1630 break;
1631 case nir_intrinsic_ssbo_atomic_and:
1632 atomic_name = "and";
1633 break;
1634 case nir_intrinsic_ssbo_atomic_or:
1635 atomic_name = "or";
1636 break;
1637 case nir_intrinsic_ssbo_atomic_xor:
1638 atomic_name = "xor";
1639 break;
1640 case nir_intrinsic_ssbo_atomic_exchange:
1641 atomic_name = "swap";
1642 break;
1643 case nir_intrinsic_ssbo_atomic_comp_swap:
1644 atomic_name = "cmpswap";
1645 break;
1646 default:
1647 abort();
1648 }
1649
1650 if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap) {
1651 params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[3]), 0);
1652 }
1653 params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
1654 params[arg_count++] = ctx->abi->load_ssbo(ctx->abi,
1655 get_src(ctx, instr->src[0]),
1656 true);
1657
1658 if (HAVE_LLVM >= 0x0800) {
1659 params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
1660 params[arg_count++] = ctx->ac.i32_0; /* soffset */
1661 params[arg_count++] = ctx->ac.i32_0; /* slc */
1662
1663 length = snprintf(intrinsic_name, sizeof(intrinsic_name),
1664 "llvm.amdgcn.raw.buffer.atomic.%s.i32",
1665 atomic_name);
1666 } else {
1667 params[arg_count++] = ctx->ac.i32_0; /* vindex */
1668 params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
1669 params[arg_count++] = ctx->ac.i1false; /* slc */
1670
1671 length = snprintf(intrinsic_name, sizeof(intrinsic_name),
1672 "llvm.amdgcn.buffer.atomic.%s", atomic_name);
1673 }
1674
1675 assert(length < sizeof(intrinsic_name));
1676 return ac_build_intrinsic(&ctx->ac, intrinsic_name, ctx->ac.i32,
1677 params, arg_count, 0);
1678 }
1679
1680 static LLVMValueRef visit_load_buffer(struct ac_nir_context *ctx,
1681 const nir_intrinsic_instr *instr)
1682 {
1683 int elem_size_bytes = instr->dest.ssa.bit_size / 8;
1684 int num_components = instr->num_components;
1685 enum gl_access_qualifier access = nir_intrinsic_access(instr);
1686 unsigned cache_policy = get_cache_policy(ctx, access, false, false);
1687
1688 LLVMValueRef offset = get_src(ctx, instr->src[1]);
1689 LLVMValueRef rsrc = ctx->abi->load_ssbo(ctx->abi,
1690 get_src(ctx, instr->src[0]), false);
1691 LLVMValueRef vindex = ctx->ac.i32_0;
1692
1693 LLVMTypeRef def_type = get_def_type(ctx, &instr->dest.ssa);
1694 LLVMTypeRef def_elem_type = num_components > 1 ? LLVMGetElementType(def_type) : def_type;
1695
1696 LLVMValueRef results[4];
1697 for (int i = 0; i < num_components;) {
1698 int num_elems = num_components - i;
1699 if (elem_size_bytes < 4 && nir_intrinsic_align(instr) % 4 != 0)
1700 num_elems = 1;
1701 if (num_elems * elem_size_bytes > 16)
1702 num_elems = 16 / elem_size_bytes;
1703 int load_bytes = num_elems * elem_size_bytes;
1704
1705 LLVMValueRef immoffset = LLVMConstInt(ctx->ac.i32, i * elem_size_bytes, false);
1706
1707 LLVMValueRef ret;
1708 if (load_bytes == 2) {
1709 ret = ac_build_tbuffer_load_short(&ctx->ac,
1710 rsrc,
1711 offset,
1712 ctx->ac.i32_0,
1713 immoffset,
1714 cache_policy & ac_glc);
1715 } else {
1716 int num_channels = util_next_power_of_two(load_bytes) / 4;
1717
1718 ret = ac_build_buffer_load(&ctx->ac, rsrc, num_channels,
1719 vindex, offset, immoffset, 0,
1720 cache_policy & ac_glc, 0,
1721 false, false);
1722 }
1723
1724 LLVMTypeRef byte_vec = LLVMVectorType(ctx->ac.i8, ac_get_type_size(LLVMTypeOf(ret)));
1725 ret = LLVMBuildBitCast(ctx->ac.builder, ret, byte_vec, "");
1726 ret = ac_trim_vector(&ctx->ac, ret, load_bytes);
1727
1728 LLVMTypeRef ret_type = LLVMVectorType(def_elem_type, num_elems);
1729 ret = LLVMBuildBitCast(ctx->ac.builder, ret, ret_type, "");
1730
1731 for (unsigned j = 0; j < num_elems; j++) {
1732 results[i + j] = LLVMBuildExtractElement(ctx->ac.builder, ret, LLVMConstInt(ctx->ac.i32, j, false), "");
1733 }
1734 i += num_elems;
1735 }
1736
1737 return ac_build_gather_values(&ctx->ac, results, num_components);
1738 }
1739
1740 static LLVMValueRef visit_load_ubo_buffer(struct ac_nir_context *ctx,
1741 const nir_intrinsic_instr *instr)
1742 {
1743 LLVMValueRef ret;
1744 LLVMValueRef rsrc = get_src(ctx, instr->src[0]);
1745 LLVMValueRef offset = get_src(ctx, instr->src[1]);
1746 int num_components = instr->num_components;
1747
1748 if (ctx->abi->load_ubo)
1749 rsrc = ctx->abi->load_ubo(ctx->abi, rsrc);
1750
1751 if (instr->dest.ssa.bit_size == 64)
1752 num_components *= 2;
1753
1754 if (instr->dest.ssa.bit_size == 16) {
1755 LLVMValueRef results[num_components];
1756 for (unsigned i = 0; i < num_components; ++i) {
1757 results[i] = ac_build_tbuffer_load_short(&ctx->ac,
1758 rsrc,
1759 offset,
1760 ctx->ac.i32_0,
1761 LLVMConstInt(ctx->ac.i32, 2 * i, 0),
1762 false);
1763 }
1764 ret = ac_build_gather_values(&ctx->ac, results, num_components);
1765 } else {
1766 ret = ac_build_buffer_load(&ctx->ac, rsrc, num_components, NULL, offset,
1767 NULL, 0, false, false, true, true);
1768
1769 ret = ac_trim_vector(&ctx->ac, ret, num_components);
1770 }
1771
1772 return LLVMBuildBitCast(ctx->ac.builder, ret,
1773 get_def_type(ctx, &instr->dest.ssa), "");
1774 }
1775
1776 static void
1777 get_deref_offset(struct ac_nir_context *ctx, nir_deref_instr *instr,
1778 bool vs_in, unsigned *vertex_index_out,
1779 LLVMValueRef *vertex_index_ref,
1780 unsigned *const_out, LLVMValueRef *indir_out)
1781 {
1782 nir_variable *var = nir_deref_instr_get_variable(instr);
1783 nir_deref_path path;
1784 unsigned idx_lvl = 1;
1785
1786 nir_deref_path_init(&path, instr, NULL);
1787
1788 if (vertex_index_out != NULL || vertex_index_ref != NULL) {
1789 if (vertex_index_ref) {
1790 *vertex_index_ref = get_src(ctx, path.path[idx_lvl]->arr.index);
1791 if (vertex_index_out)
1792 *vertex_index_out = 0;
1793 } else {
1794 nir_const_value *v = nir_src_as_const_value(path.path[idx_lvl]->arr.index);
1795 assert(v);
1796 *vertex_index_out = v->u32[0];
1797 }
1798 ++idx_lvl;
1799 }
1800
1801 uint32_t const_offset = 0;
1802 LLVMValueRef offset = NULL;
1803
1804 if (var->data.compact) {
1805 assert(instr->deref_type == nir_deref_type_array);
1806 nir_const_value *v = nir_src_as_const_value(instr->arr.index);
1807 assert(v);
1808 const_offset = v->u32[0];
1809 goto out;
1810 }
1811
1812 for (; path.path[idx_lvl]; ++idx_lvl) {
1813 const struct glsl_type *parent_type = path.path[idx_lvl - 1]->type;
1814 if (path.path[idx_lvl]->deref_type == nir_deref_type_struct) {
1815 unsigned index = path.path[idx_lvl]->strct.index;
1816
1817 for (unsigned i = 0; i < index; i++) {
1818 const struct glsl_type *ft = glsl_get_struct_field(parent_type, i);
1819 const_offset += glsl_count_attribute_slots(ft, vs_in);
1820 }
1821 } else if(path.path[idx_lvl]->deref_type == nir_deref_type_array) {
1822 unsigned size = glsl_count_attribute_slots(path.path[idx_lvl]->type, vs_in);
1823 LLVMValueRef array_off = LLVMBuildMul(ctx->ac.builder, LLVMConstInt(ctx->ac.i32, size, 0),
1824 get_src(ctx, path.path[idx_lvl]->arr.index), "");
1825 if (offset)
1826 offset = LLVMBuildAdd(ctx->ac.builder, offset, array_off, "");
1827 else
1828 offset = array_off;
1829 } else
1830 unreachable("Uhandled deref type in get_deref_instr_offset");
1831 }
1832
1833 out:
1834 nir_deref_path_finish(&path);
1835
1836 if (const_offset && offset)
1837 offset = LLVMBuildAdd(ctx->ac.builder, offset,
1838 LLVMConstInt(ctx->ac.i32, const_offset, 0),
1839 "");
1840
1841 *const_out = const_offset;
1842 *indir_out = offset;
1843 }
1844
1845 static LLVMValueRef load_tess_varyings(struct ac_nir_context *ctx,
1846 nir_intrinsic_instr *instr,
1847 bool load_inputs)
1848 {
1849 LLVMValueRef result;
1850 LLVMValueRef vertex_index = NULL;
1851 LLVMValueRef indir_index = NULL;
1852 unsigned const_index = 0;
1853
1854 nir_variable *var = nir_deref_instr_get_variable(nir_instr_as_deref(instr->src[0].ssa->parent_instr));
1855
1856 unsigned location = var->data.location;
1857 unsigned driver_location = var->data.driver_location;
1858 const bool is_patch = var->data.patch;
1859 const bool is_compact = var->data.compact;
1860
1861 get_deref_offset(ctx, nir_instr_as_deref(instr->src[0].ssa->parent_instr),
1862 false, NULL, is_patch ? NULL : &vertex_index,
1863 &const_index, &indir_index);
1864
1865 LLVMTypeRef dest_type = get_def_type(ctx, &instr->dest.ssa);
1866
1867 LLVMTypeRef src_component_type;
1868 if (LLVMGetTypeKind(dest_type) == LLVMVectorTypeKind)
1869 src_component_type = LLVMGetElementType(dest_type);
1870 else
1871 src_component_type = dest_type;
1872
1873 result = ctx->abi->load_tess_varyings(ctx->abi, src_component_type,
1874 vertex_index, indir_index,
1875 const_index, location, driver_location,
1876 var->data.location_frac,
1877 instr->num_components,
1878 is_patch, is_compact, load_inputs);
1879 if (instr->dest.ssa.bit_size == 16) {
1880 result = ac_to_integer(&ctx->ac, result);
1881 result = LLVMBuildTrunc(ctx->ac.builder, result, dest_type, "");
1882 }
1883 return LLVMBuildBitCast(ctx->ac.builder, result, dest_type, "");
1884 }
1885
1886 static unsigned
1887 type_scalar_size_bytes(const struct glsl_type *type)
1888 {
1889 assert(glsl_type_is_vector_or_scalar(type) ||
1890 glsl_type_is_matrix(type));
1891 return glsl_type_is_boolean(type) ? 4 : glsl_get_bit_size(type) / 8;
1892 }
1893
1894 static LLVMValueRef visit_load_var(struct ac_nir_context *ctx,
1895 nir_intrinsic_instr *instr)
1896 {
1897 nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
1898 nir_variable *var = nir_deref_instr_get_variable(deref);
1899
1900 LLVMValueRef values[8];
1901 int idx = 0;
1902 int ve = instr->dest.ssa.num_components;
1903 unsigned comp = 0;
1904 LLVMValueRef indir_index;
1905 LLVMValueRef ret;
1906 unsigned const_index;
1907 unsigned stride = 4;
1908 int mode = deref->mode;
1909
1910 if (var) {
1911 bool vs_in = ctx->stage == MESA_SHADER_VERTEX &&
1912 var->data.mode == nir_var_shader_in;
1913 idx = var->data.driver_location;
1914 comp = var->data.location_frac;
1915 mode = var->data.mode;
1916
1917 get_deref_offset(ctx, deref, vs_in, NULL, NULL,
1918 &const_index, &indir_index);
1919
1920 if (var->data.compact) {
1921 stride = 1;
1922 const_index += comp;
1923 comp = 0;
1924 }
1925 }
1926
1927 if (instr->dest.ssa.bit_size == 64 &&
1928 (deref->mode == nir_var_shader_in ||
1929 deref->mode == nir_var_shader_out ||
1930 deref->mode == nir_var_function_temp))
1931 ve *= 2;
1932
1933 switch (mode) {
1934 case nir_var_shader_in:
1935 if (ctx->stage == MESA_SHADER_TESS_CTRL ||
1936 ctx->stage == MESA_SHADER_TESS_EVAL) {
1937 return load_tess_varyings(ctx, instr, true);
1938 }
1939
1940 if (ctx->stage == MESA_SHADER_GEOMETRY) {
1941 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
1942 LLVMValueRef indir_index;
1943 unsigned const_index, vertex_index;
1944 get_deref_offset(ctx, deref, false, &vertex_index, NULL,
1945 &const_index, &indir_index);
1946
1947 return ctx->abi->load_inputs(ctx->abi, var->data.location,
1948 var->data.driver_location,
1949 var->data.location_frac,
1950 instr->num_components, vertex_index, const_index, type);
1951 }
1952
1953 for (unsigned chan = comp; chan < ve + comp; chan++) {
1954 if (indir_index) {
1955 unsigned count = glsl_count_attribute_slots(
1956 var->type,
1957 ctx->stage == MESA_SHADER_VERTEX);
1958 count -= chan / 4;
1959 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
1960 &ctx->ac, ctx->abi->inputs + idx + chan, count,
1961 stride, false, true);
1962
1963 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
1964 tmp_vec,
1965 indir_index, "");
1966 } else
1967 values[chan] = ctx->abi->inputs[idx + chan + const_index * stride];
1968 }
1969 break;
1970 case nir_var_function_temp:
1971 for (unsigned chan = 0; chan < ve; chan++) {
1972 if (indir_index) {
1973 unsigned count = glsl_count_attribute_slots(
1974 var->type, false);
1975 count -= chan / 4;
1976 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
1977 &ctx->ac, ctx->locals + idx + chan, count,
1978 stride, true, true);
1979
1980 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
1981 tmp_vec,
1982 indir_index, "");
1983 } else {
1984 values[chan] = LLVMBuildLoad(ctx->ac.builder, ctx->locals[idx + chan + const_index * stride], "");
1985 }
1986 }
1987 break;
1988 case nir_var_mem_shared: {
1989 LLVMValueRef address = get_src(ctx, instr->src[0]);
1990 LLVMValueRef val = LLVMBuildLoad(ctx->ac.builder, address, "");
1991 return LLVMBuildBitCast(ctx->ac.builder, val,
1992 get_def_type(ctx, &instr->dest.ssa),
1993 "");
1994 }
1995 case nir_var_shader_out:
1996 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
1997 return load_tess_varyings(ctx, instr, false);
1998 }
1999
2000 for (unsigned chan = comp; chan < ve + comp; chan++) {
2001 if (indir_index) {
2002 unsigned count = glsl_count_attribute_slots(
2003 var->type, false);
2004 count -= chan / 4;
2005 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2006 &ctx->ac, ctx->abi->outputs + idx + chan, count,
2007 stride, true, true);
2008
2009 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2010 tmp_vec,
2011 indir_index, "");
2012 } else {
2013 values[chan] = LLVMBuildLoad(ctx->ac.builder,
2014 ctx->abi->outputs[idx + chan + const_index * stride],
2015 "");
2016 }
2017 }
2018 break;
2019 case nir_var_mem_global: {
2020 LLVMValueRef address = get_src(ctx, instr->src[0]);
2021 unsigned explicit_stride = glsl_get_explicit_stride(deref->type);
2022 unsigned natural_stride = type_scalar_size_bytes(deref->type);
2023 unsigned stride = explicit_stride ? explicit_stride : natural_stride;
2024
2025 LLVMTypeRef result_type = get_def_type(ctx, &instr->dest.ssa);
2026 if (stride != natural_stride) {
2027 LLVMTypeRef ptr_type = LLVMPointerType(LLVMGetElementType(result_type),
2028 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2029 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2030
2031 for (unsigned i = 0; i < instr->dest.ssa.num_components; ++i) {
2032 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, i * stride / natural_stride, 0);
2033 values[i] = LLVMBuildLoad(ctx->ac.builder,
2034 ac_build_gep_ptr(&ctx->ac, address, offset), "");
2035 }
2036 return ac_build_gather_values(&ctx->ac, values, instr->dest.ssa.num_components);
2037 } else {
2038 LLVMTypeRef ptr_type = LLVMPointerType(result_type,
2039 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2040 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2041 LLVMValueRef val = LLVMBuildLoad(ctx->ac.builder, address, "");
2042 return val;
2043 }
2044 }
2045 default:
2046 unreachable("unhandle variable mode");
2047 }
2048 ret = ac_build_varying_gather_values(&ctx->ac, values, ve, comp);
2049 return LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->dest.ssa), "");
2050 }
2051
2052 static void
2053 visit_store_var(struct ac_nir_context *ctx,
2054 nir_intrinsic_instr *instr)
2055 {
2056 nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2057 nir_variable *var = nir_deref_instr_get_variable(deref);
2058
2059 LLVMValueRef temp_ptr, value;
2060 int idx = 0;
2061 unsigned comp = 0;
2062 LLVMValueRef src = ac_to_float(&ctx->ac, get_src(ctx, instr->src[1]));
2063 int writemask = instr->const_index[0];
2064 LLVMValueRef indir_index;
2065 unsigned const_index;
2066
2067 if (var) {
2068 get_deref_offset(ctx, deref, false,
2069 NULL, NULL, &const_index, &indir_index);
2070 idx = var->data.driver_location;
2071 comp = var->data.location_frac;
2072
2073 if (var->data.compact) {
2074 const_index += comp;
2075 comp = 0;
2076 }
2077 }
2078
2079 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src)) == 64 &&
2080 (deref->mode == nir_var_shader_out ||
2081 deref->mode == nir_var_function_temp)) {
2082
2083 src = LLVMBuildBitCast(ctx->ac.builder, src,
2084 LLVMVectorType(ctx->ac.f32, ac_get_llvm_num_components(src) * 2),
2085 "");
2086
2087 writemask = widen_mask(writemask, 2);
2088 }
2089
2090 writemask = writemask << comp;
2091
2092 switch (deref->mode) {
2093 case nir_var_shader_out:
2094
2095 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
2096 LLVMValueRef vertex_index = NULL;
2097 LLVMValueRef indir_index = NULL;
2098 unsigned const_index = 0;
2099 const bool is_patch = var->data.patch;
2100
2101 get_deref_offset(ctx, deref, false, NULL,
2102 is_patch ? NULL : &vertex_index,
2103 &const_index, &indir_index);
2104
2105 ctx->abi->store_tcs_outputs(ctx->abi, var,
2106 vertex_index, indir_index,
2107 const_index, src, writemask);
2108 return;
2109 }
2110
2111 for (unsigned chan = 0; chan < 8; chan++) {
2112 int stride = 4;
2113 if (!(writemask & (1 << chan)))
2114 continue;
2115
2116 value = ac_llvm_extract_elem(&ctx->ac, src, chan - comp);
2117
2118 if (var->data.compact)
2119 stride = 1;
2120 if (indir_index) {
2121 unsigned count = glsl_count_attribute_slots(
2122 var->type, false);
2123 count -= chan / 4;
2124 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2125 &ctx->ac, ctx->abi->outputs + idx + chan, count,
2126 stride, true, true);
2127
2128 tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
2129 value, indir_index, "");
2130 build_store_values_extended(&ctx->ac, ctx->abi->outputs + idx + chan,
2131 count, stride, tmp_vec);
2132
2133 } else {
2134 temp_ptr = ctx->abi->outputs[idx + chan + const_index * stride];
2135
2136 LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
2137 }
2138 }
2139 break;
2140 case nir_var_function_temp:
2141 for (unsigned chan = 0; chan < 8; chan++) {
2142 if (!(writemask & (1 << chan)))
2143 continue;
2144
2145 value = ac_llvm_extract_elem(&ctx->ac, src, chan);
2146 if (indir_index) {
2147 unsigned count = glsl_count_attribute_slots(
2148 var->type, false);
2149 count -= chan / 4;
2150 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2151 &ctx->ac, ctx->locals + idx + chan, count,
2152 4, true, true);
2153
2154 tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
2155 value, indir_index, "");
2156 build_store_values_extended(&ctx->ac, ctx->locals + idx + chan,
2157 count, 4, tmp_vec);
2158 } else {
2159 temp_ptr = ctx->locals[idx + chan + const_index * 4];
2160
2161 LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
2162 }
2163 }
2164 break;
2165
2166 case nir_var_mem_global:
2167 case nir_var_mem_shared: {
2168 int writemask = instr->const_index[0];
2169 LLVMValueRef address = get_src(ctx, instr->src[0]);
2170 LLVMValueRef val = get_src(ctx, instr->src[1]);
2171
2172 unsigned explicit_stride = glsl_get_explicit_stride(deref->type);
2173 unsigned natural_stride = type_scalar_size_bytes(deref->type);
2174 unsigned stride = explicit_stride ? explicit_stride : natural_stride;
2175
2176 LLVMTypeRef ptr_type = LLVMPointerType(LLVMTypeOf(val),
2177 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2178 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2179
2180 if (writemask == (1u << ac_get_llvm_num_components(val)) - 1 &&
2181 stride == natural_stride) {
2182 LLVMTypeRef ptr_type = LLVMPointerType(LLVMTypeOf(val),
2183 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2184 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2185
2186 val = LLVMBuildBitCast(ctx->ac.builder, val,
2187 LLVMGetElementType(LLVMTypeOf(address)), "");
2188 LLVMBuildStore(ctx->ac.builder, val, address);
2189 } else {
2190 LLVMTypeRef ptr_type = LLVMPointerType(LLVMGetElementType(LLVMTypeOf(val)),
2191 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2192 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2193 for (unsigned chan = 0; chan < 4; chan++) {
2194 if (!(writemask & (1 << chan)))
2195 continue;
2196
2197 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, chan * stride / natural_stride, 0);
2198
2199 LLVMValueRef ptr = ac_build_gep_ptr(&ctx->ac, address, offset);
2200 LLVMValueRef src = ac_llvm_extract_elem(&ctx->ac, val,
2201 chan);
2202 src = LLVMBuildBitCast(ctx->ac.builder, src,
2203 LLVMGetElementType(LLVMTypeOf(ptr)), "");
2204 LLVMBuildStore(ctx->ac.builder, src, ptr);
2205 }
2206 }
2207 break;
2208 }
2209 default:
2210 abort();
2211 break;
2212 }
2213 }
2214
2215 static int image_type_to_components_count(enum glsl_sampler_dim dim, bool array)
2216 {
2217 switch (dim) {
2218 case GLSL_SAMPLER_DIM_BUF:
2219 return 1;
2220 case GLSL_SAMPLER_DIM_1D:
2221 return array ? 2 : 1;
2222 case GLSL_SAMPLER_DIM_2D:
2223 return array ? 3 : 2;
2224 case GLSL_SAMPLER_DIM_MS:
2225 return array ? 4 : 3;
2226 case GLSL_SAMPLER_DIM_3D:
2227 case GLSL_SAMPLER_DIM_CUBE:
2228 return 3;
2229 case GLSL_SAMPLER_DIM_RECT:
2230 case GLSL_SAMPLER_DIM_SUBPASS:
2231 return 2;
2232 case GLSL_SAMPLER_DIM_SUBPASS_MS:
2233 return 3;
2234 default:
2235 break;
2236 }
2237 return 0;
2238 }
2239
2240
2241 /* Adjust the sample index according to FMASK.
2242 *
2243 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
2244 * which is the identity mapping. Each nibble says which physical sample
2245 * should be fetched to get that sample.
2246 *
2247 * For example, 0x11111100 means there are only 2 samples stored and
2248 * the second sample covers 3/4 of the pixel. When reading samples 0
2249 * and 1, return physical sample 0 (determined by the first two 0s
2250 * in FMASK), otherwise return physical sample 1.
2251 *
2252 * The sample index should be adjusted as follows:
2253 * sample_index = (fmask >> (sample_index * 4)) & 0xF;
2254 */
2255 static LLVMValueRef adjust_sample_index_using_fmask(struct ac_llvm_context *ctx,
2256 LLVMValueRef coord_x, LLVMValueRef coord_y,
2257 LLVMValueRef coord_z,
2258 LLVMValueRef sample_index,
2259 LLVMValueRef fmask_desc_ptr)
2260 {
2261 struct ac_image_args args = {0};
2262 LLVMValueRef res;
2263
2264 args.coords[0] = coord_x;
2265 args.coords[1] = coord_y;
2266 if (coord_z)
2267 args.coords[2] = coord_z;
2268
2269 args.opcode = ac_image_load;
2270 args.dim = coord_z ? ac_image_2darray : ac_image_2d;
2271 args.resource = fmask_desc_ptr;
2272 args.dmask = 0xf;
2273 args.attributes = AC_FUNC_ATTR_READNONE;
2274
2275 res = ac_build_image_opcode(ctx, &args);
2276
2277 res = ac_to_integer(ctx, res);
2278 LLVMValueRef four = LLVMConstInt(ctx->i32, 4, false);
2279 LLVMValueRef F = LLVMConstInt(ctx->i32, 0xf, false);
2280
2281 LLVMValueRef fmask = LLVMBuildExtractElement(ctx->builder,
2282 res,
2283 ctx->i32_0, "");
2284
2285 LLVMValueRef sample_index4 =
2286 LLVMBuildMul(ctx->builder, sample_index, four, "");
2287 LLVMValueRef shifted_fmask =
2288 LLVMBuildLShr(ctx->builder, fmask, sample_index4, "");
2289 LLVMValueRef final_sample =
2290 LLVMBuildAnd(ctx->builder, shifted_fmask, F, "");
2291
2292 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
2293 * resource descriptor is 0 (invalid),
2294 */
2295 LLVMValueRef fmask_desc =
2296 LLVMBuildBitCast(ctx->builder, fmask_desc_ptr,
2297 ctx->v8i32, "");
2298
2299 LLVMValueRef fmask_word1 =
2300 LLVMBuildExtractElement(ctx->builder, fmask_desc,
2301 ctx->i32_1, "");
2302
2303 LLVMValueRef word1_is_nonzero =
2304 LLVMBuildICmp(ctx->builder, LLVMIntNE,
2305 fmask_word1, ctx->i32_0, "");
2306
2307 /* Replace the MSAA sample index. */
2308 sample_index =
2309 LLVMBuildSelect(ctx->builder, word1_is_nonzero,
2310 final_sample, sample_index, "");
2311 return sample_index;
2312 }
2313
2314 static nir_deref_instr *get_image_deref(const nir_intrinsic_instr *instr)
2315 {
2316 assert(instr->src[0].is_ssa);
2317 return nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2318 }
2319
2320 static LLVMValueRef get_image_descriptor(struct ac_nir_context *ctx,
2321 const nir_intrinsic_instr *instr,
2322 enum ac_descriptor_type desc_type,
2323 bool write)
2324 {
2325 return get_sampler_desc(ctx, nir_instr_as_deref(instr->src[0].ssa->parent_instr), desc_type, NULL, true, write);
2326 }
2327
2328 static void get_image_coords(struct ac_nir_context *ctx,
2329 const nir_intrinsic_instr *instr,
2330 struct ac_image_args *args)
2331 {
2332 const struct glsl_type *type = get_image_deref(instr)->type;
2333
2334 LLVMValueRef src0 = get_src(ctx, instr->src[1]);
2335 LLVMValueRef masks[] = {
2336 LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
2337 LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
2338 };
2339 LLVMValueRef sample_index = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
2340
2341 int count;
2342 enum glsl_sampler_dim dim = glsl_get_sampler_dim(type);
2343 bool is_array = glsl_sampler_type_is_array(type);
2344 bool add_frag_pos = (dim == GLSL_SAMPLER_DIM_SUBPASS ||
2345 dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2346 bool is_ms = (dim == GLSL_SAMPLER_DIM_MS ||
2347 dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2348 bool gfx9_1d = ctx->ac.chip_class >= GFX9 && dim == GLSL_SAMPLER_DIM_1D;
2349 count = image_type_to_components_count(dim, is_array);
2350
2351 if (is_ms && instr->intrinsic == nir_intrinsic_image_deref_load) {
2352 LLVMValueRef fmask_load_address[3];
2353 int chan;
2354
2355 fmask_load_address[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
2356 fmask_load_address[1] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[1], "");
2357 if (is_array)
2358 fmask_load_address[2] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[2], "");
2359 else
2360 fmask_load_address[2] = NULL;
2361 if (add_frag_pos) {
2362 for (chan = 0; chan < 2; ++chan)
2363 fmask_load_address[chan] =
2364 LLVMBuildAdd(ctx->ac.builder, fmask_load_address[chan],
2365 LLVMBuildFPToUI(ctx->ac.builder, ctx->abi->frag_pos[chan],
2366 ctx->ac.i32, ""), "");
2367 fmask_load_address[2] = ac_to_integer(&ctx->ac, ctx->abi->inputs[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER, 0)]);
2368 }
2369 sample_index = adjust_sample_index_using_fmask(&ctx->ac,
2370 fmask_load_address[0],
2371 fmask_load_address[1],
2372 fmask_load_address[2],
2373 sample_index,
2374 get_sampler_desc(ctx, nir_instr_as_deref(instr->src[0].ssa->parent_instr),
2375 AC_DESC_FMASK, NULL, false, false));
2376 }
2377 if (count == 1 && !gfx9_1d) {
2378 if (instr->src[1].ssa->num_components)
2379 args->coords[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
2380 else
2381 args->coords[0] = src0;
2382 } else {
2383 int chan;
2384 if (is_ms)
2385 count--;
2386 for (chan = 0; chan < count; ++chan) {
2387 args->coords[chan] = ac_llvm_extract_elem(&ctx->ac, src0, chan);
2388 }
2389 if (add_frag_pos) {
2390 for (chan = 0; chan < 2; ++chan) {
2391 args->coords[chan] = LLVMBuildAdd(
2392 ctx->ac.builder, args->coords[chan],
2393 LLVMBuildFPToUI(
2394 ctx->ac.builder, ctx->abi->frag_pos[chan],
2395 ctx->ac.i32, ""), "");
2396 }
2397 args->coords[2] = ac_to_integer(&ctx->ac,
2398 ctx->abi->inputs[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER, 0)]);
2399 count++;
2400 }
2401
2402 if (gfx9_1d) {
2403 if (is_array) {
2404 args->coords[2] = args->coords[1];
2405 args->coords[1] = ctx->ac.i32_0;
2406 } else
2407 args->coords[1] = ctx->ac.i32_0;
2408 count++;
2409 }
2410
2411 if (is_ms) {
2412 args->coords[count] = sample_index;
2413 count++;
2414 }
2415 }
2416 }
2417
2418 static LLVMValueRef get_image_buffer_descriptor(struct ac_nir_context *ctx,
2419 const nir_intrinsic_instr *instr, bool write)
2420 {
2421 LLVMValueRef rsrc = get_image_descriptor(ctx, instr, AC_DESC_BUFFER, write);
2422 if (ctx->abi->gfx9_stride_size_workaround) {
2423 LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 2, 0), "");
2424 LLVMValueRef stride = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 1, 0), "");
2425 stride = LLVMBuildLShr(ctx->ac.builder, stride, LLVMConstInt(ctx->ac.i32, 16, 0), "");
2426
2427 LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->ac.builder,
2428 LLVMBuildICmp(ctx->ac.builder, LLVMIntUGT, elem_count, stride, ""),
2429 elem_count, stride, "");
2430
2431 rsrc = LLVMBuildInsertElement(ctx->ac.builder, rsrc, new_elem_count,
2432 LLVMConstInt(ctx->ac.i32, 2, 0), "");
2433 }
2434 return rsrc;
2435 }
2436
2437 static LLVMValueRef visit_image_load(struct ac_nir_context *ctx,
2438 const nir_intrinsic_instr *instr)
2439 {
2440 LLVMValueRef res;
2441 const nir_deref_instr *image_deref = get_image_deref(instr);
2442 const struct glsl_type *type = image_deref->type;
2443 const nir_variable *var = nir_deref_instr_get_variable(image_deref);
2444 struct ac_image_args args = {};
2445
2446 args.cache_policy =
2447 get_cache_policy(ctx, var->data.image.access, false, false);
2448
2449 const enum glsl_sampler_dim dim = glsl_get_sampler_dim(type);
2450 if (dim == GLSL_SAMPLER_DIM_BUF) {
2451 unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
2452 unsigned num_channels = util_last_bit(mask);
2453 LLVMValueRef rsrc, vindex;
2454
2455 rsrc = get_image_buffer_descriptor(ctx, instr, false);
2456 vindex = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
2457 ctx->ac.i32_0, "");
2458
2459 /* TODO: set "can_speculate" when OpenGL needs it. */
2460 res = ac_build_buffer_load_format(&ctx->ac, rsrc, vindex,
2461 ctx->ac.i32_0, num_channels,
2462 !!(args.cache_policy & ac_glc),
2463 false);
2464 res = ac_build_expand_to_vec4(&ctx->ac, res, num_channels);
2465
2466 res = ac_trim_vector(&ctx->ac, res, instr->dest.ssa.num_components);
2467 res = ac_to_integer(&ctx->ac, res);
2468 } else {
2469 args.opcode = ac_image_load;
2470 get_image_coords(ctx, instr, &args);
2471 args.resource = get_image_descriptor(ctx, instr, AC_DESC_IMAGE, false);
2472 args.dim = get_ac_image_dim(&ctx->ac, glsl_get_sampler_dim(type),
2473 glsl_sampler_type_is_array(type));
2474 args.dmask = 15;
2475 args.attributes = AC_FUNC_ATTR_READONLY;
2476
2477 res = ac_build_image_opcode(&ctx->ac, &args);
2478 }
2479 return ac_to_integer(&ctx->ac, res);
2480 }
2481
2482 static void visit_image_store(struct ac_nir_context *ctx,
2483 nir_intrinsic_instr *instr)
2484 {
2485 const nir_deref_instr *image_deref = get_image_deref(instr);
2486 const struct glsl_type *type = image_deref->type;
2487 const nir_variable *var = nir_deref_instr_get_variable(image_deref);
2488 const enum glsl_sampler_dim dim = glsl_get_sampler_dim(type);
2489 bool writeonly_memory = var->data.image.access & ACCESS_NON_READABLE;
2490 struct ac_image_args args = {};
2491
2492 args.cache_policy = get_cache_policy(ctx, var->data.image.access, true,
2493 writeonly_memory);
2494
2495 if (dim == GLSL_SAMPLER_DIM_BUF) {
2496 LLVMValueRef rsrc = get_image_buffer_descriptor(ctx, instr, true);
2497 LLVMValueRef src = ac_to_float(&ctx->ac, get_src(ctx, instr->src[3]));
2498 unsigned src_channels = ac_get_llvm_num_components(src);
2499 LLVMValueRef vindex;
2500
2501 if (src_channels == 3)
2502 src = ac_build_expand_to_vec4(&ctx->ac, src, 3);
2503
2504 vindex = LLVMBuildExtractElement(ctx->ac.builder,
2505 get_src(ctx, instr->src[1]),
2506 ctx->ac.i32_0, "");
2507
2508 ac_build_buffer_store_format(&ctx->ac, rsrc, src, vindex,
2509 ctx->ac.i32_0, src_channels,
2510 args.cache_policy & ac_glc,
2511 writeonly_memory);
2512 } else {
2513 args.opcode = ac_image_store;
2514 args.data[0] = ac_to_float(&ctx->ac, get_src(ctx, instr->src[3]));
2515 get_image_coords(ctx, instr, &args);
2516 args.resource = get_image_descriptor(ctx, instr, AC_DESC_IMAGE, true);
2517 args.dim = get_ac_image_dim(&ctx->ac, glsl_get_sampler_dim(type),
2518 glsl_sampler_type_is_array(type));
2519 args.dmask = 15;
2520
2521 ac_build_image_opcode(&ctx->ac, &args);
2522 }
2523
2524 }
2525
2526 static LLVMValueRef visit_image_atomic(struct ac_nir_context *ctx,
2527 const nir_intrinsic_instr *instr)
2528 {
2529 LLVMValueRef params[7];
2530 int param_count = 0;
2531 const struct glsl_type *type = get_image_deref(instr)->type;
2532
2533 bool cmpswap = instr->intrinsic == nir_intrinsic_image_deref_atomic_comp_swap;
2534 const char *atomic_name;
2535 char intrinsic_name[64];
2536 enum ac_atomic_op atomic_subop;
2537 MAYBE_UNUSED int length;
2538
2539 bool is_unsigned = glsl_get_sampler_result_type(type) == GLSL_TYPE_UINT;
2540
2541 switch (instr->intrinsic) {
2542 case nir_intrinsic_image_deref_atomic_add:
2543 atomic_name = "add";
2544 atomic_subop = ac_atomic_add;
2545 break;
2546 case nir_intrinsic_image_deref_atomic_min:
2547 atomic_name = is_unsigned ? "umin" : "smin";
2548 atomic_subop = is_unsigned ? ac_atomic_umin : ac_atomic_smin;
2549 break;
2550 case nir_intrinsic_image_deref_atomic_max:
2551 atomic_name = is_unsigned ? "umax" : "smax";
2552 atomic_subop = is_unsigned ? ac_atomic_umax : ac_atomic_smax;
2553 break;
2554 case nir_intrinsic_image_deref_atomic_and:
2555 atomic_name = "and";
2556 atomic_subop = ac_atomic_and;
2557 break;
2558 case nir_intrinsic_image_deref_atomic_or:
2559 atomic_name = "or";
2560 atomic_subop = ac_atomic_or;
2561 break;
2562 case nir_intrinsic_image_deref_atomic_xor:
2563 atomic_name = "xor";
2564 atomic_subop = ac_atomic_xor;
2565 break;
2566 case nir_intrinsic_image_deref_atomic_exchange:
2567 atomic_name = "swap";
2568 atomic_subop = ac_atomic_swap;
2569 break;
2570 case nir_intrinsic_image_deref_atomic_comp_swap:
2571 atomic_name = "cmpswap";
2572 atomic_subop = 0; /* not used */
2573 break;
2574 default:
2575 abort();
2576 }
2577
2578 if (cmpswap)
2579 params[param_count++] = get_src(ctx, instr->src[4]);
2580 params[param_count++] = get_src(ctx, instr->src[3]);
2581
2582 if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_BUF) {
2583 params[param_count++] = get_image_buffer_descriptor(ctx, instr, true);
2584 params[param_count++] = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
2585 ctx->ac.i32_0, ""); /* vindex */
2586 params[param_count++] = ctx->ac.i32_0; /* voffset */
2587 if (HAVE_LLVM >= 0x800) {
2588 params[param_count++] = ctx->ac.i32_0; /* soffset */
2589 params[param_count++] = ctx->ac.i32_0; /* slc */
2590
2591 length = snprintf(intrinsic_name, sizeof(intrinsic_name),
2592 "llvm.amdgcn.struct.buffer.atomic.%s.i32", atomic_name);
2593 } else {
2594 params[param_count++] = ctx->ac.i1false; /* slc */
2595
2596 length = snprintf(intrinsic_name, sizeof(intrinsic_name),
2597 "llvm.amdgcn.buffer.atomic.%s", atomic_name);
2598 }
2599
2600 assert(length < sizeof(intrinsic_name));
2601 return ac_build_intrinsic(&ctx->ac, intrinsic_name, ctx->ac.i32,
2602 params, param_count, 0);
2603 } else {
2604 struct ac_image_args args = {};
2605 args.opcode = cmpswap ? ac_image_atomic_cmpswap : ac_image_atomic;
2606 args.atomic = atomic_subop;
2607 args.data[0] = params[0];
2608 if (cmpswap)
2609 args.data[1] = params[1];
2610 get_image_coords(ctx, instr, &args);
2611 args.resource = get_image_descriptor(ctx, instr, AC_DESC_IMAGE, true);
2612 args.dim = get_ac_image_dim(&ctx->ac, glsl_get_sampler_dim(type),
2613 glsl_sampler_type_is_array(type));
2614
2615 return ac_build_image_opcode(&ctx->ac, &args);
2616 }
2617 }
2618
2619 static LLVMValueRef visit_image_samples(struct ac_nir_context *ctx,
2620 const nir_intrinsic_instr *instr)
2621 {
2622 const struct glsl_type *type = get_image_deref(instr)->type;
2623
2624 struct ac_image_args args = { 0 };
2625 args.dim = get_ac_sampler_dim(&ctx->ac, glsl_get_sampler_dim(type),
2626 glsl_sampler_type_is_array(type));
2627 args.dmask = 0xf;
2628 args.resource = get_image_descriptor(ctx, instr, AC_DESC_IMAGE, false);
2629 args.opcode = ac_image_get_resinfo;
2630 args.lod = ctx->ac.i32_0;
2631 args.attributes = AC_FUNC_ATTR_READNONE;
2632
2633 return ac_build_image_opcode(&ctx->ac, &args);
2634 }
2635
2636 static LLVMValueRef visit_image_size(struct ac_nir_context *ctx,
2637 const nir_intrinsic_instr *instr)
2638 {
2639 LLVMValueRef res;
2640 const struct glsl_type *type = get_image_deref(instr)->type;
2641
2642 if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_BUF)
2643 return get_buffer_size(ctx, get_image_descriptor(ctx, instr, AC_DESC_BUFFER, false), true);
2644
2645 struct ac_image_args args = { 0 };
2646
2647 args.dim = get_ac_image_dim(&ctx->ac, glsl_get_sampler_dim(type),
2648 glsl_sampler_type_is_array(type));
2649 args.dmask = 0xf;
2650 args.resource = get_image_descriptor(ctx, instr, AC_DESC_IMAGE, false);
2651 args.opcode = ac_image_get_resinfo;
2652 args.lod = ctx->ac.i32_0;
2653 args.attributes = AC_FUNC_ATTR_READNONE;
2654
2655 res = ac_build_image_opcode(&ctx->ac, &args);
2656
2657 LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
2658
2659 if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_CUBE &&
2660 glsl_sampler_type_is_array(type)) {
2661 LLVMValueRef six = LLVMConstInt(ctx->ac.i32, 6, false);
2662 LLVMValueRef z = LLVMBuildExtractElement(ctx->ac.builder, res, two, "");
2663 z = LLVMBuildSDiv(ctx->ac.builder, z, six, "");
2664 res = LLVMBuildInsertElement(ctx->ac.builder, res, z, two, "");
2665 }
2666 if (ctx->ac.chip_class >= GFX9 &&
2667 glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_1D &&
2668 glsl_sampler_type_is_array(type)) {
2669 LLVMValueRef layers = LLVMBuildExtractElement(ctx->ac.builder, res, two, "");
2670 res = LLVMBuildInsertElement(ctx->ac.builder, res, layers,
2671 ctx->ac.i32_1, "");
2672
2673 }
2674 return res;
2675 }
2676
2677 static void emit_membar(struct ac_llvm_context *ac,
2678 const nir_intrinsic_instr *instr)
2679 {
2680 unsigned waitcnt = NOOP_WAITCNT;
2681
2682 switch (instr->intrinsic) {
2683 case nir_intrinsic_memory_barrier:
2684 case nir_intrinsic_group_memory_barrier:
2685 waitcnt &= VM_CNT & LGKM_CNT;
2686 break;
2687 case nir_intrinsic_memory_barrier_atomic_counter:
2688 case nir_intrinsic_memory_barrier_buffer:
2689 case nir_intrinsic_memory_barrier_image:
2690 waitcnt &= VM_CNT;
2691 break;
2692 case nir_intrinsic_memory_barrier_shared:
2693 waitcnt &= LGKM_CNT;
2694 break;
2695 default:
2696 break;
2697 }
2698 if (waitcnt != NOOP_WAITCNT)
2699 ac_build_waitcnt(ac, waitcnt);
2700 }
2701
2702 void ac_emit_barrier(struct ac_llvm_context *ac, gl_shader_stage stage)
2703 {
2704 /* SI only (thanks to a hw bug workaround):
2705 * The real barrier instruction isn’t needed, because an entire patch
2706 * always fits into a single wave.
2707 */
2708 if (ac->chip_class == SI && stage == MESA_SHADER_TESS_CTRL) {
2709 ac_build_waitcnt(ac, LGKM_CNT & VM_CNT);
2710 return;
2711 }
2712 ac_build_s_barrier(ac);
2713 }
2714
2715 static void emit_discard(struct ac_nir_context *ctx,
2716 const nir_intrinsic_instr *instr)
2717 {
2718 LLVMValueRef cond;
2719
2720 if (instr->intrinsic == nir_intrinsic_discard_if) {
2721 cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
2722 get_src(ctx, instr->src[0]),
2723 ctx->ac.i32_0, "");
2724 } else {
2725 assert(instr->intrinsic == nir_intrinsic_discard);
2726 cond = ctx->ac.i1false;
2727 }
2728
2729 ctx->abi->emit_kill(ctx->abi, cond);
2730 }
2731
2732 static LLVMValueRef
2733 visit_load_helper_invocation(struct ac_nir_context *ctx)
2734 {
2735 LLVMValueRef result = ac_build_intrinsic(&ctx->ac,
2736 "llvm.amdgcn.ps.live",
2737 ctx->ac.i1, NULL, 0,
2738 AC_FUNC_ATTR_READNONE);
2739 result = LLVMBuildNot(ctx->ac.builder, result, "");
2740 return LLVMBuildSExt(ctx->ac.builder, result, ctx->ac.i32, "");
2741 }
2742
2743 static LLVMValueRef
2744 visit_load_local_invocation_index(struct ac_nir_context *ctx)
2745 {
2746 LLVMValueRef result;
2747 LLVMValueRef thread_id = ac_get_thread_id(&ctx->ac);
2748 result = LLVMBuildAnd(ctx->ac.builder, ctx->abi->tg_size,
2749 LLVMConstInt(ctx->ac.i32, 0xfc0, false), "");
2750
2751 return LLVMBuildAdd(ctx->ac.builder, result, thread_id, "");
2752 }
2753
2754 static LLVMValueRef
2755 visit_load_subgroup_id(struct ac_nir_context *ctx)
2756 {
2757 if (ctx->stage == MESA_SHADER_COMPUTE) {
2758 LLVMValueRef result;
2759 result = LLVMBuildAnd(ctx->ac.builder, ctx->abi->tg_size,
2760 LLVMConstInt(ctx->ac.i32, 0xfc0, false), "");
2761 return LLVMBuildLShr(ctx->ac.builder, result, LLVMConstInt(ctx->ac.i32, 6, false), "");
2762 } else {
2763 return LLVMConstInt(ctx->ac.i32, 0, false);
2764 }
2765 }
2766
2767 static LLVMValueRef
2768 visit_load_num_subgroups(struct ac_nir_context *ctx)
2769 {
2770 if (ctx->stage == MESA_SHADER_COMPUTE) {
2771 return LLVMBuildAnd(ctx->ac.builder, ctx->abi->tg_size,
2772 LLVMConstInt(ctx->ac.i32, 0x3f, false), "");
2773 } else {
2774 return LLVMConstInt(ctx->ac.i32, 1, false);
2775 }
2776 }
2777
2778 static LLVMValueRef
2779 visit_first_invocation(struct ac_nir_context *ctx)
2780 {
2781 LLVMValueRef active_set = ac_build_ballot(&ctx->ac, ctx->ac.i32_1);
2782
2783 /* The second argument is whether cttz(0) should be defined, but we do not care. */
2784 LLVMValueRef args[] = {active_set, ctx->ac.i1false};
2785 LLVMValueRef result = ac_build_intrinsic(&ctx->ac,
2786 "llvm.cttz.i64",
2787 ctx->ac.i64, args, 2,
2788 AC_FUNC_ATTR_NOUNWIND |
2789 AC_FUNC_ATTR_READNONE);
2790
2791 return LLVMBuildTrunc(ctx->ac.builder, result, ctx->ac.i32, "");
2792 }
2793
2794 static LLVMValueRef
2795 visit_load_shared(struct ac_nir_context *ctx,
2796 const nir_intrinsic_instr *instr)
2797 {
2798 LLVMValueRef values[4], derived_ptr, index, ret;
2799
2800 LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[0]);
2801
2802 for (int chan = 0; chan < instr->num_components; chan++) {
2803 index = LLVMConstInt(ctx->ac.i32, chan, 0);
2804 derived_ptr = LLVMBuildGEP(ctx->ac.builder, ptr, &index, 1, "");
2805 values[chan] = LLVMBuildLoad(ctx->ac.builder, derived_ptr, "");
2806 }
2807
2808 ret = ac_build_gather_values(&ctx->ac, values, instr->num_components);
2809 return LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->dest.ssa), "");
2810 }
2811
2812 static void
2813 visit_store_shared(struct ac_nir_context *ctx,
2814 const nir_intrinsic_instr *instr)
2815 {
2816 LLVMValueRef derived_ptr, data,index;
2817 LLVMBuilderRef builder = ctx->ac.builder;
2818
2819 LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[1]);
2820 LLVMValueRef src = get_src(ctx, instr->src[0]);
2821
2822 int writemask = nir_intrinsic_write_mask(instr);
2823 for (int chan = 0; chan < 4; chan++) {
2824 if (!(writemask & (1 << chan))) {
2825 continue;
2826 }
2827 data = ac_llvm_extract_elem(&ctx->ac, src, chan);
2828 index = LLVMConstInt(ctx->ac.i32, chan, 0);
2829 derived_ptr = LLVMBuildGEP(builder, ptr, &index, 1, "");
2830 LLVMBuildStore(builder, data, derived_ptr);
2831 }
2832 }
2833
2834 static LLVMValueRef visit_var_atomic(struct ac_nir_context *ctx,
2835 const nir_intrinsic_instr *instr,
2836 LLVMValueRef ptr, int src_idx)
2837 {
2838 LLVMValueRef result;
2839 LLVMValueRef src = get_src(ctx, instr->src[src_idx]);
2840
2841 if (instr->intrinsic == nir_intrinsic_shared_atomic_comp_swap ||
2842 instr->intrinsic == nir_intrinsic_deref_atomic_comp_swap) {
2843 LLVMValueRef src1 = get_src(ctx, instr->src[src_idx + 1]);
2844 result = LLVMBuildAtomicCmpXchg(ctx->ac.builder,
2845 ptr, src, src1,
2846 LLVMAtomicOrderingSequentiallyConsistent,
2847 LLVMAtomicOrderingSequentiallyConsistent,
2848 false);
2849 result = LLVMBuildExtractValue(ctx->ac.builder, result, 0, "");
2850 } else {
2851 LLVMAtomicRMWBinOp op;
2852 switch (instr->intrinsic) {
2853 case nir_intrinsic_shared_atomic_add:
2854 case nir_intrinsic_deref_atomic_add:
2855 op = LLVMAtomicRMWBinOpAdd;
2856 break;
2857 case nir_intrinsic_shared_atomic_umin:
2858 case nir_intrinsic_deref_atomic_umin:
2859 op = LLVMAtomicRMWBinOpUMin;
2860 break;
2861 case nir_intrinsic_shared_atomic_umax:
2862 case nir_intrinsic_deref_atomic_umax:
2863 op = LLVMAtomicRMWBinOpUMax;
2864 break;
2865 case nir_intrinsic_shared_atomic_imin:
2866 case nir_intrinsic_deref_atomic_imin:
2867 op = LLVMAtomicRMWBinOpMin;
2868 break;
2869 case nir_intrinsic_shared_atomic_imax:
2870 case nir_intrinsic_deref_atomic_imax:
2871 op = LLVMAtomicRMWBinOpMax;
2872 break;
2873 case nir_intrinsic_shared_atomic_and:
2874 case nir_intrinsic_deref_atomic_and:
2875 op = LLVMAtomicRMWBinOpAnd;
2876 break;
2877 case nir_intrinsic_shared_atomic_or:
2878 case nir_intrinsic_deref_atomic_or:
2879 op = LLVMAtomicRMWBinOpOr;
2880 break;
2881 case nir_intrinsic_shared_atomic_xor:
2882 case nir_intrinsic_deref_atomic_xor:
2883 op = LLVMAtomicRMWBinOpXor;
2884 break;
2885 case nir_intrinsic_shared_atomic_exchange:
2886 case nir_intrinsic_deref_atomic_exchange:
2887 op = LLVMAtomicRMWBinOpXchg;
2888 break;
2889 default:
2890 return NULL;
2891 }
2892
2893 result = LLVMBuildAtomicRMW(ctx->ac.builder, op, ptr, ac_to_integer(&ctx->ac, src),
2894 LLVMAtomicOrderingSequentiallyConsistent,
2895 false);
2896 }
2897 return result;
2898 }
2899
2900 static LLVMValueRef load_sample_pos(struct ac_nir_context *ctx)
2901 {
2902 LLVMValueRef values[2];
2903 LLVMValueRef pos[2];
2904
2905 pos[0] = ac_to_float(&ctx->ac, ctx->abi->frag_pos[0]);
2906 pos[1] = ac_to_float(&ctx->ac, ctx->abi->frag_pos[1]);
2907
2908 values[0] = ac_build_fract(&ctx->ac, pos[0], 32);
2909 values[1] = ac_build_fract(&ctx->ac, pos[1], 32);
2910 return ac_build_gather_values(&ctx->ac, values, 2);
2911 }
2912
2913 static LLVMValueRef visit_interp(struct ac_nir_context *ctx,
2914 const nir_intrinsic_instr *instr)
2915 {
2916 LLVMValueRef result[4];
2917 LLVMValueRef interp_param;
2918 unsigned location;
2919 unsigned chan;
2920 LLVMValueRef src_c0 = NULL;
2921 LLVMValueRef src_c1 = NULL;
2922 LLVMValueRef src0 = NULL;
2923
2924 nir_deref_instr *deref_instr = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2925 nir_variable *var = nir_deref_instr_get_variable(deref_instr);
2926 int input_base = ctx->abi->fs_input_attr_indices[var->data.location - VARYING_SLOT_VAR0];
2927 switch (instr->intrinsic) {
2928 case nir_intrinsic_interp_deref_at_centroid:
2929 location = INTERP_CENTROID;
2930 break;
2931 case nir_intrinsic_interp_deref_at_sample:
2932 case nir_intrinsic_interp_deref_at_offset:
2933 location = INTERP_CENTER;
2934 src0 = get_src(ctx, instr->src[1]);
2935 break;
2936 default:
2937 break;
2938 }
2939
2940 if (instr->intrinsic == nir_intrinsic_interp_deref_at_offset) {
2941 src_c0 = ac_to_float(&ctx->ac, LLVMBuildExtractElement(ctx->ac.builder, src0, ctx->ac.i32_0, ""));
2942 src_c1 = ac_to_float(&ctx->ac, LLVMBuildExtractElement(ctx->ac.builder, src0, ctx->ac.i32_1, ""));
2943 } else if (instr->intrinsic == nir_intrinsic_interp_deref_at_sample) {
2944 LLVMValueRef sample_position;
2945 LLVMValueRef halfval = LLVMConstReal(ctx->ac.f32, 0.5f);
2946
2947 /* fetch sample ID */
2948 sample_position = ctx->abi->load_sample_position(ctx->abi, src0);
2949
2950 src_c0 = LLVMBuildExtractElement(ctx->ac.builder, sample_position, ctx->ac.i32_0, "");
2951 src_c0 = LLVMBuildFSub(ctx->ac.builder, src_c0, halfval, "");
2952 src_c1 = LLVMBuildExtractElement(ctx->ac.builder, sample_position, ctx->ac.i32_1, "");
2953 src_c1 = LLVMBuildFSub(ctx->ac.builder, src_c1, halfval, "");
2954 }
2955 interp_param = ctx->abi->lookup_interp_param(ctx->abi, var->data.interpolation, location);
2956
2957 if (location == INTERP_CENTER) {
2958 LLVMValueRef ij_out[2];
2959 LLVMValueRef ddxy_out = emit_ddxy_interp(ctx, interp_param);
2960
2961 /*
2962 * take the I then J parameters, and the DDX/Y for it, and
2963 * calculate the IJ inputs for the interpolator.
2964 * temp1 = ddx * offset/sample.x + I;
2965 * interp_param.I = ddy * offset/sample.y + temp1;
2966 * temp1 = ddx * offset/sample.x + J;
2967 * interp_param.J = ddy * offset/sample.y + temp1;
2968 */
2969 for (unsigned i = 0; i < 2; i++) {
2970 LLVMValueRef ix_ll = LLVMConstInt(ctx->ac.i32, i, false);
2971 LLVMValueRef iy_ll = LLVMConstInt(ctx->ac.i32, i + 2, false);
2972 LLVMValueRef ddx_el = LLVMBuildExtractElement(ctx->ac.builder,
2973 ddxy_out, ix_ll, "");
2974 LLVMValueRef ddy_el = LLVMBuildExtractElement(ctx->ac.builder,
2975 ddxy_out, iy_ll, "");
2976 LLVMValueRef interp_el = LLVMBuildExtractElement(ctx->ac.builder,
2977 interp_param, ix_ll, "");
2978 LLVMValueRef temp1, temp2;
2979
2980 interp_el = LLVMBuildBitCast(ctx->ac.builder, interp_el,
2981 ctx->ac.f32, "");
2982
2983 temp1 = ac_build_fmad(&ctx->ac, ddx_el, src_c0, interp_el);
2984 temp2 = ac_build_fmad(&ctx->ac, ddy_el, src_c1, temp1);
2985
2986 ij_out[i] = LLVMBuildBitCast(ctx->ac.builder,
2987 temp2, ctx->ac.i32, "");
2988 }
2989 interp_param = ac_build_gather_values(&ctx->ac, ij_out, 2);
2990
2991 }
2992
2993 LLVMValueRef attrib_idx = ctx->ac.i32_0;
2994 while(deref_instr->deref_type != nir_deref_type_var) {
2995 if (deref_instr->deref_type == nir_deref_type_array) {
2996 unsigned array_size = glsl_count_attribute_slots(deref_instr->type, false);
2997
2998 LLVMValueRef offset;
2999 nir_const_value *const_value = nir_src_as_const_value(deref_instr->arr.index);
3000 if (const_value) {
3001 offset = LLVMConstInt(ctx->ac.i32, array_size * const_value->u32[0], false);
3002 } else {
3003 LLVMValueRef indirect = get_src(ctx, deref_instr->arr.index);
3004
3005 offset = LLVMBuildMul(ctx->ac.builder, indirect,
3006 LLVMConstInt(ctx->ac.i32, array_size, false), "");
3007 }
3008
3009 attrib_idx = LLVMBuildAdd(ctx->ac.builder, attrib_idx, offset, "");
3010 deref_instr = nir_src_as_deref(deref_instr->parent);
3011 } else if (deref_instr->deref_type == nir_deref_type_struct) {
3012 LLVMValueRef offset;
3013 unsigned sidx = deref_instr->strct.index;
3014 deref_instr = nir_src_as_deref(deref_instr->parent);
3015 offset = LLVMConstInt(ctx->ac.i32, glsl_get_struct_location_offset(deref_instr->type, sidx), false);
3016 attrib_idx = LLVMBuildAdd(ctx->ac.builder, attrib_idx, offset, "");
3017 } else {
3018 unreachable("Unsupported deref type");
3019 }
3020
3021 }
3022
3023 unsigned attrib_size = glsl_count_attribute_slots(var->type, false);
3024 for (chan = 0; chan < 4; chan++) {
3025 LLVMValueRef gather = LLVMGetUndef(LLVMVectorType(ctx->ac.f32, attrib_size));
3026 LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, chan, false);
3027
3028 for (unsigned idx = 0; idx < attrib_size; ++idx) {
3029 LLVMValueRef v, attr_number;
3030
3031 attr_number = LLVMConstInt(ctx->ac.i32, input_base + idx, false);
3032 if (interp_param) {
3033 interp_param = LLVMBuildBitCast(ctx->ac.builder,
3034 interp_param, ctx->ac.v2f32, "");
3035 LLVMValueRef i = LLVMBuildExtractElement(
3036 ctx->ac.builder, interp_param, ctx->ac.i32_0, "");
3037 LLVMValueRef j = LLVMBuildExtractElement(
3038 ctx->ac.builder, interp_param, ctx->ac.i32_1, "");
3039
3040 v = ac_build_fs_interp(&ctx->ac, llvm_chan, attr_number,
3041 ctx->abi->prim_mask, i, j);
3042 } else {
3043 v = ac_build_fs_interp_mov(&ctx->ac, LLVMConstInt(ctx->ac.i32, 2, false),
3044 llvm_chan, attr_number, ctx->abi->prim_mask);
3045 }
3046
3047 gather = LLVMBuildInsertElement(ctx->ac.builder, gather, v,
3048 LLVMConstInt(ctx->ac.i32, idx, false), "");
3049 }
3050
3051 result[chan] = LLVMBuildExtractElement(ctx->ac.builder, gather, attrib_idx, "");
3052
3053 }
3054 return ac_build_varying_gather_values(&ctx->ac, result, instr->num_components,
3055 var->data.location_frac);
3056 }
3057
3058 static void visit_intrinsic(struct ac_nir_context *ctx,
3059 nir_intrinsic_instr *instr)
3060 {
3061 LLVMValueRef result = NULL;
3062
3063 switch (instr->intrinsic) {
3064 case nir_intrinsic_ballot:
3065 result = ac_build_ballot(&ctx->ac, get_src(ctx, instr->src[0]));
3066 break;
3067 case nir_intrinsic_read_invocation:
3068 result = ac_build_readlane(&ctx->ac, get_src(ctx, instr->src[0]),
3069 get_src(ctx, instr->src[1]));
3070 break;
3071 case nir_intrinsic_read_first_invocation:
3072 result = ac_build_readlane(&ctx->ac, get_src(ctx, instr->src[0]), NULL);
3073 break;
3074 case nir_intrinsic_load_subgroup_invocation:
3075 result = ac_get_thread_id(&ctx->ac);
3076 break;
3077 case nir_intrinsic_load_work_group_id: {
3078 LLVMValueRef values[3];
3079
3080 for (int i = 0; i < 3; i++) {
3081 values[i] = ctx->abi->workgroup_ids[i] ?
3082 ctx->abi->workgroup_ids[i] : ctx->ac.i32_0;
3083 }
3084
3085 result = ac_build_gather_values(&ctx->ac, values, 3);
3086 break;
3087 }
3088 case nir_intrinsic_load_base_vertex:
3089 case nir_intrinsic_load_first_vertex:
3090 result = ctx->abi->load_base_vertex(ctx->abi);
3091 break;
3092 case nir_intrinsic_load_local_group_size:
3093 result = ctx->abi->load_local_group_size(ctx->abi);
3094 break;
3095 case nir_intrinsic_load_vertex_id:
3096 result = LLVMBuildAdd(ctx->ac.builder, ctx->abi->vertex_id,
3097 ctx->abi->base_vertex, "");
3098 break;
3099 case nir_intrinsic_load_vertex_id_zero_base: {
3100 result = ctx->abi->vertex_id;
3101 break;
3102 }
3103 case nir_intrinsic_load_local_invocation_id: {
3104 result = ctx->abi->local_invocation_ids;
3105 break;
3106 }
3107 case nir_intrinsic_load_base_instance:
3108 result = ctx->abi->start_instance;
3109 break;
3110 case nir_intrinsic_load_draw_id:
3111 result = ctx->abi->draw_id;
3112 break;
3113 case nir_intrinsic_load_view_index:
3114 result = ctx->abi->view_index;
3115 break;
3116 case nir_intrinsic_load_invocation_id:
3117 if (ctx->stage == MESA_SHADER_TESS_CTRL)
3118 result = ac_unpack_param(&ctx->ac, ctx->abi->tcs_rel_ids, 8, 5);
3119 else
3120 result = ctx->abi->gs_invocation_id;
3121 break;
3122 case nir_intrinsic_load_primitive_id:
3123 if (ctx->stage == MESA_SHADER_GEOMETRY) {
3124 result = ctx->abi->gs_prim_id;
3125 } else if (ctx->stage == MESA_SHADER_TESS_CTRL) {
3126 result = ctx->abi->tcs_patch_id;
3127 } else if (ctx->stage == MESA_SHADER_TESS_EVAL) {
3128 result = ctx->abi->tes_patch_id;
3129 } else
3130 fprintf(stderr, "Unknown primitive id intrinsic: %d", ctx->stage);
3131 break;
3132 case nir_intrinsic_load_sample_id:
3133 result = ac_unpack_param(&ctx->ac, ctx->abi->ancillary, 8, 4);
3134 break;
3135 case nir_intrinsic_load_sample_pos:
3136 result = load_sample_pos(ctx);
3137 break;
3138 case nir_intrinsic_load_sample_mask_in:
3139 result = ctx->abi->load_sample_mask_in(ctx->abi);
3140 break;
3141 case nir_intrinsic_load_frag_coord: {
3142 LLVMValueRef values[4] = {
3143 ctx->abi->frag_pos[0],
3144 ctx->abi->frag_pos[1],
3145 ctx->abi->frag_pos[2],
3146 ac_build_fdiv(&ctx->ac, ctx->ac.f32_1, ctx->abi->frag_pos[3])
3147 };
3148 result = ac_build_gather_values(&ctx->ac, values, 4);
3149 break;
3150 }
3151 case nir_intrinsic_load_front_face:
3152 result = ctx->abi->front_face;
3153 break;
3154 case nir_intrinsic_load_helper_invocation:
3155 result = visit_load_helper_invocation(ctx);
3156 break;
3157 case nir_intrinsic_load_instance_id:
3158 result = ctx->abi->instance_id;
3159 break;
3160 case nir_intrinsic_load_num_work_groups:
3161 result = ctx->abi->num_work_groups;
3162 break;
3163 case nir_intrinsic_load_local_invocation_index:
3164 result = visit_load_local_invocation_index(ctx);
3165 break;
3166 case nir_intrinsic_load_subgroup_id:
3167 result = visit_load_subgroup_id(ctx);
3168 break;
3169 case nir_intrinsic_load_num_subgroups:
3170 result = visit_load_num_subgroups(ctx);
3171 break;
3172 case nir_intrinsic_first_invocation:
3173 result = visit_first_invocation(ctx);
3174 break;
3175 case nir_intrinsic_load_push_constant:
3176 result = visit_load_push_constant(ctx, instr);
3177 break;
3178 case nir_intrinsic_vulkan_resource_index: {
3179 LLVMValueRef index = get_src(ctx, instr->src[0]);
3180 unsigned desc_set = nir_intrinsic_desc_set(instr);
3181 unsigned binding = nir_intrinsic_binding(instr);
3182
3183 result = ctx->abi->load_resource(ctx->abi, index, desc_set,
3184 binding);
3185 break;
3186 }
3187 case nir_intrinsic_vulkan_resource_reindex:
3188 result = visit_vulkan_resource_reindex(ctx, instr);
3189 break;
3190 case nir_intrinsic_store_ssbo:
3191 visit_store_ssbo(ctx, instr);
3192 break;
3193 case nir_intrinsic_load_ssbo:
3194 result = visit_load_buffer(ctx, instr);
3195 break;
3196 case nir_intrinsic_ssbo_atomic_add:
3197 case nir_intrinsic_ssbo_atomic_imin:
3198 case nir_intrinsic_ssbo_atomic_umin:
3199 case nir_intrinsic_ssbo_atomic_imax:
3200 case nir_intrinsic_ssbo_atomic_umax:
3201 case nir_intrinsic_ssbo_atomic_and:
3202 case nir_intrinsic_ssbo_atomic_or:
3203 case nir_intrinsic_ssbo_atomic_xor:
3204 case nir_intrinsic_ssbo_atomic_exchange:
3205 case nir_intrinsic_ssbo_atomic_comp_swap:
3206 result = visit_atomic_ssbo(ctx, instr);
3207 break;
3208 case nir_intrinsic_load_ubo:
3209 result = visit_load_ubo_buffer(ctx, instr);
3210 break;
3211 case nir_intrinsic_get_buffer_size:
3212 result = visit_get_buffer_size(ctx, instr);
3213 break;
3214 case nir_intrinsic_load_deref:
3215 result = visit_load_var(ctx, instr);
3216 break;
3217 case nir_intrinsic_store_deref:
3218 visit_store_var(ctx, instr);
3219 break;
3220 case nir_intrinsic_load_shared:
3221 result = visit_load_shared(ctx, instr);
3222 break;
3223 case nir_intrinsic_store_shared:
3224 visit_store_shared(ctx, instr);
3225 break;
3226 case nir_intrinsic_image_deref_samples:
3227 result = visit_image_samples(ctx, instr);
3228 break;
3229 case nir_intrinsic_image_deref_load:
3230 result = visit_image_load(ctx, instr);
3231 break;
3232 case nir_intrinsic_image_deref_store:
3233 visit_image_store(ctx, instr);
3234 break;
3235 case nir_intrinsic_image_deref_atomic_add:
3236 case nir_intrinsic_image_deref_atomic_min:
3237 case nir_intrinsic_image_deref_atomic_max:
3238 case nir_intrinsic_image_deref_atomic_and:
3239 case nir_intrinsic_image_deref_atomic_or:
3240 case nir_intrinsic_image_deref_atomic_xor:
3241 case nir_intrinsic_image_deref_atomic_exchange:
3242 case nir_intrinsic_image_deref_atomic_comp_swap:
3243 result = visit_image_atomic(ctx, instr);
3244 break;
3245 case nir_intrinsic_image_deref_size:
3246 result = visit_image_size(ctx, instr);
3247 break;
3248 case nir_intrinsic_shader_clock:
3249 result = ac_build_shader_clock(&ctx->ac);
3250 break;
3251 case nir_intrinsic_discard:
3252 case nir_intrinsic_discard_if:
3253 emit_discard(ctx, instr);
3254 break;
3255 case nir_intrinsic_memory_barrier:
3256 case nir_intrinsic_group_memory_barrier:
3257 case nir_intrinsic_memory_barrier_atomic_counter:
3258 case nir_intrinsic_memory_barrier_buffer:
3259 case nir_intrinsic_memory_barrier_image:
3260 case nir_intrinsic_memory_barrier_shared:
3261 emit_membar(&ctx->ac, instr);
3262 break;
3263 case nir_intrinsic_barrier:
3264 ac_emit_barrier(&ctx->ac, ctx->stage);
3265 break;
3266 case nir_intrinsic_shared_atomic_add:
3267 case nir_intrinsic_shared_atomic_imin:
3268 case nir_intrinsic_shared_atomic_umin:
3269 case nir_intrinsic_shared_atomic_imax:
3270 case nir_intrinsic_shared_atomic_umax:
3271 case nir_intrinsic_shared_atomic_and:
3272 case nir_intrinsic_shared_atomic_or:
3273 case nir_intrinsic_shared_atomic_xor:
3274 case nir_intrinsic_shared_atomic_exchange:
3275 case nir_intrinsic_shared_atomic_comp_swap: {
3276 LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[0]);
3277 result = visit_var_atomic(ctx, instr, ptr, 1);
3278 break;
3279 }
3280 case nir_intrinsic_deref_atomic_add:
3281 case nir_intrinsic_deref_atomic_imin:
3282 case nir_intrinsic_deref_atomic_umin:
3283 case nir_intrinsic_deref_atomic_imax:
3284 case nir_intrinsic_deref_atomic_umax:
3285 case nir_intrinsic_deref_atomic_and:
3286 case nir_intrinsic_deref_atomic_or:
3287 case nir_intrinsic_deref_atomic_xor:
3288 case nir_intrinsic_deref_atomic_exchange:
3289 case nir_intrinsic_deref_atomic_comp_swap: {
3290 LLVMValueRef ptr = get_src(ctx, instr->src[0]);
3291 result = visit_var_atomic(ctx, instr, ptr, 1);
3292 break;
3293 }
3294 case nir_intrinsic_interp_deref_at_centroid:
3295 case nir_intrinsic_interp_deref_at_sample:
3296 case nir_intrinsic_interp_deref_at_offset:
3297 result = visit_interp(ctx, instr);
3298 break;
3299 case nir_intrinsic_emit_vertex:
3300 ctx->abi->emit_vertex(ctx->abi, nir_intrinsic_stream_id(instr), ctx->abi->outputs);
3301 break;
3302 case nir_intrinsic_end_primitive:
3303 ctx->abi->emit_primitive(ctx->abi, nir_intrinsic_stream_id(instr));
3304 break;
3305 case nir_intrinsic_load_tess_coord:
3306 result = ctx->abi->load_tess_coord(ctx->abi);
3307 break;
3308 case nir_intrinsic_load_tess_level_outer:
3309 result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_OUTER);
3310 break;
3311 case nir_intrinsic_load_tess_level_inner:
3312 result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_INNER);
3313 break;
3314 case nir_intrinsic_load_patch_vertices_in:
3315 result = ctx->abi->load_patch_vertices_in(ctx->abi);
3316 break;
3317 case nir_intrinsic_vote_all: {
3318 LLVMValueRef tmp = ac_build_vote_all(&ctx->ac, get_src(ctx, instr->src[0]));
3319 result = LLVMBuildSExt(ctx->ac.builder, tmp, ctx->ac.i32, "");
3320 break;
3321 }
3322 case nir_intrinsic_vote_any: {
3323 LLVMValueRef tmp = ac_build_vote_any(&ctx->ac, get_src(ctx, instr->src[0]));
3324 result = LLVMBuildSExt(ctx->ac.builder, tmp, ctx->ac.i32, "");
3325 break;
3326 }
3327 case nir_intrinsic_shuffle:
3328 result = ac_build_shuffle(&ctx->ac, get_src(ctx, instr->src[0]),
3329 get_src(ctx, instr->src[1]));
3330 break;
3331 case nir_intrinsic_reduce:
3332 result = ac_build_reduce(&ctx->ac,
3333 get_src(ctx, instr->src[0]),
3334 instr->const_index[0],
3335 instr->const_index[1]);
3336 break;
3337 case nir_intrinsic_inclusive_scan:
3338 result = ac_build_inclusive_scan(&ctx->ac,
3339 get_src(ctx, instr->src[0]),
3340 instr->const_index[0]);
3341 break;
3342 case nir_intrinsic_exclusive_scan:
3343 result = ac_build_exclusive_scan(&ctx->ac,
3344 get_src(ctx, instr->src[0]),
3345 instr->const_index[0]);
3346 break;
3347 case nir_intrinsic_quad_broadcast: {
3348 unsigned lane = nir_src_as_const_value(instr->src[1])->u32[0];
3349 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]),
3350 lane, lane, lane, lane);
3351 break;
3352 }
3353 case nir_intrinsic_quad_swap_horizontal:
3354 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 1, 0, 3 ,2);
3355 break;
3356 case nir_intrinsic_quad_swap_vertical:
3357 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 2, 3, 0 ,1);
3358 break;
3359 case nir_intrinsic_quad_swap_diagonal:
3360 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 3, 2, 1 ,0);
3361 break;
3362 default:
3363 fprintf(stderr, "Unknown intrinsic: ");
3364 nir_print_instr(&instr->instr, stderr);
3365 fprintf(stderr, "\n");
3366 break;
3367 }
3368 if (result) {
3369 ctx->ssa_defs[instr->dest.ssa.index] = result;
3370 }
3371 }
3372
3373 static LLVMValueRef get_bindless_index_from_uniform(struct ac_nir_context *ctx,
3374 unsigned base_index,
3375 unsigned constant_index,
3376 LLVMValueRef dynamic_index)
3377 {
3378 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, base_index * 4, 0);
3379 LLVMValueRef index = LLVMBuildAdd(ctx->ac.builder, dynamic_index,
3380 LLVMConstInt(ctx->ac.i32, constant_index, 0), "");
3381
3382 /* Bindless uniforms are 64bit so multiple index by 8 */
3383 index = LLVMBuildMul(ctx->ac.builder, index, LLVMConstInt(ctx->ac.i32, 8, 0), "");
3384 offset = LLVMBuildAdd(ctx->ac.builder, offset, index, "");
3385
3386 LLVMValueRef ubo_index = ctx->abi->load_ubo(ctx->abi, ctx->ac.i32_0);
3387
3388 LLVMValueRef ret = ac_build_buffer_load(&ctx->ac, ubo_index, 1, NULL, offset,
3389 NULL, 0, false, false, true, true);
3390
3391 return LLVMBuildBitCast(ctx->ac.builder, ret, ctx->ac.i32, "");
3392 }
3393
3394 static LLVMValueRef get_sampler_desc(struct ac_nir_context *ctx,
3395 nir_deref_instr *deref_instr,
3396 enum ac_descriptor_type desc_type,
3397 const nir_tex_instr *tex_instr,
3398 bool image, bool write)
3399 {
3400 LLVMValueRef index = NULL;
3401 unsigned constant_index = 0;
3402 unsigned descriptor_set;
3403 unsigned base_index;
3404 bool bindless = false;
3405
3406 if (!deref_instr) {
3407 assert(tex_instr && !image);
3408 descriptor_set = 0;
3409 base_index = tex_instr->sampler_index;
3410 } else {
3411 while(deref_instr->deref_type != nir_deref_type_var) {
3412 if (deref_instr->deref_type == nir_deref_type_array) {
3413 unsigned array_size = glsl_get_aoa_size(deref_instr->type);
3414 if (!array_size)
3415 array_size = 1;
3416
3417 nir_const_value *const_value = nir_src_as_const_value(deref_instr->arr.index);
3418 if (const_value) {
3419 constant_index += array_size * const_value->u32[0];
3420 } else {
3421 LLVMValueRef indirect = get_src(ctx, deref_instr->arr.index);
3422
3423 indirect = LLVMBuildMul(ctx->ac.builder, indirect,
3424 LLVMConstInt(ctx->ac.i32, array_size, false), "");
3425
3426 if (!index)
3427 index = indirect;
3428 else
3429 index = LLVMBuildAdd(ctx->ac.builder, index, indirect, "");
3430 }
3431
3432 deref_instr = nir_src_as_deref(deref_instr->parent);
3433 } else if (deref_instr->deref_type == nir_deref_type_struct) {
3434 unsigned sidx = deref_instr->strct.index;
3435 deref_instr = nir_src_as_deref(deref_instr->parent);
3436 constant_index += glsl_get_struct_location_offset(deref_instr->type, sidx);
3437 } else {
3438 unreachable("Unsupported deref type");
3439 }
3440 }
3441 descriptor_set = deref_instr->var->data.descriptor_set;
3442
3443 if (deref_instr->var->data.bindless) {
3444 /* For now just assert on unhandled variable types */
3445 assert(deref_instr->var->data.mode == nir_var_uniform);
3446
3447 base_index = deref_instr->var->data.driver_location;
3448 bindless = true;
3449
3450 index = index ? index : ctx->ac.i32_0;
3451 index = get_bindless_index_from_uniform(ctx, base_index,
3452 constant_index, index);
3453 } else
3454 base_index = deref_instr->var->data.binding;
3455 }
3456
3457 return ctx->abi->load_sampler_desc(ctx->abi,
3458 descriptor_set,
3459 base_index,
3460 constant_index, index,
3461 desc_type, image, write, bindless);
3462 }
3463
3464 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
3465 *
3466 * SI-CI:
3467 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
3468 * filtering manually. The driver sets img7 to a mask clearing
3469 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
3470 * s_and_b32 samp0, samp0, img7
3471 *
3472 * VI:
3473 * The ANISO_OVERRIDE sampler field enables this fix in TA.
3474 */
3475 static LLVMValueRef sici_fix_sampler_aniso(struct ac_nir_context *ctx,
3476 LLVMValueRef res, LLVMValueRef samp)
3477 {
3478 LLVMBuilderRef builder = ctx->ac.builder;
3479 LLVMValueRef img7, samp0;
3480
3481 if (ctx->ac.chip_class >= VI)
3482 return samp;
3483
3484 img7 = LLVMBuildExtractElement(builder, res,
3485 LLVMConstInt(ctx->ac.i32, 7, 0), "");
3486 samp0 = LLVMBuildExtractElement(builder, samp,
3487 LLVMConstInt(ctx->ac.i32, 0, 0), "");
3488 samp0 = LLVMBuildAnd(builder, samp0, img7, "");
3489 return LLVMBuildInsertElement(builder, samp, samp0,
3490 LLVMConstInt(ctx->ac.i32, 0, 0), "");
3491 }
3492
3493 static void tex_fetch_ptrs(struct ac_nir_context *ctx,
3494 nir_tex_instr *instr,
3495 LLVMValueRef *res_ptr, LLVMValueRef *samp_ptr,
3496 LLVMValueRef *fmask_ptr)
3497 {
3498 nir_deref_instr *texture_deref_instr = NULL;
3499 nir_deref_instr *sampler_deref_instr = NULL;
3500
3501 for (unsigned i = 0; i < instr->num_srcs; i++) {
3502 switch (instr->src[i].src_type) {
3503 case nir_tex_src_texture_deref:
3504 texture_deref_instr = nir_src_as_deref(instr->src[i].src);
3505 break;
3506 case nir_tex_src_sampler_deref:
3507 sampler_deref_instr = nir_src_as_deref(instr->src[i].src);
3508 break;
3509 default:
3510 break;
3511 }
3512 }
3513
3514 if (!sampler_deref_instr)
3515 sampler_deref_instr = texture_deref_instr;
3516
3517 if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF)
3518 *res_ptr = get_sampler_desc(ctx, texture_deref_instr, AC_DESC_BUFFER, instr, false, false);
3519 else
3520 *res_ptr = get_sampler_desc(ctx, texture_deref_instr, AC_DESC_IMAGE, instr, false, false);
3521 if (samp_ptr) {
3522 *samp_ptr = get_sampler_desc(ctx, sampler_deref_instr, AC_DESC_SAMPLER, instr, false, false);
3523 if (instr->sampler_dim < GLSL_SAMPLER_DIM_RECT)
3524 *samp_ptr = sici_fix_sampler_aniso(ctx, *res_ptr, *samp_ptr);
3525 }
3526 if (fmask_ptr && (instr->op == nir_texop_txf_ms ||
3527 instr->op == nir_texop_samples_identical))
3528 *fmask_ptr = get_sampler_desc(ctx, texture_deref_instr, AC_DESC_FMASK, instr, false, false);
3529 }
3530
3531 static LLVMValueRef apply_round_slice(struct ac_llvm_context *ctx,
3532 LLVMValueRef coord)
3533 {
3534 coord = ac_to_float(ctx, coord);
3535 coord = ac_build_round(ctx, coord);
3536 coord = ac_to_integer(ctx, coord);
3537 return coord;
3538 }
3539
3540 static void visit_tex(struct ac_nir_context *ctx, nir_tex_instr *instr)
3541 {
3542 LLVMValueRef result = NULL;
3543 struct ac_image_args args = { 0 };
3544 LLVMValueRef fmask_ptr = NULL, sample_index = NULL;
3545 LLVMValueRef ddx = NULL, ddy = NULL;
3546 unsigned offset_src = 0;
3547
3548 tex_fetch_ptrs(ctx, instr, &args.resource, &args.sampler, &fmask_ptr);
3549
3550 for (unsigned i = 0; i < instr->num_srcs; i++) {
3551 switch (instr->src[i].src_type) {
3552 case nir_tex_src_coord: {
3553 LLVMValueRef coord = get_src(ctx, instr->src[i].src);
3554 for (unsigned chan = 0; chan < instr->coord_components; ++chan)
3555 args.coords[chan] = ac_llvm_extract_elem(&ctx->ac, coord, chan);
3556 break;
3557 }
3558 case nir_tex_src_projector:
3559 break;
3560 case nir_tex_src_comparator:
3561 if (instr->is_shadow)
3562 args.compare = get_src(ctx, instr->src[i].src);
3563 break;
3564 case nir_tex_src_offset:
3565 args.offset = get_src(ctx, instr->src[i].src);
3566 offset_src = i;
3567 break;
3568 case nir_tex_src_bias:
3569 if (instr->op == nir_texop_txb)
3570 args.bias = get_src(ctx, instr->src[i].src);
3571 break;
3572 case nir_tex_src_lod: {
3573 nir_const_value *val = nir_src_as_const_value(instr->src[i].src);
3574
3575 if (val && val->i32[0] == 0)
3576 args.level_zero = true;
3577 else
3578 args.lod = get_src(ctx, instr->src[i].src);
3579 break;
3580 }
3581 case nir_tex_src_ms_index:
3582 sample_index = get_src(ctx, instr->src[i].src);
3583 break;
3584 case nir_tex_src_ms_mcs:
3585 break;
3586 case nir_tex_src_ddx:
3587 ddx = get_src(ctx, instr->src[i].src);
3588 break;
3589 case nir_tex_src_ddy:
3590 ddy = get_src(ctx, instr->src[i].src);
3591 break;
3592 case nir_tex_src_texture_offset:
3593 case nir_tex_src_sampler_offset:
3594 case nir_tex_src_plane:
3595 default:
3596 break;
3597 }
3598 }
3599
3600 if (instr->op == nir_texop_txs && instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
3601 result = get_buffer_size(ctx, args.resource, true);
3602 goto write_result;
3603 }
3604
3605 if (instr->op == nir_texop_texture_samples) {
3606 LLVMValueRef res, samples, is_msaa;
3607 res = LLVMBuildBitCast(ctx->ac.builder, args.resource, ctx->ac.v8i32, "");
3608 samples = LLVMBuildExtractElement(ctx->ac.builder, res,
3609 LLVMConstInt(ctx->ac.i32, 3, false), "");
3610 is_msaa = LLVMBuildLShr(ctx->ac.builder, samples,
3611 LLVMConstInt(ctx->ac.i32, 28, false), "");
3612 is_msaa = LLVMBuildAnd(ctx->ac.builder, is_msaa,
3613 LLVMConstInt(ctx->ac.i32, 0xe, false), "");
3614 is_msaa = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, is_msaa,
3615 LLVMConstInt(ctx->ac.i32, 0xe, false), "");
3616
3617 samples = LLVMBuildLShr(ctx->ac.builder, samples,
3618 LLVMConstInt(ctx->ac.i32, 16, false), "");
3619 samples = LLVMBuildAnd(ctx->ac.builder, samples,
3620 LLVMConstInt(ctx->ac.i32, 0xf, false), "");
3621 samples = LLVMBuildShl(ctx->ac.builder, ctx->ac.i32_1,
3622 samples, "");
3623 samples = LLVMBuildSelect(ctx->ac.builder, is_msaa, samples,
3624 ctx->ac.i32_1, "");
3625 result = samples;
3626 goto write_result;
3627 }
3628
3629 if (args.offset && instr->op != nir_texop_txf) {
3630 LLVMValueRef offset[3], pack;
3631 for (unsigned chan = 0; chan < 3; ++chan)
3632 offset[chan] = ctx->ac.i32_0;
3633
3634 unsigned num_components = ac_get_llvm_num_components(args.offset);
3635 for (unsigned chan = 0; chan < num_components; chan++) {
3636 offset[chan] = ac_llvm_extract_elem(&ctx->ac, args.offset, chan);
3637 offset[chan] = LLVMBuildAnd(ctx->ac.builder, offset[chan],
3638 LLVMConstInt(ctx->ac.i32, 0x3f, false), "");
3639 if (chan)
3640 offset[chan] = LLVMBuildShl(ctx->ac.builder, offset[chan],
3641 LLVMConstInt(ctx->ac.i32, chan * 8, false), "");
3642 }
3643 pack = LLVMBuildOr(ctx->ac.builder, offset[0], offset[1], "");
3644 pack = LLVMBuildOr(ctx->ac.builder, pack, offset[2], "");
3645 args.offset = pack;
3646 }
3647
3648 /* TC-compatible HTILE on radeonsi promotes Z16 and Z24 to Z32_FLOAT,
3649 * so the depth comparison value isn't clamped for Z16 and
3650 * Z24 anymore. Do it manually here.
3651 *
3652 * It's unnecessary if the original texture format was
3653 * Z32_FLOAT, but we don't know that here.
3654 */
3655 if (args.compare && ctx->ac.chip_class >= VI && ctx->abi->clamp_shadow_reference)
3656 args.compare = ac_build_clamp(&ctx->ac, ac_to_float(&ctx->ac, args.compare));
3657
3658 /* pack derivatives */
3659 if (ddx || ddy) {
3660 int num_src_deriv_channels, num_dest_deriv_channels;
3661 switch (instr->sampler_dim) {
3662 case GLSL_SAMPLER_DIM_3D:
3663 case GLSL_SAMPLER_DIM_CUBE:
3664 num_src_deriv_channels = 3;
3665 num_dest_deriv_channels = 3;
3666 break;
3667 case GLSL_SAMPLER_DIM_2D:
3668 default:
3669 num_src_deriv_channels = 2;
3670 num_dest_deriv_channels = 2;
3671 break;
3672 case GLSL_SAMPLER_DIM_1D:
3673 num_src_deriv_channels = 1;
3674 if (ctx->ac.chip_class >= GFX9) {
3675 num_dest_deriv_channels = 2;
3676 } else {
3677 num_dest_deriv_channels = 1;
3678 }
3679 break;
3680 }
3681
3682 for (unsigned i = 0; i < num_src_deriv_channels; i++) {
3683 args.derivs[i] = ac_to_float(&ctx->ac,
3684 ac_llvm_extract_elem(&ctx->ac, ddx, i));
3685 args.derivs[num_dest_deriv_channels + i] = ac_to_float(&ctx->ac,
3686 ac_llvm_extract_elem(&ctx->ac, ddy, i));
3687 }
3688 for (unsigned i = num_src_deriv_channels; i < num_dest_deriv_channels; i++) {
3689 args.derivs[i] = ctx->ac.f32_0;
3690 args.derivs[num_dest_deriv_channels + i] = ctx->ac.f32_0;
3691 }
3692 }
3693
3694 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE && args.coords[0]) {
3695 for (unsigned chan = 0; chan < instr->coord_components; chan++)
3696 args.coords[chan] = ac_to_float(&ctx->ac, args.coords[chan]);
3697 if (instr->coord_components == 3)
3698 args.coords[3] = LLVMGetUndef(ctx->ac.f32);
3699 ac_prepare_cube_coords(&ctx->ac,
3700 instr->op == nir_texop_txd, instr->is_array,
3701 instr->op == nir_texop_lod, args.coords, args.derivs);
3702 }
3703
3704 /* Texture coordinates fixups */
3705 if (instr->coord_components > 1 &&
3706 instr->sampler_dim == GLSL_SAMPLER_DIM_1D &&
3707 instr->is_array &&
3708 instr->op != nir_texop_txf) {
3709 args.coords[1] = apply_round_slice(&ctx->ac, args.coords[1]);
3710 }
3711
3712 if (instr->coord_components > 2 &&
3713 (instr->sampler_dim == GLSL_SAMPLER_DIM_2D ||
3714 instr->sampler_dim == GLSL_SAMPLER_DIM_MS ||
3715 instr->sampler_dim == GLSL_SAMPLER_DIM_SUBPASS ||
3716 instr->sampler_dim == GLSL_SAMPLER_DIM_SUBPASS_MS) &&
3717 instr->is_array &&
3718 instr->op != nir_texop_txf && instr->op != nir_texop_txf_ms) {
3719 args.coords[2] = apply_round_slice(&ctx->ac, args.coords[2]);
3720 }
3721
3722 if (ctx->ac.chip_class >= GFX9 &&
3723 instr->sampler_dim == GLSL_SAMPLER_DIM_1D &&
3724 instr->op != nir_texop_lod) {
3725 LLVMValueRef filler;
3726 if (instr->op == nir_texop_txf)
3727 filler = ctx->ac.i32_0;
3728 else
3729 filler = LLVMConstReal(ctx->ac.f32, 0.5);
3730
3731 if (instr->is_array)
3732 args.coords[2] = args.coords[1];
3733 args.coords[1] = filler;
3734 }
3735
3736 /* Pack sample index */
3737 if (instr->op == nir_texop_txf_ms && sample_index)
3738 args.coords[instr->coord_components] = sample_index;
3739
3740 if (instr->op == nir_texop_samples_identical) {
3741 struct ac_image_args txf_args = { 0 };
3742 memcpy(txf_args.coords, args.coords, sizeof(txf_args.coords));
3743
3744 txf_args.dmask = 0xf;
3745 txf_args.resource = fmask_ptr;
3746 txf_args.dim = instr->is_array ? ac_image_2darray : ac_image_2d;
3747 result = build_tex_intrinsic(ctx, instr, &txf_args);
3748
3749 result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
3750 result = emit_int_cmp(&ctx->ac, LLVMIntEQ, result, ctx->ac.i32_0);
3751 goto write_result;
3752 }
3753
3754 if (instr->sampler_dim == GLSL_SAMPLER_DIM_MS &&
3755 instr->op != nir_texop_txs) {
3756 unsigned sample_chan = instr->is_array ? 3 : 2;
3757 args.coords[sample_chan] = adjust_sample_index_using_fmask(
3758 &ctx->ac, args.coords[0], args.coords[1],
3759 instr->is_array ? args.coords[2] : NULL,
3760 args.coords[sample_chan], fmask_ptr);
3761 }
3762
3763 if (args.offset && instr->op == nir_texop_txf) {
3764 nir_const_value *const_offset =
3765 nir_src_as_const_value(instr->src[offset_src].src);
3766 int num_offsets = instr->src[offset_src].src.ssa->num_components;
3767 assert(const_offset);
3768 num_offsets = MIN2(num_offsets, instr->coord_components);
3769 for (unsigned i = 0; i < num_offsets; ++i) {
3770 args.coords[i] = LLVMBuildAdd(
3771 ctx->ac.builder, args.coords[i],
3772 LLVMConstInt(ctx->ac.i32, const_offset->i32[i], false), "");
3773 }
3774 args.offset = NULL;
3775 }
3776
3777 /* TODO TG4 support */
3778 args.dmask = 0xf;
3779 if (instr->op == nir_texop_tg4) {
3780 if (instr->is_shadow)
3781 args.dmask = 1;
3782 else
3783 args.dmask = 1 << instr->component;
3784 }
3785
3786 if (instr->sampler_dim != GLSL_SAMPLER_DIM_BUF)
3787 args.dim = get_ac_sampler_dim(&ctx->ac, instr->sampler_dim, instr->is_array);
3788 result = build_tex_intrinsic(ctx, instr, &args);
3789
3790 if (instr->op == nir_texop_query_levels)
3791 result = LLVMBuildExtractElement(ctx->ac.builder, result, LLVMConstInt(ctx->ac.i32, 3, false), "");
3792 else if (instr->is_shadow && instr->is_new_style_shadow &&
3793 instr->op != nir_texop_txs && instr->op != nir_texop_lod &&
3794 instr->op != nir_texop_tg4)
3795 result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
3796 else if (instr->op == nir_texop_txs &&
3797 instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE &&
3798 instr->is_array) {
3799 LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
3800 LLVMValueRef six = LLVMConstInt(ctx->ac.i32, 6, false);
3801 LLVMValueRef z = LLVMBuildExtractElement(ctx->ac.builder, result, two, "");
3802 z = LLVMBuildSDiv(ctx->ac.builder, z, six, "");
3803 result = LLVMBuildInsertElement(ctx->ac.builder, result, z, two, "");
3804 } else if (ctx->ac.chip_class >= GFX9 &&
3805 instr->op == nir_texop_txs &&
3806 instr->sampler_dim == GLSL_SAMPLER_DIM_1D &&
3807 instr->is_array) {
3808 LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
3809 LLVMValueRef layers = LLVMBuildExtractElement(ctx->ac.builder, result, two, "");
3810 result = LLVMBuildInsertElement(ctx->ac.builder, result, layers,
3811 ctx->ac.i32_1, "");
3812 } else if (instr->dest.ssa.num_components != 4)
3813 result = ac_trim_vector(&ctx->ac, result, instr->dest.ssa.num_components);
3814
3815 write_result:
3816 if (result) {
3817 assert(instr->dest.is_ssa);
3818 result = ac_to_integer(&ctx->ac, result);
3819 ctx->ssa_defs[instr->dest.ssa.index] = result;
3820 }
3821 }
3822
3823
3824 static void visit_phi(struct ac_nir_context *ctx, nir_phi_instr *instr)
3825 {
3826 LLVMTypeRef type = get_def_type(ctx, &instr->dest.ssa);
3827 LLVMValueRef result = LLVMBuildPhi(ctx->ac.builder, type, "");
3828
3829 ctx->ssa_defs[instr->dest.ssa.index] = result;
3830 _mesa_hash_table_insert(ctx->phis, instr, result);
3831 }
3832
3833 static void visit_post_phi(struct ac_nir_context *ctx,
3834 nir_phi_instr *instr,
3835 LLVMValueRef llvm_phi)
3836 {
3837 nir_foreach_phi_src(src, instr) {
3838 LLVMBasicBlockRef block = get_block(ctx, src->pred);
3839 LLVMValueRef llvm_src = get_src(ctx, src->src);
3840
3841 LLVMAddIncoming(llvm_phi, &llvm_src, &block, 1);
3842 }
3843 }
3844
3845 static void phi_post_pass(struct ac_nir_context *ctx)
3846 {
3847 hash_table_foreach(ctx->phis, entry) {
3848 visit_post_phi(ctx, (nir_phi_instr*)entry->key,
3849 (LLVMValueRef)entry->data);
3850 }
3851 }
3852
3853
3854 static void visit_ssa_undef(struct ac_nir_context *ctx,
3855 const nir_ssa_undef_instr *instr)
3856 {
3857 unsigned num_components = instr->def.num_components;
3858 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
3859 LLVMValueRef undef;
3860
3861 if (num_components == 1)
3862 undef = LLVMGetUndef(type);
3863 else {
3864 undef = LLVMGetUndef(LLVMVectorType(type, num_components));
3865 }
3866 ctx->ssa_defs[instr->def.index] = undef;
3867 }
3868
3869 static void visit_jump(struct ac_llvm_context *ctx,
3870 const nir_jump_instr *instr)
3871 {
3872 switch (instr->type) {
3873 case nir_jump_break:
3874 ac_build_break(ctx);
3875 break;
3876 case nir_jump_continue:
3877 ac_build_continue(ctx);
3878 break;
3879 default:
3880 fprintf(stderr, "Unknown NIR jump instr: ");
3881 nir_print_instr(&instr->instr, stderr);
3882 fprintf(stderr, "\n");
3883 abort();
3884 }
3885 }
3886
3887 static LLVMTypeRef
3888 glsl_base_to_llvm_type(struct ac_llvm_context *ac,
3889 enum glsl_base_type type)
3890 {
3891 switch (type) {
3892 case GLSL_TYPE_INT:
3893 case GLSL_TYPE_UINT:
3894 case GLSL_TYPE_BOOL:
3895 case GLSL_TYPE_SUBROUTINE:
3896 return ac->i32;
3897 case GLSL_TYPE_INT16:
3898 case GLSL_TYPE_UINT16:
3899 return ac->i16;
3900 case GLSL_TYPE_FLOAT:
3901 return ac->f32;
3902 case GLSL_TYPE_FLOAT16:
3903 return ac->f16;
3904 case GLSL_TYPE_INT64:
3905 case GLSL_TYPE_UINT64:
3906 return ac->i64;
3907 case GLSL_TYPE_DOUBLE:
3908 return ac->f64;
3909 default:
3910 unreachable("unknown GLSL type");
3911 }
3912 }
3913
3914 static LLVMTypeRef
3915 glsl_to_llvm_type(struct ac_llvm_context *ac,
3916 const struct glsl_type *type)
3917 {
3918 if (glsl_type_is_scalar(type)) {
3919 return glsl_base_to_llvm_type(ac, glsl_get_base_type(type));
3920 }
3921
3922 if (glsl_type_is_vector(type)) {
3923 return LLVMVectorType(
3924 glsl_base_to_llvm_type(ac, glsl_get_base_type(type)),
3925 glsl_get_vector_elements(type));
3926 }
3927
3928 if (glsl_type_is_matrix(type)) {
3929 return LLVMArrayType(
3930 glsl_to_llvm_type(ac, glsl_get_column_type(type)),
3931 glsl_get_matrix_columns(type));
3932 }
3933
3934 if (glsl_type_is_array(type)) {
3935 return LLVMArrayType(
3936 glsl_to_llvm_type(ac, glsl_get_array_element(type)),
3937 glsl_get_length(type));
3938 }
3939
3940 assert(glsl_type_is_struct_or_ifc(type));
3941
3942 LLVMTypeRef member_types[glsl_get_length(type)];
3943
3944 for (unsigned i = 0; i < glsl_get_length(type); i++) {
3945 member_types[i] =
3946 glsl_to_llvm_type(ac,
3947 glsl_get_struct_field(type, i));
3948 }
3949
3950 return LLVMStructTypeInContext(ac->context, member_types,
3951 glsl_get_length(type), false);
3952 }
3953
3954 static void visit_deref(struct ac_nir_context *ctx,
3955 nir_deref_instr *instr)
3956 {
3957 if (instr->mode != nir_var_mem_shared &&
3958 instr->mode != nir_var_mem_global)
3959 return;
3960
3961 LLVMValueRef result = NULL;
3962 switch(instr->deref_type) {
3963 case nir_deref_type_var: {
3964 struct hash_entry *entry = _mesa_hash_table_search(ctx->vars, instr->var);
3965 result = entry->data;
3966 break;
3967 }
3968 case nir_deref_type_struct:
3969 if (instr->mode == nir_var_mem_global) {
3970 nir_deref_instr *parent = nir_deref_instr_parent(instr);
3971 uint64_t offset = glsl_get_struct_field_offset(parent->type,
3972 instr->strct.index);
3973 result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent),
3974 LLVMConstInt(ctx->ac.i32, offset, 0));
3975 } else {
3976 result = ac_build_gep0(&ctx->ac, get_src(ctx, instr->parent),
3977 LLVMConstInt(ctx->ac.i32, instr->strct.index, 0));
3978 }
3979 break;
3980 case nir_deref_type_array:
3981 if (instr->mode == nir_var_mem_global) {
3982 nir_deref_instr *parent = nir_deref_instr_parent(instr);
3983 unsigned stride = glsl_get_explicit_stride(parent->type);
3984
3985 if ((glsl_type_is_matrix(parent->type) &&
3986 glsl_matrix_type_is_row_major(parent->type)) ||
3987 (glsl_type_is_vector(parent->type) && stride == 0))
3988 stride = type_scalar_size_bytes(parent->type);
3989
3990 assert(stride > 0);
3991 LLVMValueRef index = get_src(ctx, instr->arr.index);
3992 if (LLVMTypeOf(index) != ctx->ac.i64)
3993 index = LLVMBuildZExt(ctx->ac.builder, index, ctx->ac.i64, "");
3994
3995 LLVMValueRef offset = LLVMBuildMul(ctx->ac.builder, index, LLVMConstInt(ctx->ac.i64, stride, 0), "");
3996
3997 result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent), offset);
3998 } else {
3999 result = ac_build_gep0(&ctx->ac, get_src(ctx, instr->parent),
4000 get_src(ctx, instr->arr.index));
4001 }
4002 break;
4003 case nir_deref_type_ptr_as_array:
4004 if (instr->mode == nir_var_mem_global) {
4005 unsigned stride = nir_deref_instr_ptr_as_array_stride(instr);
4006
4007 LLVMValueRef index = get_src(ctx, instr->arr.index);
4008 if (LLVMTypeOf(index) != ctx->ac.i64)
4009 index = LLVMBuildZExt(ctx->ac.builder, index, ctx->ac.i64, "");
4010
4011 LLVMValueRef offset = LLVMBuildMul(ctx->ac.builder, index, LLVMConstInt(ctx->ac.i64, stride, 0), "");
4012
4013 result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent), offset);
4014 } else {
4015 result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent),
4016 get_src(ctx, instr->arr.index));
4017 }
4018 break;
4019 case nir_deref_type_cast: {
4020 result = get_src(ctx, instr->parent);
4021
4022 /* We can't use the structs from LLVM because the shader
4023 * specifies its own offsets. */
4024 LLVMTypeRef pointee_type = ctx->ac.i8;
4025 if (instr->mode == nir_var_mem_shared)
4026 pointee_type = glsl_to_llvm_type(&ctx->ac, instr->type);
4027
4028 unsigned address_space;
4029
4030 switch(instr->mode) {
4031 case nir_var_mem_shared:
4032 address_space = AC_ADDR_SPACE_LDS;
4033 break;
4034 case nir_var_mem_global:
4035 address_space = AC_ADDR_SPACE_GLOBAL;
4036 break;
4037 default:
4038 unreachable("Unhandled address space");
4039 }
4040
4041 LLVMTypeRef type = LLVMPointerType(pointee_type, address_space);
4042
4043 if (LLVMTypeOf(result) != type) {
4044 if (LLVMGetTypeKind(LLVMTypeOf(result)) == LLVMVectorTypeKind) {
4045 result = LLVMBuildBitCast(ctx->ac.builder, result,
4046 type, "");
4047 } else {
4048 result = LLVMBuildIntToPtr(ctx->ac.builder, result,
4049 type, "");
4050 }
4051 }
4052 break;
4053 }
4054 default:
4055 unreachable("Unhandled deref_instr deref type");
4056 }
4057
4058 ctx->ssa_defs[instr->dest.ssa.index] = result;
4059 }
4060
4061 static void visit_cf_list(struct ac_nir_context *ctx,
4062 struct exec_list *list);
4063
4064 static void visit_block(struct ac_nir_context *ctx, nir_block *block)
4065 {
4066 LLVMBasicBlockRef llvm_block = LLVMGetInsertBlock(ctx->ac.builder);
4067 nir_foreach_instr(instr, block)
4068 {
4069 switch (instr->type) {
4070 case nir_instr_type_alu:
4071 visit_alu(ctx, nir_instr_as_alu(instr));
4072 break;
4073 case nir_instr_type_load_const:
4074 visit_load_const(ctx, nir_instr_as_load_const(instr));
4075 break;
4076 case nir_instr_type_intrinsic:
4077 visit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
4078 break;
4079 case nir_instr_type_tex:
4080 visit_tex(ctx, nir_instr_as_tex(instr));
4081 break;
4082 case nir_instr_type_phi:
4083 visit_phi(ctx, nir_instr_as_phi(instr));
4084 break;
4085 case nir_instr_type_ssa_undef:
4086 visit_ssa_undef(ctx, nir_instr_as_ssa_undef(instr));
4087 break;
4088 case nir_instr_type_jump:
4089 visit_jump(&ctx->ac, nir_instr_as_jump(instr));
4090 break;
4091 case nir_instr_type_deref:
4092 visit_deref(ctx, nir_instr_as_deref(instr));
4093 break;
4094 default:
4095 fprintf(stderr, "Unknown NIR instr type: ");
4096 nir_print_instr(instr, stderr);
4097 fprintf(stderr, "\n");
4098 abort();
4099 }
4100 }
4101
4102 _mesa_hash_table_insert(ctx->defs, block, llvm_block);
4103 }
4104
4105 static void visit_if(struct ac_nir_context *ctx, nir_if *if_stmt)
4106 {
4107 LLVMValueRef value = get_src(ctx, if_stmt->condition);
4108
4109 nir_block *then_block =
4110 (nir_block *) exec_list_get_head(&if_stmt->then_list);
4111
4112 ac_build_uif(&ctx->ac, value, then_block->index);
4113
4114 visit_cf_list(ctx, &if_stmt->then_list);
4115
4116 if (!exec_list_is_empty(&if_stmt->else_list)) {
4117 nir_block *else_block =
4118 (nir_block *) exec_list_get_head(&if_stmt->else_list);
4119
4120 ac_build_else(&ctx->ac, else_block->index);
4121 visit_cf_list(ctx, &if_stmt->else_list);
4122 }
4123
4124 ac_build_endif(&ctx->ac, then_block->index);
4125 }
4126
4127 static void visit_loop(struct ac_nir_context *ctx, nir_loop *loop)
4128 {
4129 nir_block *first_loop_block =
4130 (nir_block *) exec_list_get_head(&loop->body);
4131
4132 ac_build_bgnloop(&ctx->ac, first_loop_block->index);
4133
4134 visit_cf_list(ctx, &loop->body);
4135
4136 ac_build_endloop(&ctx->ac, first_loop_block->index);
4137 }
4138
4139 static void visit_cf_list(struct ac_nir_context *ctx,
4140 struct exec_list *list)
4141 {
4142 foreach_list_typed(nir_cf_node, node, node, list)
4143 {
4144 switch (node->type) {
4145 case nir_cf_node_block:
4146 visit_block(ctx, nir_cf_node_as_block(node));
4147 break;
4148
4149 case nir_cf_node_if:
4150 visit_if(ctx, nir_cf_node_as_if(node));
4151 break;
4152
4153 case nir_cf_node_loop:
4154 visit_loop(ctx, nir_cf_node_as_loop(node));
4155 break;
4156
4157 default:
4158 assert(0);
4159 }
4160 }
4161 }
4162
4163 void
4164 ac_handle_shader_output_decl(struct ac_llvm_context *ctx,
4165 struct ac_shader_abi *abi,
4166 struct nir_shader *nir,
4167 struct nir_variable *variable,
4168 gl_shader_stage stage)
4169 {
4170 unsigned output_loc = variable->data.driver_location / 4;
4171 unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
4172
4173 /* tess ctrl has it's own load/store paths for outputs */
4174 if (stage == MESA_SHADER_TESS_CTRL)
4175 return;
4176
4177 if (stage == MESA_SHADER_VERTEX ||
4178 stage == MESA_SHADER_TESS_EVAL ||
4179 stage == MESA_SHADER_GEOMETRY) {
4180 int idx = variable->data.location + variable->data.index;
4181 if (idx == VARYING_SLOT_CLIP_DIST0) {
4182 int length = nir->info.clip_distance_array_size +
4183 nir->info.cull_distance_array_size;
4184
4185 if (length > 4)
4186 attrib_count = 2;
4187 else
4188 attrib_count = 1;
4189 }
4190 }
4191
4192 bool is_16bit = glsl_type_is_16bit(glsl_without_array(variable->type));
4193 LLVMTypeRef type = is_16bit ? ctx->f16 : ctx->f32;
4194 for (unsigned i = 0; i < attrib_count; ++i) {
4195 for (unsigned chan = 0; chan < 4; chan++) {
4196 abi->outputs[ac_llvm_reg_index_soa(output_loc + i, chan)] =
4197 ac_build_alloca_undef(ctx, type, "");
4198 }
4199 }
4200 }
4201
4202 static void
4203 setup_locals(struct ac_nir_context *ctx,
4204 struct nir_function *func)
4205 {
4206 int i, j;
4207 ctx->num_locals = 0;
4208 nir_foreach_variable(variable, &func->impl->locals) {
4209 unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
4210 variable->data.driver_location = ctx->num_locals * 4;
4211 variable->data.location_frac = 0;
4212 ctx->num_locals += attrib_count;
4213 }
4214 ctx->locals = malloc(4 * ctx->num_locals * sizeof(LLVMValueRef));
4215 if (!ctx->locals)
4216 return;
4217
4218 for (i = 0; i < ctx->num_locals; i++) {
4219 for (j = 0; j < 4; j++) {
4220 ctx->locals[i * 4 + j] =
4221 ac_build_alloca_undef(&ctx->ac, ctx->ac.f32, "temp");
4222 }
4223 }
4224 }
4225
4226 static void
4227 setup_shared(struct ac_nir_context *ctx,
4228 struct nir_shader *nir)
4229 {
4230 nir_foreach_variable(variable, &nir->shared) {
4231 LLVMValueRef shared =
4232 LLVMAddGlobalInAddressSpace(
4233 ctx->ac.module, glsl_to_llvm_type(&ctx->ac, variable->type),
4234 variable->name ? variable->name : "",
4235 AC_ADDR_SPACE_LDS);
4236 _mesa_hash_table_insert(ctx->vars, variable, shared);
4237 }
4238 }
4239
4240 void ac_nir_translate(struct ac_llvm_context *ac, struct ac_shader_abi *abi,
4241 struct nir_shader *nir)
4242 {
4243 struct ac_nir_context ctx = {};
4244 struct nir_function *func;
4245
4246 ctx.ac = *ac;
4247 ctx.abi = abi;
4248
4249 ctx.stage = nir->info.stage;
4250
4251 ctx.main_function = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx.ac.builder));
4252
4253 nir_foreach_variable(variable, &nir->outputs)
4254 ac_handle_shader_output_decl(&ctx.ac, ctx.abi, nir, variable,
4255 ctx.stage);
4256
4257 ctx.defs = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
4258 _mesa_key_pointer_equal);
4259 ctx.phis = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
4260 _mesa_key_pointer_equal);
4261 ctx.vars = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
4262 _mesa_key_pointer_equal);
4263
4264 func = (struct nir_function *)exec_list_get_head(&nir->functions);
4265
4266 nir_index_ssa_defs(func->impl);
4267 ctx.ssa_defs = calloc(func->impl->ssa_alloc, sizeof(LLVMValueRef));
4268
4269 setup_locals(&ctx, func);
4270
4271 if (gl_shader_stage_is_compute(nir->info.stage))
4272 setup_shared(&ctx, nir);
4273
4274 visit_cf_list(&ctx, &func->impl->body);
4275 phi_post_pass(&ctx);
4276
4277 if (!gl_shader_stage_is_compute(nir->info.stage))
4278 ctx.abi->emit_outputs(ctx.abi, AC_LLVM_MAX_OUTPUTS,
4279 ctx.abi->outputs);
4280
4281 free(ctx.locals);
4282 free(ctx.ssa_defs);
4283 ralloc_free(ctx.defs);
4284 ralloc_free(ctx.phis);
4285 ralloc_free(ctx.vars);
4286 }
4287
4288 void
4289 ac_lower_indirect_derefs(struct nir_shader *nir, enum chip_class chip_class)
4290 {
4291 /* While it would be nice not to have this flag, we are constrained
4292 * by the reality that LLVM 5.0 doesn't have working VGPR indexing
4293 * on GFX9.
4294 */
4295 bool llvm_has_working_vgpr_indexing = chip_class <= VI;
4296
4297 /* TODO: Indirect indexing of GS inputs is unimplemented.
4298 *
4299 * TCS and TES load inputs directly from LDS or offchip memory, so
4300 * indirect indexing is trivial.
4301 */
4302 nir_variable_mode indirect_mask = 0;
4303 if (nir->info.stage == MESA_SHADER_GEOMETRY ||
4304 (nir->info.stage != MESA_SHADER_TESS_CTRL &&
4305 nir->info.stage != MESA_SHADER_TESS_EVAL &&
4306 !llvm_has_working_vgpr_indexing)) {
4307 indirect_mask |= nir_var_shader_in;
4308 }
4309 if (!llvm_has_working_vgpr_indexing &&
4310 nir->info.stage != MESA_SHADER_TESS_CTRL)
4311 indirect_mask |= nir_var_shader_out;
4312
4313 /* TODO: We shouldn't need to do this, however LLVM isn't currently
4314 * smart enough to handle indirects without causing excess spilling
4315 * causing the gpu to hang.
4316 *
4317 * See the following thread for more details of the problem:
4318 * https://lists.freedesktop.org/archives/mesa-dev/2017-July/162106.html
4319 */
4320 indirect_mask |= nir_var_function_temp;
4321
4322 nir_lower_indirect_derefs(nir, indirect_mask);
4323 }
4324
4325 static unsigned
4326 get_inst_tessfactor_writemask(nir_intrinsic_instr *intrin)
4327 {
4328 if (intrin->intrinsic != nir_intrinsic_store_deref)
4329 return 0;
4330
4331 nir_variable *var =
4332 nir_deref_instr_get_variable(nir_src_as_deref(intrin->src[0]));
4333
4334 if (var->data.mode != nir_var_shader_out)
4335 return 0;
4336
4337 unsigned writemask = 0;
4338 const int location = var->data.location;
4339 unsigned first_component = var->data.location_frac;
4340 unsigned num_comps = intrin->dest.ssa.num_components;
4341
4342 if (location == VARYING_SLOT_TESS_LEVEL_INNER)
4343 writemask = ((1 << (num_comps + 1)) - 1) << first_component;
4344 else if (location == VARYING_SLOT_TESS_LEVEL_OUTER)
4345 writemask = (((1 << (num_comps + 1)) - 1) << first_component) << 4;
4346
4347 return writemask;
4348 }
4349
4350 static void
4351 scan_tess_ctrl(nir_cf_node *cf_node, unsigned *upper_block_tf_writemask,
4352 unsigned *cond_block_tf_writemask,
4353 bool *tessfactors_are_def_in_all_invocs, bool is_nested_cf)
4354 {
4355 switch (cf_node->type) {
4356 case nir_cf_node_block: {
4357 nir_block *block = nir_cf_node_as_block(cf_node);
4358 nir_foreach_instr(instr, block) {
4359 if (instr->type != nir_instr_type_intrinsic)
4360 continue;
4361
4362 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
4363 if (intrin->intrinsic == nir_intrinsic_barrier) {
4364
4365 /* If we find a barrier in nested control flow put this in the
4366 * too hard basket. In GLSL this is not possible but it is in
4367 * SPIR-V.
4368 */
4369 if (is_nested_cf) {
4370 *tessfactors_are_def_in_all_invocs = false;
4371 return;
4372 }
4373
4374 /* The following case must be prevented:
4375 * gl_TessLevelInner = ...;
4376 * barrier();
4377 * if (gl_InvocationID == 1)
4378 * gl_TessLevelInner = ...;
4379 *
4380 * If you consider disjoint code segments separated by barriers, each
4381 * such segment that writes tess factor channels should write the same
4382 * channels in all codepaths within that segment.
4383 */
4384 if (upper_block_tf_writemask || cond_block_tf_writemask) {
4385 /* Accumulate the result: */
4386 *tessfactors_are_def_in_all_invocs &=
4387 !(*cond_block_tf_writemask & ~(*upper_block_tf_writemask));
4388
4389 /* Analyze the next code segment from scratch. */
4390 *upper_block_tf_writemask = 0;
4391 *cond_block_tf_writemask = 0;
4392 }
4393 } else
4394 *upper_block_tf_writemask |= get_inst_tessfactor_writemask(intrin);
4395 }
4396
4397 break;
4398 }
4399 case nir_cf_node_if: {
4400 unsigned then_tessfactor_writemask = 0;
4401 unsigned else_tessfactor_writemask = 0;
4402
4403 nir_if *if_stmt = nir_cf_node_as_if(cf_node);
4404 foreach_list_typed(nir_cf_node, nested_node, node, &if_stmt->then_list) {
4405 scan_tess_ctrl(nested_node, &then_tessfactor_writemask,
4406 cond_block_tf_writemask,
4407 tessfactors_are_def_in_all_invocs, true);
4408 }
4409
4410 foreach_list_typed(nir_cf_node, nested_node, node, &if_stmt->else_list) {
4411 scan_tess_ctrl(nested_node, &else_tessfactor_writemask,
4412 cond_block_tf_writemask,
4413 tessfactors_are_def_in_all_invocs, true);
4414 }
4415
4416 if (then_tessfactor_writemask || else_tessfactor_writemask) {
4417 /* If both statements write the same tess factor channels,
4418 * we can say that the upper block writes them too.
4419 */
4420 *upper_block_tf_writemask |= then_tessfactor_writemask &
4421 else_tessfactor_writemask;
4422 *cond_block_tf_writemask |= then_tessfactor_writemask |
4423 else_tessfactor_writemask;
4424 }
4425
4426 break;
4427 }
4428 case nir_cf_node_loop: {
4429 nir_loop *loop = nir_cf_node_as_loop(cf_node);
4430 foreach_list_typed(nir_cf_node, nested_node, node, &loop->body) {
4431 scan_tess_ctrl(nested_node, cond_block_tf_writemask,
4432 cond_block_tf_writemask,
4433 tessfactors_are_def_in_all_invocs, true);
4434 }
4435
4436 break;
4437 }
4438 default:
4439 unreachable("unknown cf node type");
4440 }
4441 }
4442
4443 bool
4444 ac_are_tessfactors_def_in_all_invocs(const struct nir_shader *nir)
4445 {
4446 assert(nir->info.stage == MESA_SHADER_TESS_CTRL);
4447
4448 /* The pass works as follows:
4449 * If all codepaths write tess factors, we can say that all
4450 * invocations define tess factors.
4451 *
4452 * Each tess factor channel is tracked separately.
4453 */
4454 unsigned main_block_tf_writemask = 0; /* if main block writes tess factors */
4455 unsigned cond_block_tf_writemask = 0; /* if cond block writes tess factors */
4456
4457 /* Initial value = true. Here the pass will accumulate results from
4458 * multiple segments surrounded by barriers. If tess factors aren't
4459 * written at all, it's a shader bug and we don't care if this will be
4460 * true.
4461 */
4462 bool tessfactors_are_def_in_all_invocs = true;
4463
4464 nir_foreach_function(function, nir) {
4465 if (function->impl) {
4466 foreach_list_typed(nir_cf_node, node, node, &function->impl->body) {
4467 scan_tess_ctrl(node, &main_block_tf_writemask,
4468 &cond_block_tf_writemask,
4469 &tessfactors_are_def_in_all_invocs,
4470 false);
4471 }
4472 }
4473 }
4474
4475 /* Accumulate the result for the last code segment separated by a
4476 * barrier.
4477 */
4478 if (main_block_tf_writemask || cond_block_tf_writemask) {
4479 tessfactors_are_def_in_all_invocs &=
4480 !(cond_block_tf_writemask & ~main_block_tf_writemask);
4481 }
4482
4483 return tessfactors_are_def_in_all_invocs;
4484 }