projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
ac/nir: add 8-bit types to glsl_base_to_llvm_type
[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 == 8) {
1436 unsigned load_dwords = instr->dest.ssa.num_components > 1 ? 2 : 1;
1437 LLVMTypeRef vec_type = LLVMVectorType(LLVMInt8TypeInContext(ctx->ac.context), 4 * load_dwords);
1438 ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
1439 LLVMValueRef res = LLVMBuildLoad(ctx->ac.builder, ptr, "");
1440
1441 LLVMValueRef params[3];
1442 if (load_dwords > 1) {
1443 LLVMValueRef res_vec = LLVMBuildBitCast(ctx->ac.builder, res, LLVMVectorType(ctx->ac.i32, 2), "");
1444 params[0] = LLVMBuildExtractElement(ctx->ac.builder, res_vec, LLVMConstInt(ctx->ac.i32, 1, false), "");
1445 params[1] = LLVMBuildExtractElement(ctx->ac.builder, res_vec, LLVMConstInt(ctx->ac.i32, 0, false), "");
1446 } else {
1447 res = LLVMBuildBitCast(ctx->ac.builder, res, ctx->ac.i32, "");
1448 params[0] = ctx->ac.i32_0;
1449 params[1] = res;
1450 }
1451 params[2] = addr;
1452 res = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.alignbyte", ctx->ac.i32, params, 3, 0);
1453
1454 res = LLVMBuildTrunc(ctx->ac.builder, res, LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.num_components * 8), "");
1455 if (instr->dest.ssa.num_components > 1)
1456 res = LLVMBuildBitCast(ctx->ac.builder, res, LLVMVectorType(LLVMInt8TypeInContext(ctx->ac.context), instr->dest.ssa.num_components), "");
1457 return res;
1458 } else if (instr->dest.ssa.bit_size == 16) {
1459 unsigned load_dwords = instr->dest.ssa.num_components / 2 + 1;
1460 LLVMTypeRef vec_type = LLVMVectorType(LLVMInt16TypeInContext(ctx->ac.context), 2 * load_dwords);
1461 ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
1462 LLVMValueRef res = LLVMBuildLoad(ctx->ac.builder, ptr, "");
1463 res = LLVMBuildBitCast(ctx->ac.builder, res, vec_type, "");
1464 LLVMValueRef cond = LLVMBuildLShr(ctx->ac.builder, addr, ctx->ac.i32_1, "");
1465 cond = LLVMBuildTrunc(ctx->ac.builder, cond, ctx->ac.i1, "");
1466 LLVMValueRef mask[] = { LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
1467 LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
1468 LLVMConstInt(ctx->ac.i32, 4, false)};
1469 LLVMValueRef swizzle_aligned = LLVMConstVector(&mask[0], instr->dest.ssa.num_components);
1470 LLVMValueRef swizzle_unaligned = LLVMConstVector(&mask[1], instr->dest.ssa.num_components);
1471 LLVMValueRef shuffle_aligned = LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_aligned, "");
1472 LLVMValueRef shuffle_unaligned = LLVMBuildShuffleVector(ctx->ac.builder, res, res, swizzle_unaligned, "");
1473 res = LLVMBuildSelect(ctx->ac.builder, cond, shuffle_unaligned, shuffle_aligned, "");
1474 return LLVMBuildBitCast(ctx->ac.builder, res, get_def_type(ctx, &instr->dest.ssa), "");
1475 }
1476
1477 ptr = ac_cast_ptr(&ctx->ac, ptr, get_def_type(ctx, &instr->dest.ssa));
1478
1479 return LLVMBuildLoad(ctx->ac.builder, ptr, "");
1480 }
1481
1482 static LLVMValueRef visit_get_buffer_size(struct ac_nir_context *ctx,
1483 const nir_intrinsic_instr *instr)
1484 {
1485 LLVMValueRef index = get_src(ctx, instr->src[0]);
1486
1487 return get_buffer_size(ctx, ctx->abi->load_ssbo(ctx->abi, index, false), false);
1488 }
1489
1490 static uint32_t widen_mask(uint32_t mask, unsigned multiplier)
1491 {
1492 uint32_t new_mask = 0;
1493 for(unsigned i = 0; i < 32 && (1u << i) <= mask; ++i)
1494 if (mask & (1u << i))
1495 new_mask |= ((1u << multiplier) - 1u) << (i * multiplier);
1496 return new_mask;
1497 }
1498
1499 static LLVMValueRef extract_vector_range(struct ac_llvm_context *ctx, LLVMValueRef src,
1500 unsigned start, unsigned count)
1501 {
1502 LLVMValueRef mask[] = {
1503 ctx->i32_0, ctx->i32_1,
1504 LLVMConstInt(ctx->i32, 2, false), LLVMConstInt(ctx->i32, 3, false) };
1505
1506 unsigned src_elements = ac_get_llvm_num_components(src);
1507
1508 if (count == src_elements) {
1509 assert(start == 0);
1510 return src;
1511 } else if (count == 1) {
1512 assert(start < src_elements);
1513 return LLVMBuildExtractElement(ctx->builder, src, mask[start], "");
1514 } else {
1515 assert(start + count <= src_elements);
1516 assert(count <= 4);
1517 LLVMValueRef swizzle = LLVMConstVector(&mask[start], count);
1518 return LLVMBuildShuffleVector(ctx->builder, src, src, swizzle, "");
1519 }
1520 }
1521
1522 static unsigned get_cache_policy(struct ac_nir_context *ctx,
1523 enum gl_access_qualifier access,
1524 bool may_store_unaligned,
1525 bool writeonly_memory)
1526 {
1527 unsigned cache_policy = 0;
1528
1529 /* SI has a TC L1 bug causing corruption of 8bit/16bit stores. All
1530 * store opcodes not aligned to a dword are affected. The only way to
1531 * get unaligned stores is through shader images.
1532 */
1533 if (((may_store_unaligned && ctx->ac.chip_class == SI) ||
1534 /* If this is write-only, don't keep data in L1 to prevent
1535 * evicting L1 cache lines that may be needed by other
1536 * instructions.
1537 */
1538 writeonly_memory ||
1539 access & (ACCESS_COHERENT | ACCESS_VOLATILE))) {
1540 cache_policy |= ac_glc;
1541 }
1542
1543 return cache_policy;
1544 }
1545
1546 static void visit_store_ssbo(struct ac_nir_context *ctx,
1547 nir_intrinsic_instr *instr)
1548 {
1549 LLVMValueRef src_data = get_src(ctx, instr->src[0]);
1550 int elem_size_bytes = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src_data)) / 8;
1551 unsigned writemask = nir_intrinsic_write_mask(instr);
1552 enum gl_access_qualifier access = nir_intrinsic_access(instr);
1553 bool writeonly_memory = access & ACCESS_NON_READABLE;
1554 unsigned cache_policy = get_cache_policy(ctx, access, false, writeonly_memory);
1555
1556 LLVMValueRef rsrc = ctx->abi->load_ssbo(ctx->abi,
1557 get_src(ctx, instr->src[1]), true);
1558 LLVMValueRef base_data = src_data;
1559 base_data = ac_trim_vector(&ctx->ac, base_data, instr->num_components);
1560 LLVMValueRef base_offset = get_src(ctx, instr->src[2]);
1561
1562 while (writemask) {
1563 int start, count;
1564 LLVMValueRef data, offset;
1565 LLVMTypeRef data_type;
1566
1567 u_bit_scan_consecutive_range(&writemask, &start, &count);
1568
1569 /* Due to an LLVM limitation, split 3-element writes
1570 * into a 2-element and a 1-element write. */
1571 if (count == 3) {
1572 writemask |= 1 << (start + 2);
1573 count = 2;
1574 }
1575 int num_bytes = count * elem_size_bytes; /* count in bytes */
1576
1577 /* we can only store 4 DWords at the same time.
1578 * can only happen for 64 Bit vectors. */
1579 if (num_bytes > 16) {
1580 writemask |= ((1u << (count - 2)) - 1u) << (start + 2);
1581 count = 2;
1582 num_bytes = 16;
1583 }
1584
1585 /* check alignment of 16 Bit stores */
1586 if (elem_size_bytes == 2 && num_bytes > 2 && (start % 2) == 1) {
1587 writemask |= ((1u << (count - 1)) - 1u) << (start + 1);
1588 count = 1;
1589 num_bytes = 2;
1590 }
1591 data = extract_vector_range(&ctx->ac, base_data, start, count);
1592
1593 offset = LLVMBuildAdd(ctx->ac.builder, base_offset,
1594 LLVMConstInt(ctx->ac.i32, start * elem_size_bytes, false), "");
1595
1596 if (num_bytes == 1) {
1597 ac_build_tbuffer_store_byte(&ctx->ac, rsrc, data,
1598 offset, ctx->ac.i32_0,
1599 cache_policy & ac_glc,
1600 writeonly_memory);
1601 } else if (num_bytes == 2) {
1602 ac_build_tbuffer_store_short(&ctx->ac, rsrc, data,
1603 offset, ctx->ac.i32_0,
1604 cache_policy & ac_glc,
1605 writeonly_memory);
1606 } else {
1607 int num_channels = num_bytes / 4;
1608
1609 switch (num_bytes) {
1610 case 16: /* v4f32 */
1611 data_type = ctx->ac.v4f32;
1612 break;
1613 case 8: /* v2f32 */
1614 data_type = ctx->ac.v2f32;
1615 break;
1616 case 4: /* f32 */
1617 data_type = ctx->ac.f32;
1618 break;
1619 default:
1620 unreachable("Malformed vector store.");
1621 }
1622 data = LLVMBuildBitCast(ctx->ac.builder, data, data_type, "");
1623
1624 ac_build_buffer_store_dword(&ctx->ac, rsrc, data,
1625 num_channels, offset,
1626 ctx->ac.i32_0, 0,
1627 cache_policy & ac_glc,
1628 false, writeonly_memory,
1629 false);
1630 }
1631 }
1632 }
1633
1634 static LLVMValueRef visit_atomic_ssbo(struct ac_nir_context *ctx,
1635 const nir_intrinsic_instr *instr)
1636 {
1637 const char *atomic_name;
1638 char intrinsic_name[64];
1639 LLVMValueRef params[7];
1640 int arg_count = 0;
1641 int length;
1642
1643 switch (instr->intrinsic) {
1644 case nir_intrinsic_ssbo_atomic_add:
1645 atomic_name = "add";
1646 break;
1647 case nir_intrinsic_ssbo_atomic_imin:
1648 atomic_name = "smin";
1649 break;
1650 case nir_intrinsic_ssbo_atomic_umin:
1651 atomic_name = "umin";
1652 break;
1653 case nir_intrinsic_ssbo_atomic_imax:
1654 atomic_name = "smax";
1655 break;
1656 case nir_intrinsic_ssbo_atomic_umax:
1657 atomic_name = "umax";
1658 break;
1659 case nir_intrinsic_ssbo_atomic_and:
1660 atomic_name = "and";
1661 break;
1662 case nir_intrinsic_ssbo_atomic_or:
1663 atomic_name = "or";
1664 break;
1665 case nir_intrinsic_ssbo_atomic_xor:
1666 atomic_name = "xor";
1667 break;
1668 case nir_intrinsic_ssbo_atomic_exchange:
1669 atomic_name = "swap";
1670 break;
1671 case nir_intrinsic_ssbo_atomic_comp_swap:
1672 atomic_name = "cmpswap";
1673 break;
1674 default:
1675 abort();
1676 }
1677
1678 if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap) {
1679 params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[3]), 0);
1680 }
1681 params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
1682 params[arg_count++] = ctx->abi->load_ssbo(ctx->abi,
1683 get_src(ctx, instr->src[0]),
1684 true);
1685
1686 if (HAVE_LLVM >= 0x0800) {
1687 params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
1688 params[arg_count++] = ctx->ac.i32_0; /* soffset */
1689 params[arg_count++] = ctx->ac.i32_0; /* slc */
1690
1691 length = snprintf(intrinsic_name, sizeof(intrinsic_name),
1692 "llvm.amdgcn.raw.buffer.atomic.%s.i32",
1693 atomic_name);
1694 } else {
1695 params[arg_count++] = ctx->ac.i32_0; /* vindex */
1696 params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
1697 params[arg_count++] = ctx->ac.i1false; /* slc */
1698
1699 length = snprintf(intrinsic_name, sizeof(intrinsic_name),
1700 "llvm.amdgcn.buffer.atomic.%s", atomic_name);
1701 }
1702
1703 assert(length < sizeof(intrinsic_name));
1704 return ac_build_intrinsic(&ctx->ac, intrinsic_name, ctx->ac.i32,
1705 params, arg_count, 0);
1706 }
1707
1708 static LLVMValueRef visit_load_buffer(struct ac_nir_context *ctx,
1709 const nir_intrinsic_instr *instr)
1710 {
1711 int elem_size_bytes = instr->dest.ssa.bit_size / 8;
1712 int num_components = instr->num_components;
1713 enum gl_access_qualifier access = nir_intrinsic_access(instr);
1714 unsigned cache_policy = get_cache_policy(ctx, access, false, false);
1715
1716 LLVMValueRef offset = get_src(ctx, instr->src[1]);
1717 LLVMValueRef rsrc = ctx->abi->load_ssbo(ctx->abi,
1718 get_src(ctx, instr->src[0]), false);
1719 LLVMValueRef vindex = ctx->ac.i32_0;
1720
1721 LLVMTypeRef def_type = get_def_type(ctx, &instr->dest.ssa);
1722 LLVMTypeRef def_elem_type = num_components > 1 ? LLVMGetElementType(def_type) : def_type;
1723
1724 LLVMValueRef results[4];
1725 for (int i = 0; i < num_components;) {
1726 int num_elems = num_components - i;
1727 if (elem_size_bytes < 4 && nir_intrinsic_align(instr) % 4 != 0)
1728 num_elems = 1;
1729 if (num_elems * elem_size_bytes > 16)
1730 num_elems = 16 / elem_size_bytes;
1731 int load_bytes = num_elems * elem_size_bytes;
1732
1733 LLVMValueRef immoffset = LLVMConstInt(ctx->ac.i32, i * elem_size_bytes, false);
1734
1735 LLVMValueRef ret;
1736
1737 if (load_bytes == 1) {
1738 ret = ac_build_tbuffer_load_byte(&ctx->ac,
1739 rsrc,
1740 offset,
1741 ctx->ac.i32_0,
1742 immoffset,
1743 cache_policy & ac_glc);
1744 } else if (load_bytes == 2) {
1745 ret = ac_build_tbuffer_load_short(&ctx->ac,
1746 rsrc,
1747 offset,
1748 ctx->ac.i32_0,
1749 immoffset,
1750 cache_policy & ac_glc);
1751 } else {
1752 int num_channels = util_next_power_of_two(load_bytes) / 4;
1753
1754 ret = ac_build_buffer_load(&ctx->ac, rsrc, num_channels,
1755 vindex, offset, immoffset, 0,
1756 cache_policy & ac_glc, 0,
1757 false, false);
1758 }
1759
1760 LLVMTypeRef byte_vec = LLVMVectorType(ctx->ac.i8, ac_get_type_size(LLVMTypeOf(ret)));
1761 ret = LLVMBuildBitCast(ctx->ac.builder, ret, byte_vec, "");
1762 ret = ac_trim_vector(&ctx->ac, ret, load_bytes);
1763
1764 LLVMTypeRef ret_type = LLVMVectorType(def_elem_type, num_elems);
1765 ret = LLVMBuildBitCast(ctx->ac.builder, ret, ret_type, "");
1766
1767 for (unsigned j = 0; j < num_elems; j++) {
1768 results[i + j] = LLVMBuildExtractElement(ctx->ac.builder, ret, LLVMConstInt(ctx->ac.i32, j, false), "");
1769 }
1770 i += num_elems;
1771 }
1772
1773 return ac_build_gather_values(&ctx->ac, results, num_components);
1774 }
1775
1776 static LLVMValueRef visit_load_ubo_buffer(struct ac_nir_context *ctx,
1777 const nir_intrinsic_instr *instr)
1778 {
1779 LLVMValueRef ret;
1780 LLVMValueRef rsrc = get_src(ctx, instr->src[0]);
1781 LLVMValueRef offset = get_src(ctx, instr->src[1]);
1782 int num_components = instr->num_components;
1783
1784 if (ctx->abi->load_ubo)
1785 rsrc = ctx->abi->load_ubo(ctx->abi, rsrc);
1786
1787 if (instr->dest.ssa.bit_size == 64)
1788 num_components *= 2;
1789
1790 if (instr->dest.ssa.bit_size == 16 || instr->dest.ssa.bit_size == 8) {
1791 unsigned load_bytes = instr->dest.ssa.bit_size / 8;
1792 LLVMValueRef results[num_components];
1793 for (unsigned i = 0; i < num_components; ++i) {
1794 LLVMValueRef immoffset = LLVMConstInt(ctx->ac.i32,
1795 load_bytes * i, 0);
1796
1797 if (load_bytes == 1) {
1798 results[i] = ac_build_tbuffer_load_byte(&ctx->ac,
1799 rsrc,
1800 offset,
1801 ctx->ac.i32_0,
1802 immoffset,
1803 false);
1804 } else {
1805 assert(load_bytes == 2);
1806 results[i] = ac_build_tbuffer_load_short(&ctx->ac,
1807 rsrc,
1808 offset,
1809 ctx->ac.i32_0,
1810 immoffset,
1811 false);
1812 }
1813 }
1814 ret = ac_build_gather_values(&ctx->ac, results, num_components);
1815 } else {
1816 ret = ac_build_buffer_load(&ctx->ac, rsrc, num_components, NULL, offset,
1817 NULL, 0, false, false, true, true);
1818
1819 ret = ac_trim_vector(&ctx->ac, ret, num_components);
1820 }
1821
1822 return LLVMBuildBitCast(ctx->ac.builder, ret,
1823 get_def_type(ctx, &instr->dest.ssa), "");
1824 }
1825
1826 static void
1827 get_deref_offset(struct ac_nir_context *ctx, nir_deref_instr *instr,
1828 bool vs_in, unsigned *vertex_index_out,
1829 LLVMValueRef *vertex_index_ref,
1830 unsigned *const_out, LLVMValueRef *indir_out)
1831 {
1832 nir_variable *var = nir_deref_instr_get_variable(instr);
1833 nir_deref_path path;
1834 unsigned idx_lvl = 1;
1835
1836 nir_deref_path_init(&path, instr, NULL);
1837
1838 if (vertex_index_out != NULL || vertex_index_ref != NULL) {
1839 if (vertex_index_ref) {
1840 *vertex_index_ref = get_src(ctx, path.path[idx_lvl]->arr.index);
1841 if (vertex_index_out)
1842 *vertex_index_out = 0;
1843 } else {
1844 nir_const_value *v = nir_src_as_const_value(path.path[idx_lvl]->arr.index);
1845 assert(v);
1846 *vertex_index_out = v->u32[0];
1847 }
1848 ++idx_lvl;
1849 }
1850
1851 uint32_t const_offset = 0;
1852 LLVMValueRef offset = NULL;
1853
1854 if (var->data.compact) {
1855 assert(instr->deref_type == nir_deref_type_array);
1856 nir_const_value *v = nir_src_as_const_value(instr->arr.index);
1857 assert(v);
1858 const_offset = v->u32[0];
1859 goto out;
1860 }
1861
1862 for (; path.path[idx_lvl]; ++idx_lvl) {
1863 const struct glsl_type *parent_type = path.path[idx_lvl - 1]->type;
1864 if (path.path[idx_lvl]->deref_type == nir_deref_type_struct) {
1865 unsigned index = path.path[idx_lvl]->strct.index;
1866
1867 for (unsigned i = 0; i < index; i++) {
1868 const struct glsl_type *ft = glsl_get_struct_field(parent_type, i);
1869 const_offset += glsl_count_attribute_slots(ft, vs_in);
1870 }
1871 } else if(path.path[idx_lvl]->deref_type == nir_deref_type_array) {
1872 unsigned size = glsl_count_attribute_slots(path.path[idx_lvl]->type, vs_in);
1873 LLVMValueRef array_off = LLVMBuildMul(ctx->ac.builder, LLVMConstInt(ctx->ac.i32, size, 0),
1874 get_src(ctx, path.path[idx_lvl]->arr.index), "");
1875 if (offset)
1876 offset = LLVMBuildAdd(ctx->ac.builder, offset, array_off, "");
1877 else
1878 offset = array_off;
1879 } else
1880 unreachable("Uhandled deref type in get_deref_instr_offset");
1881 }
1882
1883 out:
1884 nir_deref_path_finish(&path);
1885
1886 if (const_offset && offset)
1887 offset = LLVMBuildAdd(ctx->ac.builder, offset,
1888 LLVMConstInt(ctx->ac.i32, const_offset, 0),
1889 "");
1890
1891 *const_out = const_offset;
1892 *indir_out = offset;
1893 }
1894
1895 static LLVMValueRef load_tess_varyings(struct ac_nir_context *ctx,
1896 nir_intrinsic_instr *instr,
1897 bool load_inputs)
1898 {
1899 LLVMValueRef result;
1900 LLVMValueRef vertex_index = NULL;
1901 LLVMValueRef indir_index = NULL;
1902 unsigned const_index = 0;
1903
1904 nir_variable *var = nir_deref_instr_get_variable(nir_instr_as_deref(instr->src[0].ssa->parent_instr));
1905
1906 unsigned location = var->data.location;
1907 unsigned driver_location = var->data.driver_location;
1908 const bool is_patch = var->data.patch;
1909 const bool is_compact = var->data.compact;
1910
1911 get_deref_offset(ctx, nir_instr_as_deref(instr->src[0].ssa->parent_instr),
1912 false, NULL, is_patch ? NULL : &vertex_index,
1913 &const_index, &indir_index);
1914
1915 LLVMTypeRef dest_type = get_def_type(ctx, &instr->dest.ssa);
1916
1917 LLVMTypeRef src_component_type;
1918 if (LLVMGetTypeKind(dest_type) == LLVMVectorTypeKind)
1919 src_component_type = LLVMGetElementType(dest_type);
1920 else
1921 src_component_type = dest_type;
1922
1923 result = ctx->abi->load_tess_varyings(ctx->abi, src_component_type,
1924 vertex_index, indir_index,
1925 const_index, location, driver_location,
1926 var->data.location_frac,
1927 instr->num_components,
1928 is_patch, is_compact, load_inputs);
1929 if (instr->dest.ssa.bit_size == 16) {
1930 result = ac_to_integer(&ctx->ac, result);
1931 result = LLVMBuildTrunc(ctx->ac.builder, result, dest_type, "");
1932 }
1933 return LLVMBuildBitCast(ctx->ac.builder, result, dest_type, "");
1934 }
1935
1936 static unsigned
1937 type_scalar_size_bytes(const struct glsl_type *type)
1938 {
1939 assert(glsl_type_is_vector_or_scalar(type) ||
1940 glsl_type_is_matrix(type));
1941 return glsl_type_is_boolean(type) ? 4 : glsl_get_bit_size(type) / 8;
1942 }
1943
1944 static LLVMValueRef visit_load_var(struct ac_nir_context *ctx,
1945 nir_intrinsic_instr *instr)
1946 {
1947 nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
1948 nir_variable *var = nir_deref_instr_get_variable(deref);
1949
1950 LLVMValueRef values[8];
1951 int idx = 0;
1952 int ve = instr->dest.ssa.num_components;
1953 unsigned comp = 0;
1954 LLVMValueRef indir_index;
1955 LLVMValueRef ret;
1956 unsigned const_index;
1957 unsigned stride = 4;
1958 int mode = deref->mode;
1959
1960 if (var) {
1961 bool vs_in = ctx->stage == MESA_SHADER_VERTEX &&
1962 var->data.mode == nir_var_shader_in;
1963 idx = var->data.driver_location;
1964 comp = var->data.location_frac;
1965 mode = var->data.mode;
1966
1967 get_deref_offset(ctx, deref, vs_in, NULL, NULL,
1968 &const_index, &indir_index);
1969
1970 if (var->data.compact) {
1971 stride = 1;
1972 const_index += comp;
1973 comp = 0;
1974 }
1975 }
1976
1977 if (instr->dest.ssa.bit_size == 64 &&
1978 (deref->mode == nir_var_shader_in ||
1979 deref->mode == nir_var_shader_out ||
1980 deref->mode == nir_var_function_temp))
1981 ve *= 2;
1982
1983 switch (mode) {
1984 case nir_var_shader_in:
1985 if (ctx->stage == MESA_SHADER_TESS_CTRL ||
1986 ctx->stage == MESA_SHADER_TESS_EVAL) {
1987 return load_tess_varyings(ctx, instr, true);
1988 }
1989
1990 if (ctx->stage == MESA_SHADER_GEOMETRY) {
1991 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.bit_size);
1992 LLVMValueRef indir_index;
1993 unsigned const_index, vertex_index;
1994 get_deref_offset(ctx, deref, false, &vertex_index, NULL,
1995 &const_index, &indir_index);
1996
1997 return ctx->abi->load_inputs(ctx->abi, var->data.location,
1998 var->data.driver_location,
1999 var->data.location_frac,
2000 instr->num_components, vertex_index, const_index, type);
2001 }
2002
2003 for (unsigned chan = comp; chan < ve + comp; chan++) {
2004 if (indir_index) {
2005 unsigned count = glsl_count_attribute_slots(
2006 var->type,
2007 ctx->stage == MESA_SHADER_VERTEX);
2008 count -= chan / 4;
2009 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2010 &ctx->ac, ctx->abi->inputs + idx + chan, count,
2011 stride, false, true);
2012
2013 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2014 tmp_vec,
2015 indir_index, "");
2016 } else
2017 values[chan] = ctx->abi->inputs[idx + chan + const_index * stride];
2018 }
2019 break;
2020 case nir_var_function_temp:
2021 for (unsigned chan = 0; chan < ve; chan++) {
2022 if (indir_index) {
2023 unsigned count = glsl_count_attribute_slots(
2024 var->type, false);
2025 count -= chan / 4;
2026 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2027 &ctx->ac, ctx->locals + idx + chan, count,
2028 stride, true, true);
2029
2030 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2031 tmp_vec,
2032 indir_index, "");
2033 } else {
2034 values[chan] = LLVMBuildLoad(ctx->ac.builder, ctx->locals[idx + chan + const_index * stride], "");
2035 }
2036 }
2037 break;
2038 case nir_var_mem_shared: {
2039 LLVMValueRef address = get_src(ctx, instr->src[0]);
2040 LLVMValueRef val = LLVMBuildLoad(ctx->ac.builder, address, "");
2041 return LLVMBuildBitCast(ctx->ac.builder, val,
2042 get_def_type(ctx, &instr->dest.ssa),
2043 "");
2044 }
2045 case nir_var_shader_out:
2046 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
2047 return load_tess_varyings(ctx, instr, false);
2048 }
2049
2050 for (unsigned chan = comp; chan < ve + comp; chan++) {
2051 if (indir_index) {
2052 unsigned count = glsl_count_attribute_slots(
2053 var->type, false);
2054 count -= chan / 4;
2055 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2056 &ctx->ac, ctx->abi->outputs + idx + chan, count,
2057 stride, true, true);
2058
2059 values[chan] = LLVMBuildExtractElement(ctx->ac.builder,
2060 tmp_vec,
2061 indir_index, "");
2062 } else {
2063 values[chan] = LLVMBuildLoad(ctx->ac.builder,
2064 ctx->abi->outputs[idx + chan + const_index * stride],
2065 "");
2066 }
2067 }
2068 break;
2069 case nir_var_mem_global: {
2070 LLVMValueRef address = get_src(ctx, instr->src[0]);
2071 unsigned explicit_stride = glsl_get_explicit_stride(deref->type);
2072 unsigned natural_stride = type_scalar_size_bytes(deref->type);
2073 unsigned stride = explicit_stride ? explicit_stride : natural_stride;
2074
2075 LLVMTypeRef result_type = get_def_type(ctx, &instr->dest.ssa);
2076 if (stride != natural_stride) {
2077 LLVMTypeRef ptr_type = LLVMPointerType(LLVMGetElementType(result_type),
2078 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2079 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2080
2081 for (unsigned i = 0; i < instr->dest.ssa.num_components; ++i) {
2082 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, i * stride / natural_stride, 0);
2083 values[i] = LLVMBuildLoad(ctx->ac.builder,
2084 ac_build_gep_ptr(&ctx->ac, address, offset), "");
2085 }
2086 return ac_build_gather_values(&ctx->ac, values, instr->dest.ssa.num_components);
2087 } else {
2088 LLVMTypeRef ptr_type = LLVMPointerType(result_type,
2089 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2090 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2091 LLVMValueRef val = LLVMBuildLoad(ctx->ac.builder, address, "");
2092 return val;
2093 }
2094 }
2095 default:
2096 unreachable("unhandle variable mode");
2097 }
2098 ret = ac_build_varying_gather_values(&ctx->ac, values, ve, comp);
2099 return LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->dest.ssa), "");
2100 }
2101
2102 static void
2103 visit_store_var(struct ac_nir_context *ctx,
2104 nir_intrinsic_instr *instr)
2105 {
2106 nir_deref_instr *deref = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2107 nir_variable *var = nir_deref_instr_get_variable(deref);
2108
2109 LLVMValueRef temp_ptr, value;
2110 int idx = 0;
2111 unsigned comp = 0;
2112 LLVMValueRef src = ac_to_float(&ctx->ac, get_src(ctx, instr->src[1]));
2113 int writemask = instr->const_index[0];
2114 LLVMValueRef indir_index;
2115 unsigned const_index;
2116
2117 if (var) {
2118 get_deref_offset(ctx, deref, false,
2119 NULL, NULL, &const_index, &indir_index);
2120 idx = var->data.driver_location;
2121 comp = var->data.location_frac;
2122
2123 if (var->data.compact) {
2124 const_index += comp;
2125 comp = 0;
2126 }
2127 }
2128
2129 if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src)) == 64 &&
2130 (deref->mode == nir_var_shader_out ||
2131 deref->mode == nir_var_function_temp)) {
2132
2133 src = LLVMBuildBitCast(ctx->ac.builder, src,
2134 LLVMVectorType(ctx->ac.f32, ac_get_llvm_num_components(src) * 2),
2135 "");
2136
2137 writemask = widen_mask(writemask, 2);
2138 }
2139
2140 writemask = writemask << comp;
2141
2142 switch (deref->mode) {
2143 case nir_var_shader_out:
2144
2145 if (ctx->stage == MESA_SHADER_TESS_CTRL) {
2146 LLVMValueRef vertex_index = NULL;
2147 LLVMValueRef indir_index = NULL;
2148 unsigned const_index = 0;
2149 const bool is_patch = var->data.patch;
2150
2151 get_deref_offset(ctx, deref, false, NULL,
2152 is_patch ? NULL : &vertex_index,
2153 &const_index, &indir_index);
2154
2155 ctx->abi->store_tcs_outputs(ctx->abi, var,
2156 vertex_index, indir_index,
2157 const_index, src, writemask);
2158 return;
2159 }
2160
2161 for (unsigned chan = 0; chan < 8; chan++) {
2162 int stride = 4;
2163 if (!(writemask & (1 << chan)))
2164 continue;
2165
2166 value = ac_llvm_extract_elem(&ctx->ac, src, chan - comp);
2167
2168 if (var->data.compact)
2169 stride = 1;
2170 if (indir_index) {
2171 unsigned count = glsl_count_attribute_slots(
2172 var->type, false);
2173 count -= chan / 4;
2174 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2175 &ctx->ac, ctx->abi->outputs + idx + chan, count,
2176 stride, true, true);
2177
2178 tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
2179 value, indir_index, "");
2180 build_store_values_extended(&ctx->ac, ctx->abi->outputs + idx + chan,
2181 count, stride, tmp_vec);
2182
2183 } else {
2184 temp_ptr = ctx->abi->outputs[idx + chan + const_index * stride];
2185
2186 LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
2187 }
2188 }
2189 break;
2190 case nir_var_function_temp:
2191 for (unsigned chan = 0; chan < 8; chan++) {
2192 if (!(writemask & (1 << chan)))
2193 continue;
2194
2195 value = ac_llvm_extract_elem(&ctx->ac, src, chan);
2196 if (indir_index) {
2197 unsigned count = glsl_count_attribute_slots(
2198 var->type, false);
2199 count -= chan / 4;
2200 LLVMValueRef tmp_vec = ac_build_gather_values_extended(
2201 &ctx->ac, ctx->locals + idx + chan, count,
2202 4, true, true);
2203
2204 tmp_vec = LLVMBuildInsertElement(ctx->ac.builder, tmp_vec,
2205 value, indir_index, "");
2206 build_store_values_extended(&ctx->ac, ctx->locals + idx + chan,
2207 count, 4, tmp_vec);
2208 } else {
2209 temp_ptr = ctx->locals[idx + chan + const_index * 4];
2210
2211 LLVMBuildStore(ctx->ac.builder, value, temp_ptr);
2212 }
2213 }
2214 break;
2215
2216 case nir_var_mem_global:
2217 case nir_var_mem_shared: {
2218 int writemask = instr->const_index[0];
2219 LLVMValueRef address = get_src(ctx, instr->src[0]);
2220 LLVMValueRef val = get_src(ctx, instr->src[1]);
2221
2222 unsigned explicit_stride = glsl_get_explicit_stride(deref->type);
2223 unsigned natural_stride = type_scalar_size_bytes(deref->type);
2224 unsigned stride = explicit_stride ? explicit_stride : natural_stride;
2225
2226 LLVMTypeRef ptr_type = LLVMPointerType(LLVMTypeOf(val),
2227 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2228 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2229
2230 if (writemask == (1u << ac_get_llvm_num_components(val)) - 1 &&
2231 stride == natural_stride) {
2232 LLVMTypeRef ptr_type = LLVMPointerType(LLVMTypeOf(val),
2233 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2234 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2235
2236 val = LLVMBuildBitCast(ctx->ac.builder, val,
2237 LLVMGetElementType(LLVMTypeOf(address)), "");
2238 LLVMBuildStore(ctx->ac.builder, val, address);
2239 } else {
2240 LLVMTypeRef ptr_type = LLVMPointerType(LLVMGetElementType(LLVMTypeOf(val)),
2241 LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
2242 address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
2243 for (unsigned chan = 0; chan < 4; chan++) {
2244 if (!(writemask & (1 << chan)))
2245 continue;
2246
2247 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, chan * stride / natural_stride, 0);
2248
2249 LLVMValueRef ptr = ac_build_gep_ptr(&ctx->ac, address, offset);
2250 LLVMValueRef src = ac_llvm_extract_elem(&ctx->ac, val,
2251 chan);
2252 src = LLVMBuildBitCast(ctx->ac.builder, src,
2253 LLVMGetElementType(LLVMTypeOf(ptr)), "");
2254 LLVMBuildStore(ctx->ac.builder, src, ptr);
2255 }
2256 }
2257 break;
2258 }
2259 default:
2260 abort();
2261 break;
2262 }
2263 }
2264
2265 static int image_type_to_components_count(enum glsl_sampler_dim dim, bool array)
2266 {
2267 switch (dim) {
2268 case GLSL_SAMPLER_DIM_BUF:
2269 return 1;
2270 case GLSL_SAMPLER_DIM_1D:
2271 return array ? 2 : 1;
2272 case GLSL_SAMPLER_DIM_2D:
2273 return array ? 3 : 2;
2274 case GLSL_SAMPLER_DIM_MS:
2275 return array ? 4 : 3;
2276 case GLSL_SAMPLER_DIM_3D:
2277 case GLSL_SAMPLER_DIM_CUBE:
2278 return 3;
2279 case GLSL_SAMPLER_DIM_RECT:
2280 case GLSL_SAMPLER_DIM_SUBPASS:
2281 return 2;
2282 case GLSL_SAMPLER_DIM_SUBPASS_MS:
2283 return 3;
2284 default:
2285 break;
2286 }
2287 return 0;
2288 }
2289
2290
2291 /* Adjust the sample index according to FMASK.
2292 *
2293 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
2294 * which is the identity mapping. Each nibble says which physical sample
2295 * should be fetched to get that sample.
2296 *
2297 * For example, 0x11111100 means there are only 2 samples stored and
2298 * the second sample covers 3/4 of the pixel. When reading samples 0
2299 * and 1, return physical sample 0 (determined by the first two 0s
2300 * in FMASK), otherwise return physical sample 1.
2301 *
2302 * The sample index should be adjusted as follows:
2303 * sample_index = (fmask >> (sample_index * 4)) & 0xF;
2304 */
2305 static LLVMValueRef adjust_sample_index_using_fmask(struct ac_llvm_context *ctx,
2306 LLVMValueRef coord_x, LLVMValueRef coord_y,
2307 LLVMValueRef coord_z,
2308 LLVMValueRef sample_index,
2309 LLVMValueRef fmask_desc_ptr)
2310 {
2311 struct ac_image_args args = {0};
2312 LLVMValueRef res;
2313
2314 args.coords[0] = coord_x;
2315 args.coords[1] = coord_y;
2316 if (coord_z)
2317 args.coords[2] = coord_z;
2318
2319 args.opcode = ac_image_load;
2320 args.dim = coord_z ? ac_image_2darray : ac_image_2d;
2321 args.resource = fmask_desc_ptr;
2322 args.dmask = 0xf;
2323 args.attributes = AC_FUNC_ATTR_READNONE;
2324
2325 res = ac_build_image_opcode(ctx, &args);
2326
2327 res = ac_to_integer(ctx, res);
2328 LLVMValueRef four = LLVMConstInt(ctx->i32, 4, false);
2329 LLVMValueRef F = LLVMConstInt(ctx->i32, 0xf, false);
2330
2331 LLVMValueRef fmask = LLVMBuildExtractElement(ctx->builder,
2332 res,
2333 ctx->i32_0, "");
2334
2335 LLVMValueRef sample_index4 =
2336 LLVMBuildMul(ctx->builder, sample_index, four, "");
2337 LLVMValueRef shifted_fmask =
2338 LLVMBuildLShr(ctx->builder, fmask, sample_index4, "");
2339 LLVMValueRef final_sample =
2340 LLVMBuildAnd(ctx->builder, shifted_fmask, F, "");
2341
2342 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
2343 * resource descriptor is 0 (invalid),
2344 */
2345 LLVMValueRef fmask_desc =
2346 LLVMBuildBitCast(ctx->builder, fmask_desc_ptr,
2347 ctx->v8i32, "");
2348
2349 LLVMValueRef fmask_word1 =
2350 LLVMBuildExtractElement(ctx->builder, fmask_desc,
2351 ctx->i32_1, "");
2352
2353 LLVMValueRef word1_is_nonzero =
2354 LLVMBuildICmp(ctx->builder, LLVMIntNE,
2355 fmask_word1, ctx->i32_0, "");
2356
2357 /* Replace the MSAA sample index. */
2358 sample_index =
2359 LLVMBuildSelect(ctx->builder, word1_is_nonzero,
2360 final_sample, sample_index, "");
2361 return sample_index;
2362 }
2363
2364 static nir_deref_instr *get_image_deref(const nir_intrinsic_instr *instr)
2365 {
2366 assert(instr->src[0].is_ssa);
2367 return nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2368 }
2369
2370 static LLVMValueRef get_image_descriptor(struct ac_nir_context *ctx,
2371 const nir_intrinsic_instr *instr,
2372 enum ac_descriptor_type desc_type,
2373 bool write)
2374 {
2375 return get_sampler_desc(ctx, nir_instr_as_deref(instr->src[0].ssa->parent_instr), desc_type, NULL, true, write);
2376 }
2377
2378 static void get_image_coords(struct ac_nir_context *ctx,
2379 const nir_intrinsic_instr *instr,
2380 struct ac_image_args *args)
2381 {
2382 const struct glsl_type *type = get_image_deref(instr)->type;
2383
2384 LLVMValueRef src0 = get_src(ctx, instr->src[1]);
2385 LLVMValueRef masks[] = {
2386 LLVMConstInt(ctx->ac.i32, 0, false), LLVMConstInt(ctx->ac.i32, 1, false),
2387 LLVMConstInt(ctx->ac.i32, 2, false), LLVMConstInt(ctx->ac.i32, 3, false),
2388 };
2389 LLVMValueRef sample_index = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
2390
2391 int count;
2392 enum glsl_sampler_dim dim = glsl_get_sampler_dim(type);
2393 bool is_array = glsl_sampler_type_is_array(type);
2394 bool add_frag_pos = (dim == GLSL_SAMPLER_DIM_SUBPASS ||
2395 dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2396 bool is_ms = (dim == GLSL_SAMPLER_DIM_MS ||
2397 dim == GLSL_SAMPLER_DIM_SUBPASS_MS);
2398 bool gfx9_1d = ctx->ac.chip_class >= GFX9 && dim == GLSL_SAMPLER_DIM_1D;
2399 count = image_type_to_components_count(dim, is_array);
2400
2401 if (is_ms && instr->intrinsic == nir_intrinsic_image_deref_load) {
2402 LLVMValueRef fmask_load_address[3];
2403 int chan;
2404
2405 fmask_load_address[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
2406 fmask_load_address[1] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[1], "");
2407 if (is_array)
2408 fmask_load_address[2] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[2], "");
2409 else
2410 fmask_load_address[2] = NULL;
2411 if (add_frag_pos) {
2412 for (chan = 0; chan < 2; ++chan)
2413 fmask_load_address[chan] =
2414 LLVMBuildAdd(ctx->ac.builder, fmask_load_address[chan],
2415 LLVMBuildFPToUI(ctx->ac.builder, ctx->abi->frag_pos[chan],
2416 ctx->ac.i32, ""), "");
2417 fmask_load_address[2] = ac_to_integer(&ctx->ac, ctx->abi->inputs[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER, 0)]);
2418 }
2419 sample_index = adjust_sample_index_using_fmask(&ctx->ac,
2420 fmask_load_address[0],
2421 fmask_load_address[1],
2422 fmask_load_address[2],
2423 sample_index,
2424 get_sampler_desc(ctx, nir_instr_as_deref(instr->src[0].ssa->parent_instr),
2425 AC_DESC_FMASK, NULL, false, false));
2426 }
2427 if (count == 1 && !gfx9_1d) {
2428 if (instr->src[1].ssa->num_components)
2429 args->coords[0] = LLVMBuildExtractElement(ctx->ac.builder, src0, masks[0], "");
2430 else
2431 args->coords[0] = src0;
2432 } else {
2433 int chan;
2434 if (is_ms)
2435 count--;
2436 for (chan = 0; chan < count; ++chan) {
2437 args->coords[chan] = ac_llvm_extract_elem(&ctx->ac, src0, chan);
2438 }
2439 if (add_frag_pos) {
2440 for (chan = 0; chan < 2; ++chan) {
2441 args->coords[chan] = LLVMBuildAdd(
2442 ctx->ac.builder, args->coords[chan],
2443 LLVMBuildFPToUI(
2444 ctx->ac.builder, ctx->abi->frag_pos[chan],
2445 ctx->ac.i32, ""), "");
2446 }
2447 args->coords[2] = ac_to_integer(&ctx->ac,
2448 ctx->abi->inputs[ac_llvm_reg_index_soa(VARYING_SLOT_LAYER, 0)]);
2449 count++;
2450 }
2451
2452 if (gfx9_1d) {
2453 if (is_array) {
2454 args->coords[2] = args->coords[1];
2455 args->coords[1] = ctx->ac.i32_0;
2456 } else
2457 args->coords[1] = ctx->ac.i32_0;
2458 count++;
2459 }
2460
2461 if (is_ms) {
2462 args->coords[count] = sample_index;
2463 count++;
2464 }
2465 }
2466 }
2467
2468 static LLVMValueRef get_image_buffer_descriptor(struct ac_nir_context *ctx,
2469 const nir_intrinsic_instr *instr, bool write)
2470 {
2471 LLVMValueRef rsrc = get_image_descriptor(ctx, instr, AC_DESC_BUFFER, write);
2472 if (ctx->abi->gfx9_stride_size_workaround) {
2473 LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 2, 0), "");
2474 LLVMValueRef stride = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 1, 0), "");
2475 stride = LLVMBuildLShr(ctx->ac.builder, stride, LLVMConstInt(ctx->ac.i32, 16, 0), "");
2476
2477 LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->ac.builder,
2478 LLVMBuildICmp(ctx->ac.builder, LLVMIntUGT, elem_count, stride, ""),
2479 elem_count, stride, "");
2480
2481 rsrc = LLVMBuildInsertElement(ctx->ac.builder, rsrc, new_elem_count,
2482 LLVMConstInt(ctx->ac.i32, 2, 0), "");
2483 }
2484 return rsrc;
2485 }
2486
2487 static LLVMValueRef visit_image_load(struct ac_nir_context *ctx,
2488 const nir_intrinsic_instr *instr)
2489 {
2490 LLVMValueRef res;
2491 const nir_deref_instr *image_deref = get_image_deref(instr);
2492 const struct glsl_type *type = image_deref->type;
2493 const nir_variable *var = nir_deref_instr_get_variable(image_deref);
2494 struct ac_image_args args = {};
2495
2496 args.cache_policy =
2497 get_cache_policy(ctx, var->data.image.access, false, false);
2498
2499 const enum glsl_sampler_dim dim = glsl_get_sampler_dim(type);
2500 if (dim == GLSL_SAMPLER_DIM_BUF) {
2501 unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
2502 unsigned num_channels = util_last_bit(mask);
2503 LLVMValueRef rsrc, vindex;
2504
2505 rsrc = get_image_buffer_descriptor(ctx, instr, false);
2506 vindex = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
2507 ctx->ac.i32_0, "");
2508
2509 /* TODO: set "can_speculate" when OpenGL needs it. */
2510 res = ac_build_buffer_load_format(&ctx->ac, rsrc, vindex,
2511 ctx->ac.i32_0, num_channels,
2512 !!(args.cache_policy & ac_glc),
2513 false);
2514 res = ac_build_expand_to_vec4(&ctx->ac, res, num_channels);
2515
2516 res = ac_trim_vector(&ctx->ac, res, instr->dest.ssa.num_components);
2517 res = ac_to_integer(&ctx->ac, res);
2518 } else {
2519 args.opcode = ac_image_load;
2520 get_image_coords(ctx, instr, &args);
2521 args.resource = get_image_descriptor(ctx, instr, AC_DESC_IMAGE, false);
2522 args.dim = get_ac_image_dim(&ctx->ac, glsl_get_sampler_dim(type),
2523 glsl_sampler_type_is_array(type));
2524 args.dmask = 15;
2525 args.attributes = AC_FUNC_ATTR_READONLY;
2526
2527 res = ac_build_image_opcode(&ctx->ac, &args);
2528 }
2529 return ac_to_integer(&ctx->ac, res);
2530 }
2531
2532 static void visit_image_store(struct ac_nir_context *ctx,
2533 nir_intrinsic_instr *instr)
2534 {
2535 const nir_deref_instr *image_deref = get_image_deref(instr);
2536 const struct glsl_type *type = image_deref->type;
2537 const nir_variable *var = nir_deref_instr_get_variable(image_deref);
2538 const enum glsl_sampler_dim dim = glsl_get_sampler_dim(type);
2539 bool writeonly_memory = var->data.image.access & ACCESS_NON_READABLE;
2540 struct ac_image_args args = {};
2541
2542 args.cache_policy = get_cache_policy(ctx, var->data.image.access, true,
2543 writeonly_memory);
2544
2545 if (dim == GLSL_SAMPLER_DIM_BUF) {
2546 LLVMValueRef rsrc = get_image_buffer_descriptor(ctx, instr, true);
2547 LLVMValueRef src = ac_to_float(&ctx->ac, get_src(ctx, instr->src[3]));
2548 unsigned src_channels = ac_get_llvm_num_components(src);
2549 LLVMValueRef vindex;
2550
2551 if (src_channels == 3)
2552 src = ac_build_expand_to_vec4(&ctx->ac, src, 3);
2553
2554 vindex = LLVMBuildExtractElement(ctx->ac.builder,
2555 get_src(ctx, instr->src[1]),
2556 ctx->ac.i32_0, "");
2557
2558 ac_build_buffer_store_format(&ctx->ac, rsrc, src, vindex,
2559 ctx->ac.i32_0, src_channels,
2560 args.cache_policy & ac_glc,
2561 writeonly_memory);
2562 } else {
2563 args.opcode = ac_image_store;
2564 args.data[0] = ac_to_float(&ctx->ac, get_src(ctx, instr->src[3]));
2565 get_image_coords(ctx, instr, &args);
2566 args.resource = get_image_descriptor(ctx, instr, AC_DESC_IMAGE, true);
2567 args.dim = get_ac_image_dim(&ctx->ac, glsl_get_sampler_dim(type),
2568 glsl_sampler_type_is_array(type));
2569 args.dmask = 15;
2570
2571 ac_build_image_opcode(&ctx->ac, &args);
2572 }
2573
2574 }
2575
2576 static LLVMValueRef visit_image_atomic(struct ac_nir_context *ctx,
2577 const nir_intrinsic_instr *instr)
2578 {
2579 LLVMValueRef params[7];
2580 int param_count = 0;
2581 const struct glsl_type *type = get_image_deref(instr)->type;
2582
2583 bool cmpswap = instr->intrinsic == nir_intrinsic_image_deref_atomic_comp_swap;
2584 const char *atomic_name;
2585 char intrinsic_name[64];
2586 enum ac_atomic_op atomic_subop;
2587 MAYBE_UNUSED int length;
2588
2589 bool is_unsigned = glsl_get_sampler_result_type(type) == GLSL_TYPE_UINT;
2590
2591 switch (instr->intrinsic) {
2592 case nir_intrinsic_image_deref_atomic_add:
2593 atomic_name = "add";
2594 atomic_subop = ac_atomic_add;
2595 break;
2596 case nir_intrinsic_image_deref_atomic_min:
2597 atomic_name = is_unsigned ? "umin" : "smin";
2598 atomic_subop = is_unsigned ? ac_atomic_umin : ac_atomic_smin;
2599 break;
2600 case nir_intrinsic_image_deref_atomic_max:
2601 atomic_name = is_unsigned ? "umax" : "smax";
2602 atomic_subop = is_unsigned ? ac_atomic_umax : ac_atomic_smax;
2603 break;
2604 case nir_intrinsic_image_deref_atomic_and:
2605 atomic_name = "and";
2606 atomic_subop = ac_atomic_and;
2607 break;
2608 case nir_intrinsic_image_deref_atomic_or:
2609 atomic_name = "or";
2610 atomic_subop = ac_atomic_or;
2611 break;
2612 case nir_intrinsic_image_deref_atomic_xor:
2613 atomic_name = "xor";
2614 atomic_subop = ac_atomic_xor;
2615 break;
2616 case nir_intrinsic_image_deref_atomic_exchange:
2617 atomic_name = "swap";
2618 atomic_subop = ac_atomic_swap;
2619 break;
2620 case nir_intrinsic_image_deref_atomic_comp_swap:
2621 atomic_name = "cmpswap";
2622 atomic_subop = 0; /* not used */
2623 break;
2624 default:
2625 abort();
2626 }
2627
2628 if (cmpswap)
2629 params[param_count++] = get_src(ctx, instr->src[4]);
2630 params[param_count++] = get_src(ctx, instr->src[3]);
2631
2632 if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_BUF) {
2633 params[param_count++] = get_image_buffer_descriptor(ctx, instr, true);
2634 params[param_count++] = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
2635 ctx->ac.i32_0, ""); /* vindex */
2636 params[param_count++] = ctx->ac.i32_0; /* voffset */
2637 if (HAVE_LLVM >= 0x800) {
2638 params[param_count++] = ctx->ac.i32_0; /* soffset */
2639 params[param_count++] = ctx->ac.i32_0; /* slc */
2640
2641 length = snprintf(intrinsic_name, sizeof(intrinsic_name),
2642 "llvm.amdgcn.struct.buffer.atomic.%s.i32", atomic_name);
2643 } else {
2644 params[param_count++] = ctx->ac.i1false; /* slc */
2645
2646 length = snprintf(intrinsic_name, sizeof(intrinsic_name),
2647 "llvm.amdgcn.buffer.atomic.%s", atomic_name);
2648 }
2649
2650 assert(length < sizeof(intrinsic_name));
2651 return ac_build_intrinsic(&ctx->ac, intrinsic_name, ctx->ac.i32,
2652 params, param_count, 0);
2653 } else {
2654 struct ac_image_args args = {};
2655 args.opcode = cmpswap ? ac_image_atomic_cmpswap : ac_image_atomic;
2656 args.atomic = atomic_subop;
2657 args.data[0] = params[0];
2658 if (cmpswap)
2659 args.data[1] = params[1];
2660 get_image_coords(ctx, instr, &args);
2661 args.resource = get_image_descriptor(ctx, instr, AC_DESC_IMAGE, true);
2662 args.dim = get_ac_image_dim(&ctx->ac, glsl_get_sampler_dim(type),
2663 glsl_sampler_type_is_array(type));
2664
2665 return ac_build_image_opcode(&ctx->ac, &args);
2666 }
2667 }
2668
2669 static LLVMValueRef visit_image_samples(struct ac_nir_context *ctx,
2670 const nir_intrinsic_instr *instr)
2671 {
2672 const struct glsl_type *type = get_image_deref(instr)->type;
2673
2674 struct ac_image_args args = { 0 };
2675 args.dim = get_ac_sampler_dim(&ctx->ac, glsl_get_sampler_dim(type),
2676 glsl_sampler_type_is_array(type));
2677 args.dmask = 0xf;
2678 args.resource = get_image_descriptor(ctx, instr, AC_DESC_IMAGE, false);
2679 args.opcode = ac_image_get_resinfo;
2680 args.lod = ctx->ac.i32_0;
2681 args.attributes = AC_FUNC_ATTR_READNONE;
2682
2683 return ac_build_image_opcode(&ctx->ac, &args);
2684 }
2685
2686 static LLVMValueRef visit_image_size(struct ac_nir_context *ctx,
2687 const nir_intrinsic_instr *instr)
2688 {
2689 LLVMValueRef res;
2690 const struct glsl_type *type = get_image_deref(instr)->type;
2691
2692 if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_BUF)
2693 return get_buffer_size(ctx, get_image_descriptor(ctx, instr, AC_DESC_BUFFER, false), true);
2694
2695 struct ac_image_args args = { 0 };
2696
2697 args.dim = get_ac_image_dim(&ctx->ac, glsl_get_sampler_dim(type),
2698 glsl_sampler_type_is_array(type));
2699 args.dmask = 0xf;
2700 args.resource = get_image_descriptor(ctx, instr, AC_DESC_IMAGE, false);
2701 args.opcode = ac_image_get_resinfo;
2702 args.lod = ctx->ac.i32_0;
2703 args.attributes = AC_FUNC_ATTR_READNONE;
2704
2705 res = ac_build_image_opcode(&ctx->ac, &args);
2706
2707 LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
2708
2709 if (glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_CUBE &&
2710 glsl_sampler_type_is_array(type)) {
2711 LLVMValueRef six = LLVMConstInt(ctx->ac.i32, 6, false);
2712 LLVMValueRef z = LLVMBuildExtractElement(ctx->ac.builder, res, two, "");
2713 z = LLVMBuildSDiv(ctx->ac.builder, z, six, "");
2714 res = LLVMBuildInsertElement(ctx->ac.builder, res, z, two, "");
2715 }
2716 if (ctx->ac.chip_class >= GFX9 &&
2717 glsl_get_sampler_dim(type) == GLSL_SAMPLER_DIM_1D &&
2718 glsl_sampler_type_is_array(type)) {
2719 LLVMValueRef layers = LLVMBuildExtractElement(ctx->ac.builder, res, two, "");
2720 res = LLVMBuildInsertElement(ctx->ac.builder, res, layers,
2721 ctx->ac.i32_1, "");
2722
2723 }
2724 return res;
2725 }
2726
2727 static void emit_membar(struct ac_llvm_context *ac,
2728 const nir_intrinsic_instr *instr)
2729 {
2730 unsigned waitcnt = NOOP_WAITCNT;
2731
2732 switch (instr->intrinsic) {
2733 case nir_intrinsic_memory_barrier:
2734 case nir_intrinsic_group_memory_barrier:
2735 waitcnt &= VM_CNT & LGKM_CNT;
2736 break;
2737 case nir_intrinsic_memory_barrier_atomic_counter:
2738 case nir_intrinsic_memory_barrier_buffer:
2739 case nir_intrinsic_memory_barrier_image:
2740 waitcnt &= VM_CNT;
2741 break;
2742 case nir_intrinsic_memory_barrier_shared:
2743 waitcnt &= LGKM_CNT;
2744 break;
2745 default:
2746 break;
2747 }
2748 if (waitcnt != NOOP_WAITCNT)
2749 ac_build_waitcnt(ac, waitcnt);
2750 }
2751
2752 void ac_emit_barrier(struct ac_llvm_context *ac, gl_shader_stage stage)
2753 {
2754 /* SI only (thanks to a hw bug workaround):
2755 * The real barrier instruction isn’t needed, because an entire patch
2756 * always fits into a single wave.
2757 */
2758 if (ac->chip_class == SI && stage == MESA_SHADER_TESS_CTRL) {
2759 ac_build_waitcnt(ac, LGKM_CNT & VM_CNT);
2760 return;
2761 }
2762 ac_build_s_barrier(ac);
2763 }
2764
2765 static void emit_discard(struct ac_nir_context *ctx,
2766 const nir_intrinsic_instr *instr)
2767 {
2768 LLVMValueRef cond;
2769
2770 if (instr->intrinsic == nir_intrinsic_discard_if) {
2771 cond = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ,
2772 get_src(ctx, instr->src[0]),
2773 ctx->ac.i32_0, "");
2774 } else {
2775 assert(instr->intrinsic == nir_intrinsic_discard);
2776 cond = ctx->ac.i1false;
2777 }
2778
2779 ctx->abi->emit_kill(ctx->abi, cond);
2780 }
2781
2782 static LLVMValueRef
2783 visit_load_helper_invocation(struct ac_nir_context *ctx)
2784 {
2785 LLVMValueRef result = ac_build_intrinsic(&ctx->ac,
2786 "llvm.amdgcn.ps.live",
2787 ctx->ac.i1, NULL, 0,
2788 AC_FUNC_ATTR_READNONE);
2789 result = LLVMBuildNot(ctx->ac.builder, result, "");
2790 return LLVMBuildSExt(ctx->ac.builder, result, ctx->ac.i32, "");
2791 }
2792
2793 static LLVMValueRef
2794 visit_load_local_invocation_index(struct ac_nir_context *ctx)
2795 {
2796 LLVMValueRef result;
2797 LLVMValueRef thread_id = ac_get_thread_id(&ctx->ac);
2798 result = LLVMBuildAnd(ctx->ac.builder, ctx->abi->tg_size,
2799 LLVMConstInt(ctx->ac.i32, 0xfc0, false), "");
2800
2801 return LLVMBuildAdd(ctx->ac.builder, result, thread_id, "");
2802 }
2803
2804 static LLVMValueRef
2805 visit_load_subgroup_id(struct ac_nir_context *ctx)
2806 {
2807 if (ctx->stage == MESA_SHADER_COMPUTE) {
2808 LLVMValueRef result;
2809 result = LLVMBuildAnd(ctx->ac.builder, ctx->abi->tg_size,
2810 LLVMConstInt(ctx->ac.i32, 0xfc0, false), "");
2811 return LLVMBuildLShr(ctx->ac.builder, result, LLVMConstInt(ctx->ac.i32, 6, false), "");
2812 } else {
2813 return LLVMConstInt(ctx->ac.i32, 0, false);
2814 }
2815 }
2816
2817 static LLVMValueRef
2818 visit_load_num_subgroups(struct ac_nir_context *ctx)
2819 {
2820 if (ctx->stage == MESA_SHADER_COMPUTE) {
2821 return LLVMBuildAnd(ctx->ac.builder, ctx->abi->tg_size,
2822 LLVMConstInt(ctx->ac.i32, 0x3f, false), "");
2823 } else {
2824 return LLVMConstInt(ctx->ac.i32, 1, false);
2825 }
2826 }
2827
2828 static LLVMValueRef
2829 visit_first_invocation(struct ac_nir_context *ctx)
2830 {
2831 LLVMValueRef active_set = ac_build_ballot(&ctx->ac, ctx->ac.i32_1);
2832
2833 /* The second argument is whether cttz(0) should be defined, but we do not care. */
2834 LLVMValueRef args[] = {active_set, ctx->ac.i1false};
2835 LLVMValueRef result = ac_build_intrinsic(&ctx->ac,
2836 "llvm.cttz.i64",
2837 ctx->ac.i64, args, 2,
2838 AC_FUNC_ATTR_NOUNWIND |
2839 AC_FUNC_ATTR_READNONE);
2840
2841 return LLVMBuildTrunc(ctx->ac.builder, result, ctx->ac.i32, "");
2842 }
2843
2844 static LLVMValueRef
2845 visit_load_shared(struct ac_nir_context *ctx,
2846 const nir_intrinsic_instr *instr)
2847 {
2848 LLVMValueRef values[4], derived_ptr, index, ret;
2849
2850 LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[0]);
2851
2852 for (int chan = 0; chan < instr->num_components; chan++) {
2853 index = LLVMConstInt(ctx->ac.i32, chan, 0);
2854 derived_ptr = LLVMBuildGEP(ctx->ac.builder, ptr, &index, 1, "");
2855 values[chan] = LLVMBuildLoad(ctx->ac.builder, derived_ptr, "");
2856 }
2857
2858 ret = ac_build_gather_values(&ctx->ac, values, instr->num_components);
2859 return LLVMBuildBitCast(ctx->ac.builder, ret, get_def_type(ctx, &instr->dest.ssa), "");
2860 }
2861
2862 static void
2863 visit_store_shared(struct ac_nir_context *ctx,
2864 const nir_intrinsic_instr *instr)
2865 {
2866 LLVMValueRef derived_ptr, data,index;
2867 LLVMBuilderRef builder = ctx->ac.builder;
2868
2869 LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[1]);
2870 LLVMValueRef src = get_src(ctx, instr->src[0]);
2871
2872 int writemask = nir_intrinsic_write_mask(instr);
2873 for (int chan = 0; chan < 4; chan++) {
2874 if (!(writemask & (1 << chan))) {
2875 continue;
2876 }
2877 data = ac_llvm_extract_elem(&ctx->ac, src, chan);
2878 index = LLVMConstInt(ctx->ac.i32, chan, 0);
2879 derived_ptr = LLVMBuildGEP(builder, ptr, &index, 1, "");
2880 LLVMBuildStore(builder, data, derived_ptr);
2881 }
2882 }
2883
2884 static LLVMValueRef visit_var_atomic(struct ac_nir_context *ctx,
2885 const nir_intrinsic_instr *instr,
2886 LLVMValueRef ptr, int src_idx)
2887 {
2888 LLVMValueRef result;
2889 LLVMValueRef src = get_src(ctx, instr->src[src_idx]);
2890
2891 if (instr->intrinsic == nir_intrinsic_shared_atomic_comp_swap ||
2892 instr->intrinsic == nir_intrinsic_deref_atomic_comp_swap) {
2893 LLVMValueRef src1 = get_src(ctx, instr->src[src_idx + 1]);
2894 result = LLVMBuildAtomicCmpXchg(ctx->ac.builder,
2895 ptr, src, src1,
2896 LLVMAtomicOrderingSequentiallyConsistent,
2897 LLVMAtomicOrderingSequentiallyConsistent,
2898 false);
2899 result = LLVMBuildExtractValue(ctx->ac.builder, result, 0, "");
2900 } else {
2901 LLVMAtomicRMWBinOp op;
2902 switch (instr->intrinsic) {
2903 case nir_intrinsic_shared_atomic_add:
2904 case nir_intrinsic_deref_atomic_add:
2905 op = LLVMAtomicRMWBinOpAdd;
2906 break;
2907 case nir_intrinsic_shared_atomic_umin:
2908 case nir_intrinsic_deref_atomic_umin:
2909 op = LLVMAtomicRMWBinOpUMin;
2910 break;
2911 case nir_intrinsic_shared_atomic_umax:
2912 case nir_intrinsic_deref_atomic_umax:
2913 op = LLVMAtomicRMWBinOpUMax;
2914 break;
2915 case nir_intrinsic_shared_atomic_imin:
2916 case nir_intrinsic_deref_atomic_imin:
2917 op = LLVMAtomicRMWBinOpMin;
2918 break;
2919 case nir_intrinsic_shared_atomic_imax:
2920 case nir_intrinsic_deref_atomic_imax:
2921 op = LLVMAtomicRMWBinOpMax;
2922 break;
2923 case nir_intrinsic_shared_atomic_and:
2924 case nir_intrinsic_deref_atomic_and:
2925 op = LLVMAtomicRMWBinOpAnd;
2926 break;
2927 case nir_intrinsic_shared_atomic_or:
2928 case nir_intrinsic_deref_atomic_or:
2929 op = LLVMAtomicRMWBinOpOr;
2930 break;
2931 case nir_intrinsic_shared_atomic_xor:
2932 case nir_intrinsic_deref_atomic_xor:
2933 op = LLVMAtomicRMWBinOpXor;
2934 break;
2935 case nir_intrinsic_shared_atomic_exchange:
2936 case nir_intrinsic_deref_atomic_exchange:
2937 op = LLVMAtomicRMWBinOpXchg;
2938 break;
2939 default:
2940 return NULL;
2941 }
2942
2943 result = LLVMBuildAtomicRMW(ctx->ac.builder, op, ptr, ac_to_integer(&ctx->ac, src),
2944 LLVMAtomicOrderingSequentiallyConsistent,
2945 false);
2946 }
2947 return result;
2948 }
2949
2950 static LLVMValueRef load_sample_pos(struct ac_nir_context *ctx)
2951 {
2952 LLVMValueRef values[2];
2953 LLVMValueRef pos[2];
2954
2955 pos[0] = ac_to_float(&ctx->ac, ctx->abi->frag_pos[0]);
2956 pos[1] = ac_to_float(&ctx->ac, ctx->abi->frag_pos[1]);
2957
2958 values[0] = ac_build_fract(&ctx->ac, pos[0], 32);
2959 values[1] = ac_build_fract(&ctx->ac, pos[1], 32);
2960 return ac_build_gather_values(&ctx->ac, values, 2);
2961 }
2962
2963 static LLVMValueRef visit_interp(struct ac_nir_context *ctx,
2964 const nir_intrinsic_instr *instr)
2965 {
2966 LLVMValueRef result[4];
2967 LLVMValueRef interp_param;
2968 unsigned location;
2969 unsigned chan;
2970 LLVMValueRef src_c0 = NULL;
2971 LLVMValueRef src_c1 = NULL;
2972 LLVMValueRef src0 = NULL;
2973
2974 nir_deref_instr *deref_instr = nir_instr_as_deref(instr->src[0].ssa->parent_instr);
2975 nir_variable *var = nir_deref_instr_get_variable(deref_instr);
2976 int input_base = ctx->abi->fs_input_attr_indices[var->data.location - VARYING_SLOT_VAR0];
2977 switch (instr->intrinsic) {
2978 case nir_intrinsic_interp_deref_at_centroid:
2979 location = INTERP_CENTROID;
2980 break;
2981 case nir_intrinsic_interp_deref_at_sample:
2982 case nir_intrinsic_interp_deref_at_offset:
2983 location = INTERP_CENTER;
2984 src0 = get_src(ctx, instr->src[1]);
2985 break;
2986 default:
2987 break;
2988 }
2989
2990 if (instr->intrinsic == nir_intrinsic_interp_deref_at_offset) {
2991 src_c0 = ac_to_float(&ctx->ac, LLVMBuildExtractElement(ctx->ac.builder, src0, ctx->ac.i32_0, ""));
2992 src_c1 = ac_to_float(&ctx->ac, LLVMBuildExtractElement(ctx->ac.builder, src0, ctx->ac.i32_1, ""));
2993 } else if (instr->intrinsic == nir_intrinsic_interp_deref_at_sample) {
2994 LLVMValueRef sample_position;
2995 LLVMValueRef halfval = LLVMConstReal(ctx->ac.f32, 0.5f);
2996
2997 /* fetch sample ID */
2998 sample_position = ctx->abi->load_sample_position(ctx->abi, src0);
2999
3000 src_c0 = LLVMBuildExtractElement(ctx->ac.builder, sample_position, ctx->ac.i32_0, "");
3001 src_c0 = LLVMBuildFSub(ctx->ac.builder, src_c0, halfval, "");
3002 src_c1 = LLVMBuildExtractElement(ctx->ac.builder, sample_position, ctx->ac.i32_1, "");
3003 src_c1 = LLVMBuildFSub(ctx->ac.builder, src_c1, halfval, "");
3004 }
3005 interp_param = ctx->abi->lookup_interp_param(ctx->abi, var->data.interpolation, location);
3006
3007 if (location == INTERP_CENTER) {
3008 LLVMValueRef ij_out[2];
3009 LLVMValueRef ddxy_out = emit_ddxy_interp(ctx, interp_param);
3010
3011 /*
3012 * take the I then J parameters, and the DDX/Y for it, and
3013 * calculate the IJ inputs for the interpolator.
3014 * temp1 = ddx * offset/sample.x + I;
3015 * interp_param.I = ddy * offset/sample.y + temp1;
3016 * temp1 = ddx * offset/sample.x + J;
3017 * interp_param.J = ddy * offset/sample.y + temp1;
3018 */
3019 for (unsigned i = 0; i < 2; i++) {
3020 LLVMValueRef ix_ll = LLVMConstInt(ctx->ac.i32, i, false);
3021 LLVMValueRef iy_ll = LLVMConstInt(ctx->ac.i32, i + 2, false);
3022 LLVMValueRef ddx_el = LLVMBuildExtractElement(ctx->ac.builder,
3023 ddxy_out, ix_ll, "");
3024 LLVMValueRef ddy_el = LLVMBuildExtractElement(ctx->ac.builder,
3025 ddxy_out, iy_ll, "");
3026 LLVMValueRef interp_el = LLVMBuildExtractElement(ctx->ac.builder,
3027 interp_param, ix_ll, "");
3028 LLVMValueRef temp1, temp2;
3029
3030 interp_el = LLVMBuildBitCast(ctx->ac.builder, interp_el,
3031 ctx->ac.f32, "");
3032
3033 temp1 = ac_build_fmad(&ctx->ac, ddx_el, src_c0, interp_el);
3034 temp2 = ac_build_fmad(&ctx->ac, ddy_el, src_c1, temp1);
3035
3036 ij_out[i] = LLVMBuildBitCast(ctx->ac.builder,
3037 temp2, ctx->ac.i32, "");
3038 }
3039 interp_param = ac_build_gather_values(&ctx->ac, ij_out, 2);
3040
3041 }
3042
3043 LLVMValueRef attrib_idx = ctx->ac.i32_0;
3044 while(deref_instr->deref_type != nir_deref_type_var) {
3045 if (deref_instr->deref_type == nir_deref_type_array) {
3046 unsigned array_size = glsl_count_attribute_slots(deref_instr->type, false);
3047
3048 LLVMValueRef offset;
3049 nir_const_value *const_value = nir_src_as_const_value(deref_instr->arr.index);
3050 if (const_value) {
3051 offset = LLVMConstInt(ctx->ac.i32, array_size * const_value->u32[0], false);
3052 } else {
3053 LLVMValueRef indirect = get_src(ctx, deref_instr->arr.index);
3054
3055 offset = LLVMBuildMul(ctx->ac.builder, indirect,
3056 LLVMConstInt(ctx->ac.i32, array_size, false), "");
3057 }
3058
3059 attrib_idx = LLVMBuildAdd(ctx->ac.builder, attrib_idx, offset, "");
3060 deref_instr = nir_src_as_deref(deref_instr->parent);
3061 } else if (deref_instr->deref_type == nir_deref_type_struct) {
3062 LLVMValueRef offset;
3063 unsigned sidx = deref_instr->strct.index;
3064 deref_instr = nir_src_as_deref(deref_instr->parent);
3065 offset = LLVMConstInt(ctx->ac.i32, glsl_get_struct_location_offset(deref_instr->type, sidx), false);
3066 attrib_idx = LLVMBuildAdd(ctx->ac.builder, attrib_idx, offset, "");
3067 } else {
3068 unreachable("Unsupported deref type");
3069 }
3070
3071 }
3072
3073 unsigned attrib_size = glsl_count_attribute_slots(var->type, false);
3074 for (chan = 0; chan < 4; chan++) {
3075 LLVMValueRef gather = LLVMGetUndef(LLVMVectorType(ctx->ac.f32, attrib_size));
3076 LLVMValueRef llvm_chan = LLVMConstInt(ctx->ac.i32, chan, false);
3077
3078 for (unsigned idx = 0; idx < attrib_size; ++idx) {
3079 LLVMValueRef v, attr_number;
3080
3081 attr_number = LLVMConstInt(ctx->ac.i32, input_base + idx, false);
3082 if (interp_param) {
3083 interp_param = LLVMBuildBitCast(ctx->ac.builder,
3084 interp_param, ctx->ac.v2f32, "");
3085 LLVMValueRef i = LLVMBuildExtractElement(
3086 ctx->ac.builder, interp_param, ctx->ac.i32_0, "");
3087 LLVMValueRef j = LLVMBuildExtractElement(
3088 ctx->ac.builder, interp_param, ctx->ac.i32_1, "");
3089
3090 v = ac_build_fs_interp(&ctx->ac, llvm_chan, attr_number,
3091 ctx->abi->prim_mask, i, j);
3092 } else {
3093 v = ac_build_fs_interp_mov(&ctx->ac, LLVMConstInt(ctx->ac.i32, 2, false),
3094 llvm_chan, attr_number, ctx->abi->prim_mask);
3095 }
3096
3097 gather = LLVMBuildInsertElement(ctx->ac.builder, gather, v,
3098 LLVMConstInt(ctx->ac.i32, idx, false), "");
3099 }
3100
3101 result[chan] = LLVMBuildExtractElement(ctx->ac.builder, gather, attrib_idx, "");
3102
3103 }
3104 return ac_build_varying_gather_values(&ctx->ac, result, instr->num_components,
3105 var->data.location_frac);
3106 }
3107
3108 static void visit_intrinsic(struct ac_nir_context *ctx,
3109 nir_intrinsic_instr *instr)
3110 {
3111 LLVMValueRef result = NULL;
3112
3113 switch (instr->intrinsic) {
3114 case nir_intrinsic_ballot:
3115 result = ac_build_ballot(&ctx->ac, get_src(ctx, instr->src[0]));
3116 break;
3117 case nir_intrinsic_read_invocation:
3118 result = ac_build_readlane(&ctx->ac, get_src(ctx, instr->src[0]),
3119 get_src(ctx, instr->src[1]));
3120 break;
3121 case nir_intrinsic_read_first_invocation:
3122 result = ac_build_readlane(&ctx->ac, get_src(ctx, instr->src[0]), NULL);
3123 break;
3124 case nir_intrinsic_load_subgroup_invocation:
3125 result = ac_get_thread_id(&ctx->ac);
3126 break;
3127 case nir_intrinsic_load_work_group_id: {
3128 LLVMValueRef values[3];
3129
3130 for (int i = 0; i < 3; i++) {
3131 values[i] = ctx->abi->workgroup_ids[i] ?
3132 ctx->abi->workgroup_ids[i] : ctx->ac.i32_0;
3133 }
3134
3135 result = ac_build_gather_values(&ctx->ac, values, 3);
3136 break;
3137 }
3138 case nir_intrinsic_load_base_vertex:
3139 case nir_intrinsic_load_first_vertex:
3140 result = ctx->abi->load_base_vertex(ctx->abi);
3141 break;
3142 case nir_intrinsic_load_local_group_size:
3143 result = ctx->abi->load_local_group_size(ctx->abi);
3144 break;
3145 case nir_intrinsic_load_vertex_id:
3146 result = LLVMBuildAdd(ctx->ac.builder, ctx->abi->vertex_id,
3147 ctx->abi->base_vertex, "");
3148 break;
3149 case nir_intrinsic_load_vertex_id_zero_base: {
3150 result = ctx->abi->vertex_id;
3151 break;
3152 }
3153 case nir_intrinsic_load_local_invocation_id: {
3154 result = ctx->abi->local_invocation_ids;
3155 break;
3156 }
3157 case nir_intrinsic_load_base_instance:
3158 result = ctx->abi->start_instance;
3159 break;
3160 case nir_intrinsic_load_draw_id:
3161 result = ctx->abi->draw_id;
3162 break;
3163 case nir_intrinsic_load_view_index:
3164 result = ctx->abi->view_index;
3165 break;
3166 case nir_intrinsic_load_invocation_id:
3167 if (ctx->stage == MESA_SHADER_TESS_CTRL)
3168 result = ac_unpack_param(&ctx->ac, ctx->abi->tcs_rel_ids, 8, 5);
3169 else
3170 result = ctx->abi->gs_invocation_id;
3171 break;
3172 case nir_intrinsic_load_primitive_id:
3173 if (ctx->stage == MESA_SHADER_GEOMETRY) {
3174 result = ctx->abi->gs_prim_id;
3175 } else if (ctx->stage == MESA_SHADER_TESS_CTRL) {
3176 result = ctx->abi->tcs_patch_id;
3177 } else if (ctx->stage == MESA_SHADER_TESS_EVAL) {
3178 result = ctx->abi->tes_patch_id;
3179 } else
3180 fprintf(stderr, "Unknown primitive id intrinsic: %d", ctx->stage);
3181 break;
3182 case nir_intrinsic_load_sample_id:
3183 result = ac_unpack_param(&ctx->ac, ctx->abi->ancillary, 8, 4);
3184 break;
3185 case nir_intrinsic_load_sample_pos:
3186 result = load_sample_pos(ctx);
3187 break;
3188 case nir_intrinsic_load_sample_mask_in:
3189 result = ctx->abi->load_sample_mask_in(ctx->abi);
3190 break;
3191 case nir_intrinsic_load_frag_coord: {
3192 LLVMValueRef values[4] = {
3193 ctx->abi->frag_pos[0],
3194 ctx->abi->frag_pos[1],
3195 ctx->abi->frag_pos[2],
3196 ac_build_fdiv(&ctx->ac, ctx->ac.f32_1, ctx->abi->frag_pos[3])
3197 };
3198 result = ac_build_gather_values(&ctx->ac, values, 4);
3199 break;
3200 }
3201 case nir_intrinsic_load_front_face:
3202 result = ctx->abi->front_face;
3203 break;
3204 case nir_intrinsic_load_helper_invocation:
3205 result = visit_load_helper_invocation(ctx);
3206 break;
3207 case nir_intrinsic_load_instance_id:
3208 result = ctx->abi->instance_id;
3209 break;
3210 case nir_intrinsic_load_num_work_groups:
3211 result = ctx->abi->num_work_groups;
3212 break;
3213 case nir_intrinsic_load_local_invocation_index:
3214 result = visit_load_local_invocation_index(ctx);
3215 break;
3216 case nir_intrinsic_load_subgroup_id:
3217 result = visit_load_subgroup_id(ctx);
3218 break;
3219 case nir_intrinsic_load_num_subgroups:
3220 result = visit_load_num_subgroups(ctx);
3221 break;
3222 case nir_intrinsic_first_invocation:
3223 result = visit_first_invocation(ctx);
3224 break;
3225 case nir_intrinsic_load_push_constant:
3226 result = visit_load_push_constant(ctx, instr);
3227 break;
3228 case nir_intrinsic_vulkan_resource_index: {
3229 LLVMValueRef index = get_src(ctx, instr->src[0]);
3230 unsigned desc_set = nir_intrinsic_desc_set(instr);
3231 unsigned binding = nir_intrinsic_binding(instr);
3232
3233 result = ctx->abi->load_resource(ctx->abi, index, desc_set,
3234 binding);
3235 break;
3236 }
3237 case nir_intrinsic_vulkan_resource_reindex:
3238 result = visit_vulkan_resource_reindex(ctx, instr);
3239 break;
3240 case nir_intrinsic_store_ssbo:
3241 visit_store_ssbo(ctx, instr);
3242 break;
3243 case nir_intrinsic_load_ssbo:
3244 result = visit_load_buffer(ctx, instr);
3245 break;
3246 case nir_intrinsic_ssbo_atomic_add:
3247 case nir_intrinsic_ssbo_atomic_imin:
3248 case nir_intrinsic_ssbo_atomic_umin:
3249 case nir_intrinsic_ssbo_atomic_imax:
3250 case nir_intrinsic_ssbo_atomic_umax:
3251 case nir_intrinsic_ssbo_atomic_and:
3252 case nir_intrinsic_ssbo_atomic_or:
3253 case nir_intrinsic_ssbo_atomic_xor:
3254 case nir_intrinsic_ssbo_atomic_exchange:
3255 case nir_intrinsic_ssbo_atomic_comp_swap:
3256 result = visit_atomic_ssbo(ctx, instr);
3257 break;
3258 case nir_intrinsic_load_ubo:
3259 result = visit_load_ubo_buffer(ctx, instr);
3260 break;
3261 case nir_intrinsic_get_buffer_size:
3262 result = visit_get_buffer_size(ctx, instr);
3263 break;
3264 case nir_intrinsic_load_deref:
3265 result = visit_load_var(ctx, instr);
3266 break;
3267 case nir_intrinsic_store_deref:
3268 visit_store_var(ctx, instr);
3269 break;
3270 case nir_intrinsic_load_shared:
3271 result = visit_load_shared(ctx, instr);
3272 break;
3273 case nir_intrinsic_store_shared:
3274 visit_store_shared(ctx, instr);
3275 break;
3276 case nir_intrinsic_image_deref_samples:
3277 result = visit_image_samples(ctx, instr);
3278 break;
3279 case nir_intrinsic_image_deref_load:
3280 result = visit_image_load(ctx, instr);
3281 break;
3282 case nir_intrinsic_image_deref_store:
3283 visit_image_store(ctx, instr);
3284 break;
3285 case nir_intrinsic_image_deref_atomic_add:
3286 case nir_intrinsic_image_deref_atomic_min:
3287 case nir_intrinsic_image_deref_atomic_max:
3288 case nir_intrinsic_image_deref_atomic_and:
3289 case nir_intrinsic_image_deref_atomic_or:
3290 case nir_intrinsic_image_deref_atomic_xor:
3291 case nir_intrinsic_image_deref_atomic_exchange:
3292 case nir_intrinsic_image_deref_atomic_comp_swap:
3293 result = visit_image_atomic(ctx, instr);
3294 break;
3295 case nir_intrinsic_image_deref_size:
3296 result = visit_image_size(ctx, instr);
3297 break;
3298 case nir_intrinsic_shader_clock:
3299 result = ac_build_shader_clock(&ctx->ac);
3300 break;
3301 case nir_intrinsic_discard:
3302 case nir_intrinsic_discard_if:
3303 emit_discard(ctx, instr);
3304 break;
3305 case nir_intrinsic_memory_barrier:
3306 case nir_intrinsic_group_memory_barrier:
3307 case nir_intrinsic_memory_barrier_atomic_counter:
3308 case nir_intrinsic_memory_barrier_buffer:
3309 case nir_intrinsic_memory_barrier_image:
3310 case nir_intrinsic_memory_barrier_shared:
3311 emit_membar(&ctx->ac, instr);
3312 break;
3313 case nir_intrinsic_barrier:
3314 ac_emit_barrier(&ctx->ac, ctx->stage);
3315 break;
3316 case nir_intrinsic_shared_atomic_add:
3317 case nir_intrinsic_shared_atomic_imin:
3318 case nir_intrinsic_shared_atomic_umin:
3319 case nir_intrinsic_shared_atomic_imax:
3320 case nir_intrinsic_shared_atomic_umax:
3321 case nir_intrinsic_shared_atomic_and:
3322 case nir_intrinsic_shared_atomic_or:
3323 case nir_intrinsic_shared_atomic_xor:
3324 case nir_intrinsic_shared_atomic_exchange:
3325 case nir_intrinsic_shared_atomic_comp_swap: {
3326 LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[0]);
3327 result = visit_var_atomic(ctx, instr, ptr, 1);
3328 break;
3329 }
3330 case nir_intrinsic_deref_atomic_add:
3331 case nir_intrinsic_deref_atomic_imin:
3332 case nir_intrinsic_deref_atomic_umin:
3333 case nir_intrinsic_deref_atomic_imax:
3334 case nir_intrinsic_deref_atomic_umax:
3335 case nir_intrinsic_deref_atomic_and:
3336 case nir_intrinsic_deref_atomic_or:
3337 case nir_intrinsic_deref_atomic_xor:
3338 case nir_intrinsic_deref_atomic_exchange:
3339 case nir_intrinsic_deref_atomic_comp_swap: {
3340 LLVMValueRef ptr = get_src(ctx, instr->src[0]);
3341 result = visit_var_atomic(ctx, instr, ptr, 1);
3342 break;
3343 }
3344 case nir_intrinsic_interp_deref_at_centroid:
3345 case nir_intrinsic_interp_deref_at_sample:
3346 case nir_intrinsic_interp_deref_at_offset:
3347 result = visit_interp(ctx, instr);
3348 break;
3349 case nir_intrinsic_emit_vertex:
3350 ctx->abi->emit_vertex(ctx->abi, nir_intrinsic_stream_id(instr), ctx->abi->outputs);
3351 break;
3352 case nir_intrinsic_end_primitive:
3353 ctx->abi->emit_primitive(ctx->abi, nir_intrinsic_stream_id(instr));
3354 break;
3355 case nir_intrinsic_load_tess_coord:
3356 result = ctx->abi->load_tess_coord(ctx->abi);
3357 break;
3358 case nir_intrinsic_load_tess_level_outer:
3359 result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_OUTER);
3360 break;
3361 case nir_intrinsic_load_tess_level_inner:
3362 result = ctx->abi->load_tess_level(ctx->abi, VARYING_SLOT_TESS_LEVEL_INNER);
3363 break;
3364 case nir_intrinsic_load_patch_vertices_in:
3365 result = ctx->abi->load_patch_vertices_in(ctx->abi);
3366 break;
3367 case nir_intrinsic_vote_all: {
3368 LLVMValueRef tmp = ac_build_vote_all(&ctx->ac, get_src(ctx, instr->src[0]));
3369 result = LLVMBuildSExt(ctx->ac.builder, tmp, ctx->ac.i32, "");
3370 break;
3371 }
3372 case nir_intrinsic_vote_any: {
3373 LLVMValueRef tmp = ac_build_vote_any(&ctx->ac, get_src(ctx, instr->src[0]));
3374 result = LLVMBuildSExt(ctx->ac.builder, tmp, ctx->ac.i32, "");
3375 break;
3376 }
3377 case nir_intrinsic_shuffle:
3378 result = ac_build_shuffle(&ctx->ac, get_src(ctx, instr->src[0]),
3379 get_src(ctx, instr->src[1]));
3380 break;
3381 case nir_intrinsic_reduce:
3382 result = ac_build_reduce(&ctx->ac,
3383 get_src(ctx, instr->src[0]),
3384 instr->const_index[0],
3385 instr->const_index[1]);
3386 break;
3387 case nir_intrinsic_inclusive_scan:
3388 result = ac_build_inclusive_scan(&ctx->ac,
3389 get_src(ctx, instr->src[0]),
3390 instr->const_index[0]);
3391 break;
3392 case nir_intrinsic_exclusive_scan:
3393 result = ac_build_exclusive_scan(&ctx->ac,
3394 get_src(ctx, instr->src[0]),
3395 instr->const_index[0]);
3396 break;
3397 case nir_intrinsic_quad_broadcast: {
3398 unsigned lane = nir_src_as_const_value(instr->src[1])->u32[0];
3399 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]),
3400 lane, lane, lane, lane);
3401 break;
3402 }
3403 case nir_intrinsic_quad_swap_horizontal:
3404 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 1, 0, 3 ,2);
3405 break;
3406 case nir_intrinsic_quad_swap_vertical:
3407 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 2, 3, 0 ,1);
3408 break;
3409 case nir_intrinsic_quad_swap_diagonal:
3410 result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 3, 2, 1 ,0);
3411 break;
3412 default:
3413 fprintf(stderr, "Unknown intrinsic: ");
3414 nir_print_instr(&instr->instr, stderr);
3415 fprintf(stderr, "\n");
3416 break;
3417 }
3418 if (result) {
3419 ctx->ssa_defs[instr->dest.ssa.index] = result;
3420 }
3421 }
3422
3423 static LLVMValueRef get_bindless_index_from_uniform(struct ac_nir_context *ctx,
3424 unsigned base_index,
3425 unsigned constant_index,
3426 LLVMValueRef dynamic_index)
3427 {
3428 LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, base_index * 4, 0);
3429 LLVMValueRef index = LLVMBuildAdd(ctx->ac.builder, dynamic_index,
3430 LLVMConstInt(ctx->ac.i32, constant_index, 0), "");
3431
3432 /* Bindless uniforms are 64bit so multiple index by 8 */
3433 index = LLVMBuildMul(ctx->ac.builder, index, LLVMConstInt(ctx->ac.i32, 8, 0), "");
3434 offset = LLVMBuildAdd(ctx->ac.builder, offset, index, "");
3435
3436 LLVMValueRef ubo_index = ctx->abi->load_ubo(ctx->abi, ctx->ac.i32_0);
3437
3438 LLVMValueRef ret = ac_build_buffer_load(&ctx->ac, ubo_index, 1, NULL, offset,
3439 NULL, 0, false, false, true, true);
3440
3441 return LLVMBuildBitCast(ctx->ac.builder, ret, ctx->ac.i32, "");
3442 }
3443
3444 static LLVMValueRef get_sampler_desc(struct ac_nir_context *ctx,
3445 nir_deref_instr *deref_instr,
3446 enum ac_descriptor_type desc_type,
3447 const nir_tex_instr *tex_instr,
3448 bool image, bool write)
3449 {
3450 LLVMValueRef index = NULL;
3451 unsigned constant_index = 0;
3452 unsigned descriptor_set;
3453 unsigned base_index;
3454 bool bindless = false;
3455
3456 if (!deref_instr) {
3457 assert(tex_instr && !image);
3458 descriptor_set = 0;
3459 base_index = tex_instr->sampler_index;
3460 } else {
3461 while(deref_instr->deref_type != nir_deref_type_var) {
3462 if (deref_instr->deref_type == nir_deref_type_array) {
3463 unsigned array_size = glsl_get_aoa_size(deref_instr->type);
3464 if (!array_size)
3465 array_size = 1;
3466
3467 nir_const_value *const_value = nir_src_as_const_value(deref_instr->arr.index);
3468 if (const_value) {
3469 constant_index += array_size * const_value->u32[0];
3470 } else {
3471 LLVMValueRef indirect = get_src(ctx, deref_instr->arr.index);
3472
3473 indirect = LLVMBuildMul(ctx->ac.builder, indirect,
3474 LLVMConstInt(ctx->ac.i32, array_size, false), "");
3475
3476 if (!index)
3477 index = indirect;
3478 else
3479 index = LLVMBuildAdd(ctx->ac.builder, index, indirect, "");
3480 }
3481
3482 deref_instr = nir_src_as_deref(deref_instr->parent);
3483 } else if (deref_instr->deref_type == nir_deref_type_struct) {
3484 unsigned sidx = deref_instr->strct.index;
3485 deref_instr = nir_src_as_deref(deref_instr->parent);
3486 constant_index += glsl_get_struct_location_offset(deref_instr->type, sidx);
3487 } else {
3488 unreachable("Unsupported deref type");
3489 }
3490 }
3491 descriptor_set = deref_instr->var->data.descriptor_set;
3492
3493 if (deref_instr->var->data.bindless) {
3494 /* For now just assert on unhandled variable types */
3495 assert(deref_instr->var->data.mode == nir_var_uniform);
3496
3497 base_index = deref_instr->var->data.driver_location;
3498 bindless = true;
3499
3500 index = index ? index : ctx->ac.i32_0;
3501 index = get_bindless_index_from_uniform(ctx, base_index,
3502 constant_index, index);
3503 } else
3504 base_index = deref_instr->var->data.binding;
3505 }
3506
3507 return ctx->abi->load_sampler_desc(ctx->abi,
3508 descriptor_set,
3509 base_index,
3510 constant_index, index,
3511 desc_type, image, write, bindless);
3512 }
3513
3514 /* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
3515 *
3516 * SI-CI:
3517 * If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
3518 * filtering manually. The driver sets img7 to a mask clearing
3519 * MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
3520 * s_and_b32 samp0, samp0, img7
3521 *
3522 * VI:
3523 * The ANISO_OVERRIDE sampler field enables this fix in TA.
3524 */
3525 static LLVMValueRef sici_fix_sampler_aniso(struct ac_nir_context *ctx,
3526 LLVMValueRef res, LLVMValueRef samp)
3527 {
3528 LLVMBuilderRef builder = ctx->ac.builder;
3529 LLVMValueRef img7, samp0;
3530
3531 if (ctx->ac.chip_class >= VI)
3532 return samp;
3533
3534 img7 = LLVMBuildExtractElement(builder, res,
3535 LLVMConstInt(ctx->ac.i32, 7, 0), "");
3536 samp0 = LLVMBuildExtractElement(builder, samp,
3537 LLVMConstInt(ctx->ac.i32, 0, 0), "");
3538 samp0 = LLVMBuildAnd(builder, samp0, img7, "");
3539 return LLVMBuildInsertElement(builder, samp, samp0,
3540 LLVMConstInt(ctx->ac.i32, 0, 0), "");
3541 }
3542
3543 static void tex_fetch_ptrs(struct ac_nir_context *ctx,
3544 nir_tex_instr *instr,
3545 LLVMValueRef *res_ptr, LLVMValueRef *samp_ptr,
3546 LLVMValueRef *fmask_ptr)
3547 {
3548 nir_deref_instr *texture_deref_instr = NULL;
3549 nir_deref_instr *sampler_deref_instr = NULL;
3550
3551 for (unsigned i = 0; i < instr->num_srcs; i++) {
3552 switch (instr->src[i].src_type) {
3553 case nir_tex_src_texture_deref:
3554 texture_deref_instr = nir_src_as_deref(instr->src[i].src);
3555 break;
3556 case nir_tex_src_sampler_deref:
3557 sampler_deref_instr = nir_src_as_deref(instr->src[i].src);
3558 break;
3559 default:
3560 break;
3561 }
3562 }
3563
3564 if (!sampler_deref_instr)
3565 sampler_deref_instr = texture_deref_instr;
3566
3567 if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF)
3568 *res_ptr = get_sampler_desc(ctx, texture_deref_instr, AC_DESC_BUFFER, instr, false, false);
3569 else
3570 *res_ptr = get_sampler_desc(ctx, texture_deref_instr, AC_DESC_IMAGE, instr, false, false);
3571 if (samp_ptr) {
3572 *samp_ptr = get_sampler_desc(ctx, sampler_deref_instr, AC_DESC_SAMPLER, instr, false, false);
3573 if (instr->sampler_dim < GLSL_SAMPLER_DIM_RECT)
3574 *samp_ptr = sici_fix_sampler_aniso(ctx, *res_ptr, *samp_ptr);
3575 }
3576 if (fmask_ptr && (instr->op == nir_texop_txf_ms ||
3577 instr->op == nir_texop_samples_identical))
3578 *fmask_ptr = get_sampler_desc(ctx, texture_deref_instr, AC_DESC_FMASK, instr, false, false);
3579 }
3580
3581 static LLVMValueRef apply_round_slice(struct ac_llvm_context *ctx,
3582 LLVMValueRef coord)
3583 {
3584 coord = ac_to_float(ctx, coord);
3585 coord = ac_build_round(ctx, coord);
3586 coord = ac_to_integer(ctx, coord);
3587 return coord;
3588 }
3589
3590 static void visit_tex(struct ac_nir_context *ctx, nir_tex_instr *instr)
3591 {
3592 LLVMValueRef result = NULL;
3593 struct ac_image_args args = { 0 };
3594 LLVMValueRef fmask_ptr = NULL, sample_index = NULL;
3595 LLVMValueRef ddx = NULL, ddy = NULL;
3596 unsigned offset_src = 0;
3597
3598 tex_fetch_ptrs(ctx, instr, &args.resource, &args.sampler, &fmask_ptr);
3599
3600 for (unsigned i = 0; i < instr->num_srcs; i++) {
3601 switch (instr->src[i].src_type) {
3602 case nir_tex_src_coord: {
3603 LLVMValueRef coord = get_src(ctx, instr->src[i].src);
3604 for (unsigned chan = 0; chan < instr->coord_components; ++chan)
3605 args.coords[chan] = ac_llvm_extract_elem(&ctx->ac, coord, chan);
3606 break;
3607 }
3608 case nir_tex_src_projector:
3609 break;
3610 case nir_tex_src_comparator:
3611 if (instr->is_shadow)
3612 args.compare = get_src(ctx, instr->src[i].src);
3613 break;
3614 case nir_tex_src_offset:
3615 args.offset = get_src(ctx, instr->src[i].src);
3616 offset_src = i;
3617 break;
3618 case nir_tex_src_bias:
3619 if (instr->op == nir_texop_txb)
3620 args.bias = get_src(ctx, instr->src[i].src);
3621 break;
3622 case nir_tex_src_lod: {
3623 nir_const_value *val = nir_src_as_const_value(instr->src[i].src);
3624
3625 if (val && val->i32[0] == 0)
3626 args.level_zero = true;
3627 else
3628 args.lod = get_src(ctx, instr->src[i].src);
3629 break;
3630 }
3631 case nir_tex_src_ms_index:
3632 sample_index = get_src(ctx, instr->src[i].src);
3633 break;
3634 case nir_tex_src_ms_mcs:
3635 break;
3636 case nir_tex_src_ddx:
3637 ddx = get_src(ctx, instr->src[i].src);
3638 break;
3639 case nir_tex_src_ddy:
3640 ddy = get_src(ctx, instr->src[i].src);
3641 break;
3642 case nir_tex_src_texture_offset:
3643 case nir_tex_src_sampler_offset:
3644 case nir_tex_src_plane:
3645 default:
3646 break;
3647 }
3648 }
3649
3650 if (instr->op == nir_texop_txs && instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
3651 result = get_buffer_size(ctx, args.resource, true);
3652 goto write_result;
3653 }
3654
3655 if (instr->op == nir_texop_texture_samples) {
3656 LLVMValueRef res, samples, is_msaa;
3657 res = LLVMBuildBitCast(ctx->ac.builder, args.resource, ctx->ac.v8i32, "");
3658 samples = LLVMBuildExtractElement(ctx->ac.builder, res,
3659 LLVMConstInt(ctx->ac.i32, 3, false), "");
3660 is_msaa = LLVMBuildLShr(ctx->ac.builder, samples,
3661 LLVMConstInt(ctx->ac.i32, 28, false), "");
3662 is_msaa = LLVMBuildAnd(ctx->ac.builder, is_msaa,
3663 LLVMConstInt(ctx->ac.i32, 0xe, false), "");
3664 is_msaa = LLVMBuildICmp(ctx->ac.builder, LLVMIntEQ, is_msaa,
3665 LLVMConstInt(ctx->ac.i32, 0xe, false), "");
3666
3667 samples = LLVMBuildLShr(ctx->ac.builder, samples,
3668 LLVMConstInt(ctx->ac.i32, 16, false), "");
3669 samples = LLVMBuildAnd(ctx->ac.builder, samples,
3670 LLVMConstInt(ctx->ac.i32, 0xf, false), "");
3671 samples = LLVMBuildShl(ctx->ac.builder, ctx->ac.i32_1,
3672 samples, "");
3673 samples = LLVMBuildSelect(ctx->ac.builder, is_msaa, samples,
3674 ctx->ac.i32_1, "");
3675 result = samples;
3676 goto write_result;
3677 }
3678
3679 if (args.offset && instr->op != nir_texop_txf) {
3680 LLVMValueRef offset[3], pack;
3681 for (unsigned chan = 0; chan < 3; ++chan)
3682 offset[chan] = ctx->ac.i32_0;
3683
3684 unsigned num_components = ac_get_llvm_num_components(args.offset);
3685 for (unsigned chan = 0; chan < num_components; chan++) {
3686 offset[chan] = ac_llvm_extract_elem(&ctx->ac, args.offset, chan);
3687 offset[chan] = LLVMBuildAnd(ctx->ac.builder, offset[chan],
3688 LLVMConstInt(ctx->ac.i32, 0x3f, false), "");
3689 if (chan)
3690 offset[chan] = LLVMBuildShl(ctx->ac.builder, offset[chan],
3691 LLVMConstInt(ctx->ac.i32, chan * 8, false), "");
3692 }
3693 pack = LLVMBuildOr(ctx->ac.builder, offset[0], offset[1], "");
3694 pack = LLVMBuildOr(ctx->ac.builder, pack, offset[2], "");
3695 args.offset = pack;
3696 }
3697
3698 /* TC-compatible HTILE on radeonsi promotes Z16 and Z24 to Z32_FLOAT,
3699 * so the depth comparison value isn't clamped for Z16 and
3700 * Z24 anymore. Do it manually here.
3701 *
3702 * It's unnecessary if the original texture format was
3703 * Z32_FLOAT, but we don't know that here.
3704 */
3705 if (args.compare && ctx->ac.chip_class >= VI && ctx->abi->clamp_shadow_reference)
3706 args.compare = ac_build_clamp(&ctx->ac, ac_to_float(&ctx->ac, args.compare));
3707
3708 /* pack derivatives */
3709 if (ddx || ddy) {
3710 int num_src_deriv_channels, num_dest_deriv_channels;
3711 switch (instr->sampler_dim) {
3712 case GLSL_SAMPLER_DIM_3D:
3713 case GLSL_SAMPLER_DIM_CUBE:
3714 num_src_deriv_channels = 3;
3715 num_dest_deriv_channels = 3;
3716 break;
3717 case GLSL_SAMPLER_DIM_2D:
3718 default:
3719 num_src_deriv_channels = 2;
3720 num_dest_deriv_channels = 2;
3721 break;
3722 case GLSL_SAMPLER_DIM_1D:
3723 num_src_deriv_channels = 1;
3724 if (ctx->ac.chip_class >= GFX9) {
3725 num_dest_deriv_channels = 2;
3726 } else {
3727 num_dest_deriv_channels = 1;
3728 }
3729 break;
3730 }
3731
3732 for (unsigned i = 0; i < num_src_deriv_channels; i++) {
3733 args.derivs[i] = ac_to_float(&ctx->ac,
3734 ac_llvm_extract_elem(&ctx->ac, ddx, i));
3735 args.derivs[num_dest_deriv_channels + i] = ac_to_float(&ctx->ac,
3736 ac_llvm_extract_elem(&ctx->ac, ddy, i));
3737 }
3738 for (unsigned i = num_src_deriv_channels; i < num_dest_deriv_channels; i++) {
3739 args.derivs[i] = ctx->ac.f32_0;
3740 args.derivs[num_dest_deriv_channels + i] = ctx->ac.f32_0;
3741 }
3742 }
3743
3744 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE && args.coords[0]) {
3745 for (unsigned chan = 0; chan < instr->coord_components; chan++)
3746 args.coords[chan] = ac_to_float(&ctx->ac, args.coords[chan]);
3747 if (instr->coord_components == 3)
3748 args.coords[3] = LLVMGetUndef(ctx->ac.f32);
3749 ac_prepare_cube_coords(&ctx->ac,
3750 instr->op == nir_texop_txd, instr->is_array,
3751 instr->op == nir_texop_lod, args.coords, args.derivs);
3752 }
3753
3754 /* Texture coordinates fixups */
3755 if (instr->coord_components > 1 &&
3756 instr->sampler_dim == GLSL_SAMPLER_DIM_1D &&
3757 instr->is_array &&
3758 instr->op != nir_texop_txf) {
3759 args.coords[1] = apply_round_slice(&ctx->ac, args.coords[1]);
3760 }
3761
3762 if (instr->coord_components > 2 &&
3763 (instr->sampler_dim == GLSL_SAMPLER_DIM_2D ||
3764 instr->sampler_dim == GLSL_SAMPLER_DIM_MS ||
3765 instr->sampler_dim == GLSL_SAMPLER_DIM_SUBPASS ||
3766 instr->sampler_dim == GLSL_SAMPLER_DIM_SUBPASS_MS) &&
3767 instr->is_array &&
3768 instr->op != nir_texop_txf && instr->op != nir_texop_txf_ms) {
3769 args.coords[2] = apply_round_slice(&ctx->ac, args.coords[2]);
3770 }
3771
3772 if (ctx->ac.chip_class >= GFX9 &&
3773 instr->sampler_dim == GLSL_SAMPLER_DIM_1D &&
3774 instr->op != nir_texop_lod) {
3775 LLVMValueRef filler;
3776 if (instr->op == nir_texop_txf)
3777 filler = ctx->ac.i32_0;
3778 else
3779 filler = LLVMConstReal(ctx->ac.f32, 0.5);
3780
3781 if (instr->is_array)
3782 args.coords[2] = args.coords[1];
3783 args.coords[1] = filler;
3784 }
3785
3786 /* Pack sample index */
3787 if (instr->op == nir_texop_txf_ms && sample_index)
3788 args.coords[instr->coord_components] = sample_index;
3789
3790 if (instr->op == nir_texop_samples_identical) {
3791 struct ac_image_args txf_args = { 0 };
3792 memcpy(txf_args.coords, args.coords, sizeof(txf_args.coords));
3793
3794 txf_args.dmask = 0xf;
3795 txf_args.resource = fmask_ptr;
3796 txf_args.dim = instr->is_array ? ac_image_2darray : ac_image_2d;
3797 result = build_tex_intrinsic(ctx, instr, &txf_args);
3798
3799 result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
3800 result = emit_int_cmp(&ctx->ac, LLVMIntEQ, result, ctx->ac.i32_0);
3801 goto write_result;
3802 }
3803
3804 if (instr->sampler_dim == GLSL_SAMPLER_DIM_MS &&
3805 instr->op != nir_texop_txs) {
3806 unsigned sample_chan = instr->is_array ? 3 : 2;
3807 args.coords[sample_chan] = adjust_sample_index_using_fmask(
3808 &ctx->ac, args.coords[0], args.coords[1],
3809 instr->is_array ? args.coords[2] : NULL,
3810 args.coords[sample_chan], fmask_ptr);
3811 }
3812
3813 if (args.offset && instr->op == nir_texop_txf) {
3814 nir_const_value *const_offset =
3815 nir_src_as_const_value(instr->src[offset_src].src);
3816 int num_offsets = instr->src[offset_src].src.ssa->num_components;
3817 assert(const_offset);
3818 num_offsets = MIN2(num_offsets, instr->coord_components);
3819 for (unsigned i = 0; i < num_offsets; ++i) {
3820 args.coords[i] = LLVMBuildAdd(
3821 ctx->ac.builder, args.coords[i],
3822 LLVMConstInt(ctx->ac.i32, const_offset->i32[i], false), "");
3823 }
3824 args.offset = NULL;
3825 }
3826
3827 /* TODO TG4 support */
3828 args.dmask = 0xf;
3829 if (instr->op == nir_texop_tg4) {
3830 if (instr->is_shadow)
3831 args.dmask = 1;
3832 else
3833 args.dmask = 1 << instr->component;
3834 }
3835
3836 if (instr->sampler_dim != GLSL_SAMPLER_DIM_BUF)
3837 args.dim = get_ac_sampler_dim(&ctx->ac, instr->sampler_dim, instr->is_array);
3838 result = build_tex_intrinsic(ctx, instr, &args);
3839
3840 if (instr->op == nir_texop_query_levels)
3841 result = LLVMBuildExtractElement(ctx->ac.builder, result, LLVMConstInt(ctx->ac.i32, 3, false), "");
3842 else if (instr->is_shadow && instr->is_new_style_shadow &&
3843 instr->op != nir_texop_txs && instr->op != nir_texop_lod &&
3844 instr->op != nir_texop_tg4)
3845 result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
3846 else if (instr->op == nir_texop_txs &&
3847 instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE &&
3848 instr->is_array) {
3849 LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
3850 LLVMValueRef six = LLVMConstInt(ctx->ac.i32, 6, false);
3851 LLVMValueRef z = LLVMBuildExtractElement(ctx->ac.builder, result, two, "");
3852 z = LLVMBuildSDiv(ctx->ac.builder, z, six, "");
3853 result = LLVMBuildInsertElement(ctx->ac.builder, result, z, two, "");
3854 } else if (ctx->ac.chip_class >= GFX9 &&
3855 instr->op == nir_texop_txs &&
3856 instr->sampler_dim == GLSL_SAMPLER_DIM_1D &&
3857 instr->is_array) {
3858 LLVMValueRef two = LLVMConstInt(ctx->ac.i32, 2, false);
3859 LLVMValueRef layers = LLVMBuildExtractElement(ctx->ac.builder, result, two, "");
3860 result = LLVMBuildInsertElement(ctx->ac.builder, result, layers,
3861 ctx->ac.i32_1, "");
3862 } else if (instr->dest.ssa.num_components != 4)
3863 result = ac_trim_vector(&ctx->ac, result, instr->dest.ssa.num_components);
3864
3865 write_result:
3866 if (result) {
3867 assert(instr->dest.is_ssa);
3868 result = ac_to_integer(&ctx->ac, result);
3869 ctx->ssa_defs[instr->dest.ssa.index] = result;
3870 }
3871 }
3872
3873
3874 static void visit_phi(struct ac_nir_context *ctx, nir_phi_instr *instr)
3875 {
3876 LLVMTypeRef type = get_def_type(ctx, &instr->dest.ssa);
3877 LLVMValueRef result = LLVMBuildPhi(ctx->ac.builder, type, "");
3878
3879 ctx->ssa_defs[instr->dest.ssa.index] = result;
3880 _mesa_hash_table_insert(ctx->phis, instr, result);
3881 }
3882
3883 static void visit_post_phi(struct ac_nir_context *ctx,
3884 nir_phi_instr *instr,
3885 LLVMValueRef llvm_phi)
3886 {
3887 nir_foreach_phi_src(src, instr) {
3888 LLVMBasicBlockRef block = get_block(ctx, src->pred);
3889 LLVMValueRef llvm_src = get_src(ctx, src->src);
3890
3891 LLVMAddIncoming(llvm_phi, &llvm_src, &block, 1);
3892 }
3893 }
3894
3895 static void phi_post_pass(struct ac_nir_context *ctx)
3896 {
3897 hash_table_foreach(ctx->phis, entry) {
3898 visit_post_phi(ctx, (nir_phi_instr*)entry->key,
3899 (LLVMValueRef)entry->data);
3900 }
3901 }
3902
3903
3904 static void visit_ssa_undef(struct ac_nir_context *ctx,
3905 const nir_ssa_undef_instr *instr)
3906 {
3907 unsigned num_components = instr->def.num_components;
3908 LLVMTypeRef type = LLVMIntTypeInContext(ctx->ac.context, instr->def.bit_size);
3909 LLVMValueRef undef;
3910
3911 if (num_components == 1)
3912 undef = LLVMGetUndef(type);
3913 else {
3914 undef = LLVMGetUndef(LLVMVectorType(type, num_components));
3915 }
3916 ctx->ssa_defs[instr->def.index] = undef;
3917 }
3918
3919 static void visit_jump(struct ac_llvm_context *ctx,
3920 const nir_jump_instr *instr)
3921 {
3922 switch (instr->type) {
3923 case nir_jump_break:
3924 ac_build_break(ctx);
3925 break;
3926 case nir_jump_continue:
3927 ac_build_continue(ctx);
3928 break;
3929 default:
3930 fprintf(stderr, "Unknown NIR jump instr: ");
3931 nir_print_instr(&instr->instr, stderr);
3932 fprintf(stderr, "\n");
3933 abort();
3934 }
3935 }
3936
3937 static LLVMTypeRef
3938 glsl_base_to_llvm_type(struct ac_llvm_context *ac,
3939 enum glsl_base_type type)
3940 {
3941 switch (type) {
3942 case GLSL_TYPE_INT:
3943 case GLSL_TYPE_UINT:
3944 case GLSL_TYPE_BOOL:
3945 case GLSL_TYPE_SUBROUTINE:
3946 return ac->i32;
3947 case GLSL_TYPE_INT8:
3948 case GLSL_TYPE_UINT8:
3949 return ac->i8;
3950 case GLSL_TYPE_INT16:
3951 case GLSL_TYPE_UINT16:
3952 return ac->i16;
3953 case GLSL_TYPE_FLOAT:
3954 return ac->f32;
3955 case GLSL_TYPE_FLOAT16:
3956 return ac->f16;
3957 case GLSL_TYPE_INT64:
3958 case GLSL_TYPE_UINT64:
3959 return ac->i64;
3960 case GLSL_TYPE_DOUBLE:
3961 return ac->f64;
3962 default:
3963 unreachable("unknown GLSL type");
3964 }
3965 }
3966
3967 static LLVMTypeRef
3968 glsl_to_llvm_type(struct ac_llvm_context *ac,
3969 const struct glsl_type *type)
3970 {
3971 if (glsl_type_is_scalar(type)) {
3972 return glsl_base_to_llvm_type(ac, glsl_get_base_type(type));
3973 }
3974
3975 if (glsl_type_is_vector(type)) {
3976 return LLVMVectorType(
3977 glsl_base_to_llvm_type(ac, glsl_get_base_type(type)),
3978 glsl_get_vector_elements(type));
3979 }
3980
3981 if (glsl_type_is_matrix(type)) {
3982 return LLVMArrayType(
3983 glsl_to_llvm_type(ac, glsl_get_column_type(type)),
3984 glsl_get_matrix_columns(type));
3985 }
3986
3987 if (glsl_type_is_array(type)) {
3988 return LLVMArrayType(
3989 glsl_to_llvm_type(ac, glsl_get_array_element(type)),
3990 glsl_get_length(type));
3991 }
3992
3993 assert(glsl_type_is_struct_or_ifc(type));
3994
3995 LLVMTypeRef member_types[glsl_get_length(type)];
3996
3997 for (unsigned i = 0; i < glsl_get_length(type); i++) {
3998 member_types[i] =
3999 glsl_to_llvm_type(ac,
4000 glsl_get_struct_field(type, i));
4001 }
4002
4003 return LLVMStructTypeInContext(ac->context, member_types,
4004 glsl_get_length(type), false);
4005 }
4006
4007 static void visit_deref(struct ac_nir_context *ctx,
4008 nir_deref_instr *instr)
4009 {
4010 if (instr->mode != nir_var_mem_shared &&
4011 instr->mode != nir_var_mem_global)
4012 return;
4013
4014 LLVMValueRef result = NULL;
4015 switch(instr->deref_type) {
4016 case nir_deref_type_var: {
4017 struct hash_entry *entry = _mesa_hash_table_search(ctx->vars, instr->var);
4018 result = entry->data;
4019 break;
4020 }
4021 case nir_deref_type_struct:
4022 if (instr->mode == nir_var_mem_global) {
4023 nir_deref_instr *parent = nir_deref_instr_parent(instr);
4024 uint64_t offset = glsl_get_struct_field_offset(parent->type,
4025 instr->strct.index);
4026 result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent),
4027 LLVMConstInt(ctx->ac.i32, offset, 0));
4028 } else {
4029 result = ac_build_gep0(&ctx->ac, get_src(ctx, instr->parent),
4030 LLVMConstInt(ctx->ac.i32, instr->strct.index, 0));
4031 }
4032 break;
4033 case nir_deref_type_array:
4034 if (instr->mode == nir_var_mem_global) {
4035 nir_deref_instr *parent = nir_deref_instr_parent(instr);
4036 unsigned stride = glsl_get_explicit_stride(parent->type);
4037
4038 if ((glsl_type_is_matrix(parent->type) &&
4039 glsl_matrix_type_is_row_major(parent->type)) ||
4040 (glsl_type_is_vector(parent->type) && stride == 0))
4041 stride = type_scalar_size_bytes(parent->type);
4042
4043 assert(stride > 0);
4044 LLVMValueRef index = get_src(ctx, instr->arr.index);
4045 if (LLVMTypeOf(index) != ctx->ac.i64)
4046 index = LLVMBuildZExt(ctx->ac.builder, index, ctx->ac.i64, "");
4047
4048 LLVMValueRef offset = LLVMBuildMul(ctx->ac.builder, index, LLVMConstInt(ctx->ac.i64, stride, 0), "");
4049
4050 result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent), offset);
4051 } else {
4052 result = ac_build_gep0(&ctx->ac, get_src(ctx, instr->parent),
4053 get_src(ctx, instr->arr.index));
4054 }
4055 break;
4056 case nir_deref_type_ptr_as_array:
4057 if (instr->mode == nir_var_mem_global) {
4058 unsigned stride = nir_deref_instr_ptr_as_array_stride(instr);
4059
4060 LLVMValueRef index = get_src(ctx, instr->arr.index);
4061 if (LLVMTypeOf(index) != ctx->ac.i64)
4062 index = LLVMBuildZExt(ctx->ac.builder, index, ctx->ac.i64, "");
4063
4064 LLVMValueRef offset = LLVMBuildMul(ctx->ac.builder, index, LLVMConstInt(ctx->ac.i64, stride, 0), "");
4065
4066 result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent), offset);
4067 } else {
4068 result = ac_build_gep_ptr(&ctx->ac, get_src(ctx, instr->parent),
4069 get_src(ctx, instr->arr.index));
4070 }
4071 break;
4072 case nir_deref_type_cast: {
4073 result = get_src(ctx, instr->parent);
4074
4075 /* We can't use the structs from LLVM because the shader
4076 * specifies its own offsets. */
4077 LLVMTypeRef pointee_type = ctx->ac.i8;
4078 if (instr->mode == nir_var_mem_shared)
4079 pointee_type = glsl_to_llvm_type(&ctx->ac, instr->type);
4080
4081 unsigned address_space;
4082
4083 switch(instr->mode) {
4084 case nir_var_mem_shared:
4085 address_space = AC_ADDR_SPACE_LDS;
4086 break;
4087 case nir_var_mem_global:
4088 address_space = AC_ADDR_SPACE_GLOBAL;
4089 break;
4090 default:
4091 unreachable("Unhandled address space");
4092 }
4093
4094 LLVMTypeRef type = LLVMPointerType(pointee_type, address_space);
4095
4096 if (LLVMTypeOf(result) != type) {
4097 if (LLVMGetTypeKind(LLVMTypeOf(result)) == LLVMVectorTypeKind) {
4098 result = LLVMBuildBitCast(ctx->ac.builder, result,
4099 type, "");
4100 } else {
4101 result = LLVMBuildIntToPtr(ctx->ac.builder, result,
4102 type, "");
4103 }
4104 }
4105 break;
4106 }
4107 default:
4108 unreachable("Unhandled deref_instr deref type");
4109 }
4110
4111 ctx->ssa_defs[instr->dest.ssa.index] = result;
4112 }
4113
4114 static void visit_cf_list(struct ac_nir_context *ctx,
4115 struct exec_list *list);
4116
4117 static void visit_block(struct ac_nir_context *ctx, nir_block *block)
4118 {
4119 LLVMBasicBlockRef llvm_block = LLVMGetInsertBlock(ctx->ac.builder);
4120 nir_foreach_instr(instr, block)
4121 {
4122 switch (instr->type) {
4123 case nir_instr_type_alu:
4124 visit_alu(ctx, nir_instr_as_alu(instr));
4125 break;
4126 case nir_instr_type_load_const:
4127 visit_load_const(ctx, nir_instr_as_load_const(instr));
4128 break;
4129 case nir_instr_type_intrinsic:
4130 visit_intrinsic(ctx, nir_instr_as_intrinsic(instr));
4131 break;
4132 case nir_instr_type_tex:
4133 visit_tex(ctx, nir_instr_as_tex(instr));
4134 break;
4135 case nir_instr_type_phi:
4136 visit_phi(ctx, nir_instr_as_phi(instr));
4137 break;
4138 case nir_instr_type_ssa_undef:
4139 visit_ssa_undef(ctx, nir_instr_as_ssa_undef(instr));
4140 break;
4141 case nir_instr_type_jump:
4142 visit_jump(&ctx->ac, nir_instr_as_jump(instr));
4143 break;
4144 case nir_instr_type_deref:
4145 visit_deref(ctx, nir_instr_as_deref(instr));
4146 break;
4147 default:
4148 fprintf(stderr, "Unknown NIR instr type: ");
4149 nir_print_instr(instr, stderr);
4150 fprintf(stderr, "\n");
4151 abort();
4152 }
4153 }
4154
4155 _mesa_hash_table_insert(ctx->defs, block, llvm_block);
4156 }
4157
4158 static void visit_if(struct ac_nir_context *ctx, nir_if *if_stmt)
4159 {
4160 LLVMValueRef value = get_src(ctx, if_stmt->condition);
4161
4162 nir_block *then_block =
4163 (nir_block *) exec_list_get_head(&if_stmt->then_list);
4164
4165 ac_build_uif(&ctx->ac, value, then_block->index);
4166
4167 visit_cf_list(ctx, &if_stmt->then_list);
4168
4169 if (!exec_list_is_empty(&if_stmt->else_list)) {
4170 nir_block *else_block =
4171 (nir_block *) exec_list_get_head(&if_stmt->else_list);
4172
4173 ac_build_else(&ctx->ac, else_block->index);
4174 visit_cf_list(ctx, &if_stmt->else_list);
4175 }
4176
4177 ac_build_endif(&ctx->ac, then_block->index);
4178 }
4179
4180 static void visit_loop(struct ac_nir_context *ctx, nir_loop *loop)
4181 {
4182 nir_block *first_loop_block =
4183 (nir_block *) exec_list_get_head(&loop->body);
4184
4185 ac_build_bgnloop(&ctx->ac, first_loop_block->index);
4186
4187 visit_cf_list(ctx, &loop->body);
4188
4189 ac_build_endloop(&ctx->ac, first_loop_block->index);
4190 }
4191
4192 static void visit_cf_list(struct ac_nir_context *ctx,
4193 struct exec_list *list)
4194 {
4195 foreach_list_typed(nir_cf_node, node, node, list)
4196 {
4197 switch (node->type) {
4198 case nir_cf_node_block:
4199 visit_block(ctx, nir_cf_node_as_block(node));
4200 break;
4201
4202 case nir_cf_node_if:
4203 visit_if(ctx, nir_cf_node_as_if(node));
4204 break;
4205
4206 case nir_cf_node_loop:
4207 visit_loop(ctx, nir_cf_node_as_loop(node));
4208 break;
4209
4210 default:
4211 assert(0);
4212 }
4213 }
4214 }
4215
4216 void
4217 ac_handle_shader_output_decl(struct ac_llvm_context *ctx,
4218 struct ac_shader_abi *abi,
4219 struct nir_shader *nir,
4220 struct nir_variable *variable,
4221 gl_shader_stage stage)
4222 {
4223 unsigned output_loc = variable->data.driver_location / 4;
4224 unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
4225
4226 /* tess ctrl has it's own load/store paths for outputs */
4227 if (stage == MESA_SHADER_TESS_CTRL)
4228 return;
4229
4230 if (stage == MESA_SHADER_VERTEX ||
4231 stage == MESA_SHADER_TESS_EVAL ||
4232 stage == MESA_SHADER_GEOMETRY) {
4233 int idx = variable->data.location + variable->data.index;
4234 if (idx == VARYING_SLOT_CLIP_DIST0) {
4235 int length = nir->info.clip_distance_array_size +
4236 nir->info.cull_distance_array_size;
4237
4238 if (length > 4)
4239 attrib_count = 2;
4240 else
4241 attrib_count = 1;
4242 }
4243 }
4244
4245 bool is_16bit = glsl_type_is_16bit(glsl_without_array(variable->type));
4246 LLVMTypeRef type = is_16bit ? ctx->f16 : ctx->f32;
4247 for (unsigned i = 0; i < attrib_count; ++i) {
4248 for (unsigned chan = 0; chan < 4; chan++) {
4249 abi->outputs[ac_llvm_reg_index_soa(output_loc + i, chan)] =
4250 ac_build_alloca_undef(ctx, type, "");
4251 }
4252 }
4253 }
4254
4255 static void
4256 setup_locals(struct ac_nir_context *ctx,
4257 struct nir_function *func)
4258 {
4259 int i, j;
4260 ctx->num_locals = 0;
4261 nir_foreach_variable(variable, &func->impl->locals) {
4262 unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
4263 variable->data.driver_location = ctx->num_locals * 4;
4264 variable->data.location_frac = 0;
4265 ctx->num_locals += attrib_count;
4266 }
4267 ctx->locals = malloc(4 * ctx->num_locals * sizeof(LLVMValueRef));
4268 if (!ctx->locals)
4269 return;
4270
4271 for (i = 0; i < ctx->num_locals; i++) {
4272 for (j = 0; j < 4; j++) {
4273 ctx->locals[i * 4 + j] =
4274 ac_build_alloca_undef(&ctx->ac, ctx->ac.f32, "temp");
4275 }
4276 }
4277 }
4278
4279 static void
4280 setup_shared(struct ac_nir_context *ctx,
4281 struct nir_shader *nir)
4282 {
4283 nir_foreach_variable(variable, &nir->shared) {
4284 LLVMValueRef shared =
4285 LLVMAddGlobalInAddressSpace(
4286 ctx->ac.module, glsl_to_llvm_type(&ctx->ac, variable->type),
4287 variable->name ? variable->name : "",
4288 AC_ADDR_SPACE_LDS);
4289 _mesa_hash_table_insert(ctx->vars, variable, shared);
4290 }
4291 }
4292
4293 void ac_nir_translate(struct ac_llvm_context *ac, struct ac_shader_abi *abi,
4294 struct nir_shader *nir)
4295 {
4296 struct ac_nir_context ctx = {};
4297 struct nir_function *func;
4298
4299 ctx.ac = *ac;
4300 ctx.abi = abi;
4301
4302 ctx.stage = nir->info.stage;
4303
4304 ctx.main_function = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx.ac.builder));
4305
4306 nir_foreach_variable(variable, &nir->outputs)
4307 ac_handle_shader_output_decl(&ctx.ac, ctx.abi, nir, variable,
4308 ctx.stage);
4309
4310 ctx.defs = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
4311 _mesa_key_pointer_equal);
4312 ctx.phis = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
4313 _mesa_key_pointer_equal);
4314 ctx.vars = _mesa_hash_table_create(NULL, _mesa_hash_pointer,
4315 _mesa_key_pointer_equal);
4316
4317 func = (struct nir_function *)exec_list_get_head(&nir->functions);
4318
4319 nir_index_ssa_defs(func->impl);
4320 ctx.ssa_defs = calloc(func->impl->ssa_alloc, sizeof(LLVMValueRef));
4321
4322 setup_locals(&ctx, func);
4323
4324 if (gl_shader_stage_is_compute(nir->info.stage))
4325 setup_shared(&ctx, nir);
4326
4327 visit_cf_list(&ctx, &func->impl->body);
4328 phi_post_pass(&ctx);
4329
4330 if (!gl_shader_stage_is_compute(nir->info.stage))
4331 ctx.abi->emit_outputs(ctx.abi, AC_LLVM_MAX_OUTPUTS,
4332 ctx.abi->outputs);
4333
4334 free(ctx.locals);
4335 free(ctx.ssa_defs);
4336 ralloc_free(ctx.defs);
4337 ralloc_free(ctx.phis);
4338 ralloc_free(ctx.vars);
4339 }
4340
4341 void
4342 ac_lower_indirect_derefs(struct nir_shader *nir, enum chip_class chip_class)
4343 {
4344 /* While it would be nice not to have this flag, we are constrained
4345 * by the reality that LLVM 5.0 doesn't have working VGPR indexing
4346 * on GFX9.
4347 */
4348 bool llvm_has_working_vgpr_indexing = chip_class <= VI;
4349
4350 /* TODO: Indirect indexing of GS inputs is unimplemented.
4351 *
4352 * TCS and TES load inputs directly from LDS or offchip memory, so
4353 * indirect indexing is trivial.
4354 */
4355 nir_variable_mode indirect_mask = 0;
4356 if (nir->info.stage == MESA_SHADER_GEOMETRY ||
4357 (nir->info.stage != MESA_SHADER_TESS_CTRL &&
4358 nir->info.stage != MESA_SHADER_TESS_EVAL &&
4359 !llvm_has_working_vgpr_indexing)) {
4360 indirect_mask |= nir_var_shader_in;
4361 }
4362 if (!llvm_has_working_vgpr_indexing &&
4363 nir->info.stage != MESA_SHADER_TESS_CTRL)
4364 indirect_mask |= nir_var_shader_out;
4365
4366 /* TODO: We shouldn't need to do this, however LLVM isn't currently
4367 * smart enough to handle indirects without causing excess spilling
4368 * causing the gpu to hang.
4369 *
4370 * See the following thread for more details of the problem:
4371 * https://lists.freedesktop.org/archives/mesa-dev/2017-July/162106.html
4372 */
4373 indirect_mask |= nir_var_function_temp;
4374
4375 nir_lower_indirect_derefs(nir, indirect_mask);
4376 }
4377
4378 static unsigned
4379 get_inst_tessfactor_writemask(nir_intrinsic_instr *intrin)
4380 {
4381 if (intrin->intrinsic != nir_intrinsic_store_deref)
4382 return 0;
4383
4384 nir_variable *var =
4385 nir_deref_instr_get_variable(nir_src_as_deref(intrin->src[0]));
4386
4387 if (var->data.mode != nir_var_shader_out)
4388 return 0;
4389
4390 unsigned writemask = 0;
4391 const int location = var->data.location;
4392 unsigned first_component = var->data.location_frac;
4393 unsigned num_comps = intrin->dest.ssa.num_components;
4394
4395 if (location == VARYING_SLOT_TESS_LEVEL_INNER)
4396 writemask = ((1 << (num_comps + 1)) - 1) << first_component;
4397 else if (location == VARYING_SLOT_TESS_LEVEL_OUTER)
4398 writemask = (((1 << (num_comps + 1)) - 1) << first_component) << 4;
4399
4400 return writemask;
4401 }
4402
4403 static void
4404 scan_tess_ctrl(nir_cf_node *cf_node, unsigned *upper_block_tf_writemask,
4405 unsigned *cond_block_tf_writemask,
4406 bool *tessfactors_are_def_in_all_invocs, bool is_nested_cf)
4407 {
4408 switch (cf_node->type) {
4409 case nir_cf_node_block: {
4410 nir_block *block = nir_cf_node_as_block(cf_node);
4411 nir_foreach_instr(instr, block) {
4412 if (instr->type != nir_instr_type_intrinsic)
4413 continue;
4414
4415 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
4416 if (intrin->intrinsic == nir_intrinsic_barrier) {
4417
4418 /* If we find a barrier in nested control flow put this in the
4419 * too hard basket. In GLSL this is not possible but it is in
4420 * SPIR-V.
4421 */
4422 if (is_nested_cf) {
4423 *tessfactors_are_def_in_all_invocs = false;
4424 return;
4425 }
4426
4427 /* The following case must be prevented:
4428 * gl_TessLevelInner = ...;
4429 * barrier();
4430 * if (gl_InvocationID == 1)
4431 * gl_TessLevelInner = ...;
4432 *
4433 * If you consider disjoint code segments separated by barriers, each
4434 * such segment that writes tess factor channels should write the same
4435 * channels in all codepaths within that segment.
4436 */
4437 if (upper_block_tf_writemask || cond_block_tf_writemask) {
4438 /* Accumulate the result: */
4439 *tessfactors_are_def_in_all_invocs &=
4440 !(*cond_block_tf_writemask & ~(*upper_block_tf_writemask));
4441
4442 /* Analyze the next code segment from scratch. */
4443 *upper_block_tf_writemask = 0;
4444 *cond_block_tf_writemask = 0;
4445 }
4446 } else
4447 *upper_block_tf_writemask |= get_inst_tessfactor_writemask(intrin);
4448 }
4449
4450 break;
4451 }
4452 case nir_cf_node_if: {
4453 unsigned then_tessfactor_writemask = 0;
4454 unsigned else_tessfactor_writemask = 0;
4455
4456 nir_if *if_stmt = nir_cf_node_as_if(cf_node);
4457 foreach_list_typed(nir_cf_node, nested_node, node, &if_stmt->then_list) {
4458 scan_tess_ctrl(nested_node, &then_tessfactor_writemask,
4459 cond_block_tf_writemask,
4460 tessfactors_are_def_in_all_invocs, true);
4461 }
4462
4463 foreach_list_typed(nir_cf_node, nested_node, node, &if_stmt->else_list) {
4464 scan_tess_ctrl(nested_node, &else_tessfactor_writemask,
4465 cond_block_tf_writemask,
4466 tessfactors_are_def_in_all_invocs, true);
4467 }
4468
4469 if (then_tessfactor_writemask || else_tessfactor_writemask) {
4470 /* If both statements write the same tess factor channels,
4471 * we can say that the upper block writes them too.
4472 */
4473 *upper_block_tf_writemask |= then_tessfactor_writemask &
4474 else_tessfactor_writemask;
4475 *cond_block_tf_writemask |= then_tessfactor_writemask |
4476 else_tessfactor_writemask;
4477 }
4478
4479 break;
4480 }
4481 case nir_cf_node_loop: {
4482 nir_loop *loop = nir_cf_node_as_loop(cf_node);
4483 foreach_list_typed(nir_cf_node, nested_node, node, &loop->body) {
4484 scan_tess_ctrl(nested_node, cond_block_tf_writemask,
4485 cond_block_tf_writemask,
4486 tessfactors_are_def_in_all_invocs, true);
4487 }
4488
4489 break;
4490 }
4491 default:
4492 unreachable("unknown cf node type");
4493 }
4494 }
4495
4496 bool
4497 ac_are_tessfactors_def_in_all_invocs(const struct nir_shader *nir)
4498 {
4499 assert(nir->info.stage == MESA_SHADER_TESS_CTRL);
4500
4501 /* The pass works as follows:
4502 * If all codepaths write tess factors, we can say that all
4503 * invocations define tess factors.
4504 *
4505 * Each tess factor channel is tracked separately.
4506 */
4507 unsigned main_block_tf_writemask = 0; /* if main block writes tess factors */
4508 unsigned cond_block_tf_writemask = 0; /* if cond block writes tess factors */
4509
4510 /* Initial value = true. Here the pass will accumulate results from
4511 * multiple segments surrounded by barriers. If tess factors aren't
4512 * written at all, it's a shader bug and we don't care if this will be
4513 * true.
4514 */
4515 bool tessfactors_are_def_in_all_invocs = true;
4516
4517 nir_foreach_function(function, nir) {
4518 if (function->impl) {
4519 foreach_list_typed(nir_cf_node, node, node, &function->impl->body) {
4520 scan_tess_ctrl(node, &main_block_tf_writemask,
4521 &cond_block_tf_writemask,
4522 &tessfactors_are_def_in_all_invocs,
4523 false);
4524 }
4525 }
4526 }
4527
4528 /* Accumulate the result for the last code segment separated by a
4529 * barrier.
4530 */
4531 if (main_block_tf_writemask || cond_block_tf_writemask) {
4532 tessfactors_are_def_in_all_invocs &=
4533 !(cond_block_tf_writemask & ~main_block_tf_writemask);
4534 }
4535
4536 return tessfactors_are_def_in_all_invocs;
4537 }