projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
ac/nir: support vector types in the type suffix of overloaded intrinsics
[mesa.git] / src / amd / llvm / ac_llvm_build.c
1 /*
2 * Copyright 2014 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the
6 * "Software"), to deal in the Software without restriction, including
7 * without limitation the rights to use, copy, modify, merge, publish,
8 * distribute, sub license, and/or sell copies of the Software, and to
9 * permit persons to whom the Software is furnished to do so, subject to
10 * the following conditions:
11 *
12 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
13 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
14 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
15 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
16 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
17 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
18 * USE OR OTHER DEALINGS IN THE SOFTWARE.
19 *
20 * The above copyright notice and this permission notice (including the
21 * next paragraph) shall be included in all copies or substantial portions
22 * of the Software.
23 *
24 */
25 /* based on pieces from si_pipe.c and radeon_llvm_emit.c */
26 #include "ac_llvm_build.h"
27
28 #include <llvm-c/Core.h>
29 #include <llvm/Config/llvm-config.h>
30
31 #include "c11/threads.h"
32
33 #include <assert.h>
34 #include <stdio.h>
35
36 #include "ac_llvm_util.h"
37 #include "ac_shader_util.h"
38 #include "ac_exp_param.h"
39 #include "util/bitscan.h"
40 #include "util/macros.h"
41 #include "util/u_atomic.h"
42 #include "util/u_math.h"
43 #include "sid.h"
44
45 #include "shader_enums.h"
46
47 #define AC_LLVM_INITIAL_CF_DEPTH 4
48
49 /* Data for if/else/endif and bgnloop/endloop control flow structures.
50 */
51 struct ac_llvm_flow {
52 /* Loop exit or next part of if/else/endif. */
53 LLVMBasicBlockRef next_block;
54 LLVMBasicBlockRef loop_entry_block;
55 };
56
57 /* Initialize module-independent parts of the context.
58 *
59 * The caller is responsible for initializing ctx::module and ctx::builder.
60 */
61 void
62 ac_llvm_context_init(struct ac_llvm_context *ctx,
63 struct ac_llvm_compiler *compiler,
64 enum chip_class chip_class, enum radeon_family family,
65 enum ac_float_mode float_mode, unsigned wave_size,
66 unsigned ballot_mask_bits)
67 {
68 ctx->context = LLVMContextCreate();
69
70 ctx->chip_class = chip_class;
71 ctx->family = family;
72 ctx->wave_size = wave_size;
73 ctx->ballot_mask_bits = ballot_mask_bits;
74 ctx->float_mode = float_mode;
75 ctx->module = ac_create_module(wave_size == 32 ? compiler->tm_wave32
76 : compiler->tm,
77 ctx->context);
78 ctx->builder = ac_create_builder(ctx->context, float_mode);
79
80 ctx->voidt = LLVMVoidTypeInContext(ctx->context);
81 ctx->i1 = LLVMInt1TypeInContext(ctx->context);
82 ctx->i8 = LLVMInt8TypeInContext(ctx->context);
83 ctx->i16 = LLVMIntTypeInContext(ctx->context, 16);
84 ctx->i32 = LLVMIntTypeInContext(ctx->context, 32);
85 ctx->i64 = LLVMIntTypeInContext(ctx->context, 64);
86 ctx->i128 = LLVMIntTypeInContext(ctx->context, 128);
87 ctx->intptr = ctx->i32;
88 ctx->f16 = LLVMHalfTypeInContext(ctx->context);
89 ctx->f32 = LLVMFloatTypeInContext(ctx->context);
90 ctx->f64 = LLVMDoubleTypeInContext(ctx->context);
91 ctx->v2i16 = LLVMVectorType(ctx->i16, 2);
92 ctx->v4i16 = LLVMVectorType(ctx->i16, 4);
93 ctx->v2f16 = LLVMVectorType(ctx->f16, 2);
94 ctx->v4f16 = LLVMVectorType(ctx->f16, 4);
95 ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
96 ctx->v3i32 = LLVMVectorType(ctx->i32, 3);
97 ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
98 ctx->v2f32 = LLVMVectorType(ctx->f32, 2);
99 ctx->v3f32 = LLVMVectorType(ctx->f32, 3);
100 ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
101 ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
102 ctx->iN_wavemask = LLVMIntTypeInContext(ctx->context, ctx->wave_size);
103 ctx->iN_ballotmask = LLVMIntTypeInContext(ctx->context, ballot_mask_bits);
104
105 ctx->i8_0 = LLVMConstInt(ctx->i8, 0, false);
106 ctx->i8_1 = LLVMConstInt(ctx->i8, 1, false);
107 ctx->i16_0 = LLVMConstInt(ctx->i16, 0, false);
108 ctx->i16_1 = LLVMConstInt(ctx->i16, 1, false);
109 ctx->i32_0 = LLVMConstInt(ctx->i32, 0, false);
110 ctx->i32_1 = LLVMConstInt(ctx->i32, 1, false);
111 ctx->i64_0 = LLVMConstInt(ctx->i64, 0, false);
112 ctx->i64_1 = LLVMConstInt(ctx->i64, 1, false);
113 ctx->i128_0 = LLVMConstInt(ctx->i128, 0, false);
114 ctx->i128_1 = LLVMConstInt(ctx->i128, 1, false);
115 ctx->f16_0 = LLVMConstReal(ctx->f16, 0.0);
116 ctx->f16_1 = LLVMConstReal(ctx->f16, 1.0);
117 ctx->f32_0 = LLVMConstReal(ctx->f32, 0.0);
118 ctx->f32_1 = LLVMConstReal(ctx->f32, 1.0);
119 ctx->f64_0 = LLVMConstReal(ctx->f64, 0.0);
120 ctx->f64_1 = LLVMConstReal(ctx->f64, 1.0);
121
122 ctx->i1false = LLVMConstInt(ctx->i1, 0, false);
123 ctx->i1true = LLVMConstInt(ctx->i1, 1, false);
124
125 ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context,
126 "range", 5);
127
128 ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context,
129 "invariant.load", 14);
130
131 ctx->uniform_md_kind = LLVMGetMDKindIDInContext(ctx->context,
132 "amdgpu.uniform", 14);
133
134 ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
135 ctx->flow = calloc(1, sizeof(*ctx->flow));
136 }
137
138 void
139 ac_llvm_context_dispose(struct ac_llvm_context *ctx)
140 {
141 free(ctx->flow->stack);
142 free(ctx->flow);
143 ctx->flow = NULL;
144 }
145
146 int
147 ac_get_llvm_num_components(LLVMValueRef value)
148 {
149 LLVMTypeRef type = LLVMTypeOf(value);
150 unsigned num_components = LLVMGetTypeKind(type) == LLVMVectorTypeKind
151 ? LLVMGetVectorSize(type)
152 : 1;
153 return num_components;
154 }
155
156 LLVMValueRef
157 ac_llvm_extract_elem(struct ac_llvm_context *ac,
158 LLVMValueRef value,
159 int index)
160 {
161 if (LLVMGetTypeKind(LLVMTypeOf(value)) != LLVMVectorTypeKind) {
162 assert(index == 0);
163 return value;
164 }
165
166 return LLVMBuildExtractElement(ac->builder, value,
167 LLVMConstInt(ac->i32, index, false), "");
168 }
169
170 int
171 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type)
172 {
173 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
174 type = LLVMGetElementType(type);
175
176 if (LLVMGetTypeKind(type) == LLVMIntegerTypeKind)
177 return LLVMGetIntTypeWidth(type);
178
179 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
180 if (LLVMGetPointerAddressSpace(type) == AC_ADDR_SPACE_LDS)
181 return 32;
182 }
183
184 if (type == ctx->f16)
185 return 16;
186 if (type == ctx->f32)
187 return 32;
188 if (type == ctx->f64)
189 return 64;
190
191 unreachable("Unhandled type kind in get_elem_bits");
192 }
193
194 unsigned
195 ac_get_type_size(LLVMTypeRef type)
196 {
197 LLVMTypeKind kind = LLVMGetTypeKind(type);
198
199 switch (kind) {
200 case LLVMIntegerTypeKind:
201 return LLVMGetIntTypeWidth(type) / 8;
202 case LLVMHalfTypeKind:
203 return 2;
204 case LLVMFloatTypeKind:
205 return 4;
206 case LLVMDoubleTypeKind:
207 return 8;
208 case LLVMPointerTypeKind:
209 if (LLVMGetPointerAddressSpace(type) == AC_ADDR_SPACE_CONST_32BIT)
210 return 4;
211 return 8;
212 case LLVMVectorTypeKind:
213 return LLVMGetVectorSize(type) *
214 ac_get_type_size(LLVMGetElementType(type));
215 case LLVMArrayTypeKind:
216 return LLVMGetArrayLength(type) *
217 ac_get_type_size(LLVMGetElementType(type));
218 default:
219 assert(0);
220 return 0;
221 }
222 }
223
224 static LLVMTypeRef to_integer_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
225 {
226 if (t == ctx->i8)
227 return ctx->i8;
228 else if (t == ctx->f16 || t == ctx->i16)
229 return ctx->i16;
230 else if (t == ctx->f32 || t == ctx->i32)
231 return ctx->i32;
232 else if (t == ctx->f64 || t == ctx->i64)
233 return ctx->i64;
234 else
235 unreachable("Unhandled integer size");
236 }
237
238 LLVMTypeRef
239 ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
240 {
241 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
242 LLVMTypeRef elem_type = LLVMGetElementType(t);
243 return LLVMVectorType(to_integer_type_scalar(ctx, elem_type),
244 LLVMGetVectorSize(t));
245 }
246 if (LLVMGetTypeKind(t) == LLVMPointerTypeKind) {
247 switch (LLVMGetPointerAddressSpace(t)) {
248 case AC_ADDR_SPACE_GLOBAL:
249 return ctx->i64;
250 case AC_ADDR_SPACE_CONST_32BIT:
251 case AC_ADDR_SPACE_LDS:
252 return ctx->i32;
253 default:
254 unreachable("unhandled address space");
255 }
256 }
257 return to_integer_type_scalar(ctx, t);
258 }
259
260 LLVMValueRef
261 ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v)
262 {
263 LLVMTypeRef type = LLVMTypeOf(v);
264 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
265 return LLVMBuildPtrToInt(ctx->builder, v, ac_to_integer_type(ctx, type), "");
266 }
267 return LLVMBuildBitCast(ctx->builder, v, ac_to_integer_type(ctx, type), "");
268 }
269
270 LLVMValueRef
271 ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v)
272 {
273 LLVMTypeRef type = LLVMTypeOf(v);
274 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind)
275 return v;
276 return ac_to_integer(ctx, v);
277 }
278
279 static LLVMTypeRef to_float_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
280 {
281 if (t == ctx->i8)
282 return ctx->i8;
283 else if (t == ctx->i16 || t == ctx->f16)
284 return ctx->f16;
285 else if (t == ctx->i32 || t == ctx->f32)
286 return ctx->f32;
287 else if (t == ctx->i64 || t == ctx->f64)
288 return ctx->f64;
289 else
290 unreachable("Unhandled float size");
291 }
292
293 LLVMTypeRef
294 ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
295 {
296 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
297 LLVMTypeRef elem_type = LLVMGetElementType(t);
298 return LLVMVectorType(to_float_type_scalar(ctx, elem_type),
299 LLVMGetVectorSize(t));
300 }
301 return to_float_type_scalar(ctx, t);
302 }
303
304 LLVMValueRef
305 ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v)
306 {
307 LLVMTypeRef type = LLVMTypeOf(v);
308 return LLVMBuildBitCast(ctx->builder, v, ac_to_float_type(ctx, type), "");
309 }
310
311
312 LLVMValueRef
313 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
314 LLVMTypeRef return_type, LLVMValueRef *params,
315 unsigned param_count, unsigned attrib_mask)
316 {
317 LLVMValueRef function, call;
318 bool set_callsite_attrs = !(attrib_mask & AC_FUNC_ATTR_LEGACY);
319
320 function = LLVMGetNamedFunction(ctx->module, name);
321 if (!function) {
322 LLVMTypeRef param_types[32], function_type;
323 unsigned i;
324
325 assert(param_count <= 32);
326
327 for (i = 0; i < param_count; ++i) {
328 assert(params[i]);
329 param_types[i] = LLVMTypeOf(params[i]);
330 }
331 function_type =
332 LLVMFunctionType(return_type, param_types, param_count, 0);
333 function = LLVMAddFunction(ctx->module, name, function_type);
334
335 LLVMSetFunctionCallConv(function, LLVMCCallConv);
336 LLVMSetLinkage(function, LLVMExternalLinkage);
337
338 if (!set_callsite_attrs)
339 ac_add_func_attributes(ctx->context, function, attrib_mask);
340 }
341
342 call = LLVMBuildCall(ctx->builder, function, params, param_count, "");
343 if (set_callsite_attrs)
344 ac_add_func_attributes(ctx->context, call, attrib_mask);
345 return call;
346 }
347
348 /**
349 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
350 * intrinsic names).
351 */
352 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize)
353 {
354 LLVMTypeRef elem_type = type;
355
356 assert(bufsize >= 8);
357
358 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
359 int ret = snprintf(buf, bufsize, "v%u",
360 LLVMGetVectorSize(type));
361 if (ret < 0) {
362 char *type_name = LLVMPrintTypeToString(type);
363 fprintf(stderr, "Error building type name for: %s\n",
364 type_name);
365 LLVMDisposeMessage(type_name);
366 return;
367 }
368 elem_type = LLVMGetElementType(type);
369 buf += ret;
370 bufsize -= ret;
371 }
372 switch (LLVMGetTypeKind(elem_type)) {
373 default: break;
374 case LLVMIntegerTypeKind:
375 snprintf(buf, bufsize, "i%d", LLVMGetIntTypeWidth(elem_type));
376 break;
377 case LLVMHalfTypeKind:
378 snprintf(buf, bufsize, "f16");
379 break;
380 case LLVMFloatTypeKind:
381 snprintf(buf, bufsize, "f32");
382 break;
383 case LLVMDoubleTypeKind:
384 snprintf(buf, bufsize, "f64");
385 break;
386 }
387 }
388
389 /**
390 * Helper function that builds an LLVM IR PHI node and immediately adds
391 * incoming edges.
392 */
393 LLVMValueRef
394 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
395 unsigned count_incoming, LLVMValueRef *values,
396 LLVMBasicBlockRef *blocks)
397 {
398 LLVMValueRef phi = LLVMBuildPhi(ctx->builder, type, "");
399 LLVMAddIncoming(phi, values, blocks, count_incoming);
400 return phi;
401 }
402
403 void ac_build_s_barrier(struct ac_llvm_context *ctx)
404 {
405 ac_build_intrinsic(ctx, "llvm.amdgcn.s.barrier", ctx->voidt, NULL,
406 0, AC_FUNC_ATTR_CONVERGENT);
407 }
408
409 /* Prevent optimizations (at least of memory accesses) across the current
410 * point in the program by emitting empty inline assembly that is marked as
411 * having side effects.
412 *
413 * Optionally, a value can be passed through the inline assembly to prevent
414 * LLVM from hoisting calls to ReadNone functions.
415 */
416 void
417 ac_build_optimization_barrier(struct ac_llvm_context *ctx,
418 LLVMValueRef *pvgpr)
419 {
420 static int counter = 0;
421
422 LLVMBuilderRef builder = ctx->builder;
423 char code[16];
424
425 snprintf(code, sizeof(code), "; %d", p_atomic_inc_return(&counter));
426
427 if (!pvgpr) {
428 LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
429 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "", true, false);
430 LLVMBuildCall(builder, inlineasm, NULL, 0, "");
431 } else {
432 LLVMTypeRef ftype = LLVMFunctionType(ctx->i32, &ctx->i32, 1, false);
433 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "=v,0", true, false);
434 LLVMTypeRef type = LLVMTypeOf(*pvgpr);
435 unsigned bitsize = ac_get_elem_bits(ctx, type);
436 LLVMValueRef vgpr = *pvgpr;
437 LLVMTypeRef vgpr_type;
438 unsigned vgpr_size;
439 LLVMValueRef vgpr0;
440
441 if (bitsize < 32)
442 vgpr = LLVMBuildZExt(ctx->builder, vgpr, ctx->i32, "");
443
444 vgpr_type = LLVMTypeOf(vgpr);
445 vgpr_size = ac_get_type_size(vgpr_type);
446
447 assert(vgpr_size % 4 == 0);
448
449 vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
450 vgpr0 = LLVMBuildExtractElement(builder, vgpr, ctx->i32_0, "");
451 vgpr0 = LLVMBuildCall(builder, inlineasm, &vgpr0, 1, "");
452 vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
453 vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
454
455 if (bitsize < 32)
456 vgpr = LLVMBuildTrunc(builder, vgpr, type, "");
457
458 *pvgpr = vgpr;
459 }
460 }
461
462 LLVMValueRef
463 ac_build_shader_clock(struct ac_llvm_context *ctx, nir_scope scope)
464 {
465 const char *name = scope == NIR_SCOPE_DEVICE ? "llvm.amdgcn.s.memrealtime" : "llvm.amdgcn.s.memtime";
466 LLVMValueRef tmp = ac_build_intrinsic(ctx, name, ctx->i64, NULL, 0, 0);
467 return LLVMBuildBitCast(ctx->builder, tmp, ctx->v2i32, "");
468 }
469
470 LLVMValueRef
471 ac_build_ballot(struct ac_llvm_context *ctx,
472 LLVMValueRef value)
473 {
474 const char *name;
475
476 if (LLVM_VERSION_MAJOR >= 9) {
477 if (ctx->wave_size == 64)
478 name = "llvm.amdgcn.icmp.i64.i32";
479 else
480 name = "llvm.amdgcn.icmp.i32.i32";
481 } else {
482 name = "llvm.amdgcn.icmp.i32";
483 }
484 LLVMValueRef args[3] = {
485 value,
486 ctx->i32_0,
487 LLVMConstInt(ctx->i32, LLVMIntNE, 0)
488 };
489
490 /* We currently have no other way to prevent LLVM from lifting the icmp
491 * calls to a dominating basic block.
492 */
493 ac_build_optimization_barrier(ctx, &args[0]);
494
495 args[0] = ac_to_integer(ctx, args[0]);
496
497 return ac_build_intrinsic(ctx, name, ctx->iN_wavemask, args, 3,
498 AC_FUNC_ATTR_NOUNWIND |
499 AC_FUNC_ATTR_READNONE |
500 AC_FUNC_ATTR_CONVERGENT);
501 }
502
503 LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
504 LLVMValueRef value)
505 {
506 const char *name;
507
508 if (LLVM_VERSION_MAJOR >= 9) {
509 if (ctx->wave_size == 64)
510 name = "llvm.amdgcn.icmp.i64.i1";
511 else
512 name = "llvm.amdgcn.icmp.i32.i1";
513 } else {
514 name = "llvm.amdgcn.icmp.i1";
515 }
516 LLVMValueRef args[3] = {
517 value,
518 ctx->i1false,
519 LLVMConstInt(ctx->i32, LLVMIntNE, 0),
520 };
521
522 return ac_build_intrinsic(ctx, name, ctx->iN_wavemask, args, 3,
523 AC_FUNC_ATTR_NOUNWIND |
524 AC_FUNC_ATTR_READNONE |
525 AC_FUNC_ATTR_CONVERGENT);
526 }
527
528 LLVMValueRef
529 ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value)
530 {
531 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
532 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
533 return LLVMBuildICmp(ctx->builder, LLVMIntEQ, vote_set, active_set, "");
534 }
535
536 LLVMValueRef
537 ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value)
538 {
539 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
540 return LLVMBuildICmp(ctx->builder, LLVMIntNE, vote_set,
541 LLVMConstInt(ctx->iN_wavemask, 0, 0), "");
542 }
543
544 LLVMValueRef
545 ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value)
546 {
547 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
548 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
549
550 LLVMValueRef all = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
551 vote_set, active_set, "");
552 LLVMValueRef none = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
553 vote_set,
554 LLVMConstInt(ctx->iN_wavemask, 0, 0), "");
555 return LLVMBuildOr(ctx->builder, all, none, "");
556 }
557
558 LLVMValueRef
559 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
560 unsigned value_count, unsigned component)
561 {
562 LLVMValueRef vec = NULL;
563
564 if (value_count == 1) {
565 return values[component];
566 } else if (!value_count)
567 unreachable("value_count is 0");
568
569 for (unsigned i = component; i < value_count + component; i++) {
570 LLVMValueRef value = values[i];
571
572 if (i == component)
573 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
574 LLVMValueRef index = LLVMConstInt(ctx->i32, i - component, false);
575 vec = LLVMBuildInsertElement(ctx->builder, vec, value, index, "");
576 }
577 return vec;
578 }
579
580 LLVMValueRef
581 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
582 LLVMValueRef *values,
583 unsigned value_count,
584 unsigned value_stride,
585 bool load,
586 bool always_vector)
587 {
588 LLVMBuilderRef builder = ctx->builder;
589 LLVMValueRef vec = NULL;
590 unsigned i;
591
592 if (value_count == 1 && !always_vector) {
593 if (load)
594 return LLVMBuildLoad(builder, values[0], "");
595 return values[0];
596 } else if (!value_count)
597 unreachable("value_count is 0");
598
599 for (i = 0; i < value_count; i++) {
600 LLVMValueRef value = values[i * value_stride];
601 if (load)
602 value = LLVMBuildLoad(builder, value, "");
603
604 if (!i)
605 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
606 LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
607 vec = LLVMBuildInsertElement(builder, vec, value, index, "");
608 }
609 return vec;
610 }
611
612 LLVMValueRef
613 ac_build_gather_values(struct ac_llvm_context *ctx,
614 LLVMValueRef *values,
615 unsigned value_count)
616 {
617 return ac_build_gather_values_extended(ctx, values, value_count, 1, false, false);
618 }
619
620 /* Expand a scalar or vector to <dst_channels x type> by filling the remaining
621 * channels with undef. Extract at most src_channels components from the input.
622 */
623 static LLVMValueRef
624 ac_build_expand(struct ac_llvm_context *ctx,
625 LLVMValueRef value,
626 unsigned src_channels,
627 unsigned dst_channels)
628 {
629 LLVMTypeRef elemtype;
630 LLVMValueRef chan[dst_channels];
631
632 if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMVectorTypeKind) {
633 unsigned vec_size = LLVMGetVectorSize(LLVMTypeOf(value));
634
635 if (src_channels == dst_channels && vec_size == dst_channels)
636 return value;
637
638 src_channels = MIN2(src_channels, vec_size);
639
640 for (unsigned i = 0; i < src_channels; i++)
641 chan[i] = ac_llvm_extract_elem(ctx, value, i);
642
643 elemtype = LLVMGetElementType(LLVMTypeOf(value));
644 } else {
645 if (src_channels) {
646 assert(src_channels == 1);
647 chan[0] = value;
648 }
649 elemtype = LLVMTypeOf(value);
650 }
651
652 for (unsigned i = src_channels; i < dst_channels; i++)
653 chan[i] = LLVMGetUndef(elemtype);
654
655 return ac_build_gather_values(ctx, chan, dst_channels);
656 }
657
658 /* Extract components [start, start + channels) from a vector.
659 */
660 LLVMValueRef
661 ac_extract_components(struct ac_llvm_context *ctx,
662 LLVMValueRef value,
663 unsigned start,
664 unsigned channels)
665 {
666 LLVMValueRef chan[channels];
667
668 for (unsigned i = 0; i < channels; i++)
669 chan[i] = ac_llvm_extract_elem(ctx, value, i + start);
670
671 return ac_build_gather_values(ctx, chan, channels);
672 }
673
674 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
675 * with undef. Extract at most num_channels components from the input.
676 */
677 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
678 LLVMValueRef value,
679 unsigned num_channels)
680 {
681 return ac_build_expand(ctx, value, num_channels, 4);
682 }
683
684 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value)
685 {
686 unsigned type_size = ac_get_type_size(LLVMTypeOf(value));
687 const char *name;
688
689 if (type_size == 2)
690 name = "llvm.rint.f16";
691 else if (type_size == 4)
692 name = "llvm.rint.f32";
693 else
694 name = "llvm.rint.f64";
695
696 return ac_build_intrinsic(ctx, name, LLVMTypeOf(value), &value, 1,
697 AC_FUNC_ATTR_READNONE);
698 }
699
700 LLVMValueRef
701 ac_build_fdiv(struct ac_llvm_context *ctx,
702 LLVMValueRef num,
703 LLVMValueRef den)
704 {
705 unsigned type_size = ac_get_type_size(LLVMTypeOf(den));
706 const char *name;
707
708 if (type_size == 2)
709 name = "llvm.amdgcn.rcp.f16";
710 else if (type_size == 4)
711 name = "llvm.amdgcn.rcp.f32";
712 else
713 name = "llvm.amdgcn.rcp.f64";
714
715 LLVMValueRef rcp = ac_build_intrinsic(ctx, name, LLVMTypeOf(den),
716 &den, 1, AC_FUNC_ATTR_READNONE);
717
718 return LLVMBuildFMul(ctx->builder, num, rcp, "");
719 }
720
721 /* See fast_idiv_by_const.h. */
722 /* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
723 LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
724 LLVMValueRef num,
725 LLVMValueRef multiplier,
726 LLVMValueRef pre_shift,
727 LLVMValueRef post_shift,
728 LLVMValueRef increment)
729 {
730 LLVMBuilderRef builder = ctx->builder;
731
732 num = LLVMBuildLShr(builder, num, pre_shift, "");
733 num = LLVMBuildMul(builder,
734 LLVMBuildZExt(builder, num, ctx->i64, ""),
735 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
736 num = LLVMBuildAdd(builder, num,
737 LLVMBuildZExt(builder, increment, ctx->i64, ""), "");
738 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
739 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
740 return LLVMBuildLShr(builder, num, post_shift, "");
741 }
742
743 /* See fast_idiv_by_const.h. */
744 /* If num != UINT_MAX, this more efficient version can be used. */
745 /* Set: increment = util_fast_udiv_info::increment; */
746 LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
747 LLVMValueRef num,
748 LLVMValueRef multiplier,
749 LLVMValueRef pre_shift,
750 LLVMValueRef post_shift,
751 LLVMValueRef increment)
752 {
753 LLVMBuilderRef builder = ctx->builder;
754
755 num = LLVMBuildLShr(builder, num, pre_shift, "");
756 num = LLVMBuildNUWAdd(builder, num, increment, "");
757 num = LLVMBuildMul(builder,
758 LLVMBuildZExt(builder, num, ctx->i64, ""),
759 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
760 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
761 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
762 return LLVMBuildLShr(builder, num, post_shift, "");
763 }
764
765 /* See fast_idiv_by_const.h. */
766 /* Both operands must fit in 31 bits and the divisor must not be 1. */
767 LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
768 LLVMValueRef num,
769 LLVMValueRef multiplier,
770 LLVMValueRef post_shift)
771 {
772 LLVMBuilderRef builder = ctx->builder;
773
774 num = LLVMBuildMul(builder,
775 LLVMBuildZExt(builder, num, ctx->i64, ""),
776 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
777 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
778 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
779 return LLVMBuildLShr(builder, num, post_shift, "");
780 }
781
782 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
783 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
784 * already multiplied by two. id is the cube face number.
785 */
786 struct cube_selection_coords {
787 LLVMValueRef stc[2];
788 LLVMValueRef ma;
789 LLVMValueRef id;
790 };
791
792 static void
793 build_cube_intrinsic(struct ac_llvm_context *ctx,
794 LLVMValueRef in[3],
795 struct cube_selection_coords *out)
796 {
797 LLVMTypeRef f32 = ctx->f32;
798
799 out->stc[1] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubetc",
800 f32, in, 3, AC_FUNC_ATTR_READNONE);
801 out->stc[0] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubesc",
802 f32, in, 3, AC_FUNC_ATTR_READNONE);
803 out->ma = ac_build_intrinsic(ctx, "llvm.amdgcn.cubema",
804 f32, in, 3, AC_FUNC_ATTR_READNONE);
805 out->id = ac_build_intrinsic(ctx, "llvm.amdgcn.cubeid",
806 f32, in, 3, AC_FUNC_ATTR_READNONE);
807 }
808
809 /**
810 * Build a manual selection sequence for cube face sc/tc coordinates and
811 * major axis vector (multiplied by 2 for consistency) for the given
812 * vec3 \p coords, for the face implied by \p selcoords.
813 *
814 * For the major axis, we always adjust the sign to be in the direction of
815 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
816 * the selcoords major axis.
817 */
818 static void build_cube_select(struct ac_llvm_context *ctx,
819 const struct cube_selection_coords *selcoords,
820 const LLVMValueRef *coords,
821 LLVMValueRef *out_st,
822 LLVMValueRef *out_ma)
823 {
824 LLVMBuilderRef builder = ctx->builder;
825 LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
826 LLVMValueRef is_ma_positive;
827 LLVMValueRef sgn_ma;
828 LLVMValueRef is_ma_z, is_not_ma_z;
829 LLVMValueRef is_ma_y;
830 LLVMValueRef is_ma_x;
831 LLVMValueRef sgn;
832 LLVMValueRef tmp;
833
834 is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
835 selcoords->ma, LLVMConstReal(f32, 0.0), "");
836 sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
837 LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
838
839 is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
840 is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
841 is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
842 LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
843 is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
844
845 /* Select sc */
846 tmp = LLVMBuildSelect(builder, is_ma_x, coords[2], coords[0], "");
847 sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
848 LLVMBuildSelect(builder, is_ma_z, sgn_ma,
849 LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
850 out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
851
852 /* Select tc */
853 tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
854 sgn = LLVMBuildSelect(builder, is_ma_y, sgn_ma,
855 LLVMConstReal(f32, -1.0), "");
856 out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
857
858 /* Select ma */
859 tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
860 LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
861 tmp = ac_build_intrinsic(ctx, "llvm.fabs.f32",
862 ctx->f32, &tmp, 1, AC_FUNC_ATTR_READNONE);
863 *out_ma = LLVMBuildFMul(builder, tmp, LLVMConstReal(f32, 2.0), "");
864 }
865
866 void
867 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
868 bool is_deriv, bool is_array, bool is_lod,
869 LLVMValueRef *coords_arg,
870 LLVMValueRef *derivs_arg)
871 {
872
873 LLVMBuilderRef builder = ctx->builder;
874 struct cube_selection_coords selcoords;
875 LLVMValueRef coords[3];
876 LLVMValueRef invma;
877
878 if (is_array && !is_lod) {
879 LLVMValueRef tmp = ac_build_round(ctx, coords_arg[3]);
880
881 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
882 *
883 * "For Array forms, the array layer used will be
884 *
885 * max(0, min(d−1, floor(layer+0.5)))
886 *
887 * where d is the depth of the texture array and layer
888 * comes from the component indicated in the tables below.
889 * Workaroudn for an issue where the layer is taken from a
890 * helper invocation which happens to fall on a different
891 * layer due to extrapolation."
892 *
893 * GFX8 and earlier attempt to implement this in hardware by
894 * clamping the value of coords[2] = (8 * layer) + face.
895 * Unfortunately, this means that the we end up with the wrong
896 * face when clamping occurs.
897 *
898 * Clamp the layer earlier to work around the issue.
899 */
900 if (ctx->chip_class <= GFX8) {
901 LLVMValueRef ge0;
902 ge0 = LLVMBuildFCmp(builder, LLVMRealOGE, tmp, ctx->f32_0, "");
903 tmp = LLVMBuildSelect(builder, ge0, tmp, ctx->f32_0, "");
904 }
905
906 coords_arg[3] = tmp;
907 }
908
909 build_cube_intrinsic(ctx, coords_arg, &selcoords);
910
911 invma = ac_build_intrinsic(ctx, "llvm.fabs.f32",
912 ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
913 invma = ac_build_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
914
915 for (int i = 0; i < 2; ++i)
916 coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
917
918 coords[2] = selcoords.id;
919
920 if (is_deriv && derivs_arg) {
921 LLVMValueRef derivs[4];
922 int axis;
923
924 /* Convert cube derivatives to 2D derivatives. */
925 for (axis = 0; axis < 2; axis++) {
926 LLVMValueRef deriv_st[2];
927 LLVMValueRef deriv_ma;
928
929 /* Transform the derivative alongside the texture
930 * coordinate. Mathematically, the correct formula is
931 * as follows. Assume we're projecting onto the +Z face
932 * and denote by dx/dh the derivative of the (original)
933 * X texture coordinate with respect to horizontal
934 * window coordinates. The projection onto the +Z face
935 * plane is:
936 *
937 * f(x,z) = x/z
938 *
939 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
940 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
941 *
942 * This motivatives the implementation below.
943 *
944 * Whether this actually gives the expected results for
945 * apps that might feed in derivatives obtained via
946 * finite differences is anyone's guess. The OpenGL spec
947 * seems awfully quiet about how textureGrad for cube
948 * maps should be handled.
949 */
950 build_cube_select(ctx, &selcoords, &derivs_arg[axis * 3],
951 deriv_st, &deriv_ma);
952
953 deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
954
955 for (int i = 0; i < 2; ++i)
956 derivs[axis * 2 + i] =
957 LLVMBuildFSub(builder,
958 LLVMBuildFMul(builder, deriv_st[i], invma, ""),
959 LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
960 }
961
962 memcpy(derivs_arg, derivs, sizeof(derivs));
963 }
964
965 /* Shift the texture coordinate. This must be applied after the
966 * derivative calculation.
967 */
968 for (int i = 0; i < 2; ++i)
969 coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
970
971 if (is_array) {
972 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
973 /* coords_arg.w component - array_index for cube arrays */
974 coords[2] = ac_build_fmad(ctx, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), coords[2]);
975 }
976
977 memcpy(coords_arg, coords, sizeof(coords));
978 }
979
980
981 LLVMValueRef
982 ac_build_fs_interp(struct ac_llvm_context *ctx,
983 LLVMValueRef llvm_chan,
984 LLVMValueRef attr_number,
985 LLVMValueRef params,
986 LLVMValueRef i,
987 LLVMValueRef j)
988 {
989 LLVMValueRef args[5];
990 LLVMValueRef p1;
991
992 args[0] = i;
993 args[1] = llvm_chan;
994 args[2] = attr_number;
995 args[3] = params;
996
997 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1",
998 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
999
1000 args[0] = p1;
1001 args[1] = j;
1002 args[2] = llvm_chan;
1003 args[3] = attr_number;
1004 args[4] = params;
1005
1006 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2",
1007 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
1008 }
1009
1010 LLVMValueRef
1011 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
1012 LLVMValueRef llvm_chan,
1013 LLVMValueRef attr_number,
1014 LLVMValueRef params,
1015 LLVMValueRef i,
1016 LLVMValueRef j)
1017 {
1018 LLVMValueRef args[6];
1019 LLVMValueRef p1;
1020
1021 args[0] = i;
1022 args[1] = llvm_chan;
1023 args[2] = attr_number;
1024 args[3] = ctx->i1false;
1025 args[4] = params;
1026
1027 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1.f16",
1028 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
1029
1030 args[0] = p1;
1031 args[1] = j;
1032 args[2] = llvm_chan;
1033 args[3] = attr_number;
1034 args[4] = ctx->i1false;
1035 args[5] = params;
1036
1037 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2.f16",
1038 ctx->f16, args, 6, AC_FUNC_ATTR_READNONE);
1039 }
1040
1041 LLVMValueRef
1042 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
1043 LLVMValueRef parameter,
1044 LLVMValueRef llvm_chan,
1045 LLVMValueRef attr_number,
1046 LLVMValueRef params)
1047 {
1048 LLVMValueRef args[4];
1049
1050 args[0] = parameter;
1051 args[1] = llvm_chan;
1052 args[2] = attr_number;
1053 args[3] = params;
1054
1055 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.mov",
1056 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
1057 }
1058
1059 LLVMValueRef
1060 ac_build_gep_ptr(struct ac_llvm_context *ctx,
1061 LLVMValueRef base_ptr,
1062 LLVMValueRef index)
1063 {
1064 return LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
1065 }
1066
1067 LLVMValueRef
1068 ac_build_gep0(struct ac_llvm_context *ctx,
1069 LLVMValueRef base_ptr,
1070 LLVMValueRef index)
1071 {
1072 LLVMValueRef indices[2] = {
1073 ctx->i32_0,
1074 index,
1075 };
1076 return LLVMBuildGEP(ctx->builder, base_ptr, indices, 2, "");
1077 }
1078
1079 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
1080 LLVMValueRef index)
1081 {
1082 return LLVMBuildPointerCast(ctx->builder,
1083 LLVMBuildGEP(ctx->builder, ptr, &index, 1, ""),
1084 LLVMTypeOf(ptr), "");
1085 }
1086
1087 void
1088 ac_build_indexed_store(struct ac_llvm_context *ctx,
1089 LLVMValueRef base_ptr, LLVMValueRef index,
1090 LLVMValueRef value)
1091 {
1092 LLVMBuildStore(ctx->builder, value,
1093 ac_build_gep0(ctx, base_ptr, index));
1094 }
1095
1096 /**
1097 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
1098 * It's equivalent to doing a load from &base_ptr[index].
1099 *
1100 * \param base_ptr Where the array starts.
1101 * \param index The element index into the array.
1102 * \param uniform Whether the base_ptr and index can be assumed to be
1103 * dynamically uniform (i.e. load to an SGPR)
1104 * \param invariant Whether the load is invariant (no other opcodes affect it)
1105 * \param no_unsigned_wraparound
1106 * For all possible re-associations and re-distributions of an expression
1107 * "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
1108 * without inbounds in base_ptr), this parameter is true if "addr + offset"
1109 * does not result in an unsigned integer wraparound. This is used for
1110 * optimal code generation of 32-bit pointer arithmetic.
1111 *
1112 * For example, a 32-bit immediate offset that causes a 32-bit unsigned
1113 * integer wraparound can't be an imm offset in s_load_dword, because
1114 * the instruction performs "addr + offset" in 64 bits.
1115 *
1116 * Expected usage for bindless textures by chaining GEPs:
1117 * // possible unsigned wraparound, don't use InBounds:
1118 * ptr1 = LLVMBuildGEP(base_ptr, index);
1119 * image = load(ptr1); // becomes "s_load ptr1, 0"
1120 *
1121 * ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1122 * sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1123 */
1124 static LLVMValueRef
1125 ac_build_load_custom(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1126 LLVMValueRef index, bool uniform, bool invariant,
1127 bool no_unsigned_wraparound)
1128 {
1129 LLVMValueRef pointer, result;
1130
1131 if (no_unsigned_wraparound &&
1132 LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr)) == AC_ADDR_SPACE_CONST_32BIT)
1133 pointer = LLVMBuildInBoundsGEP(ctx->builder, base_ptr, &index, 1, "");
1134 else
1135 pointer = LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
1136
1137 if (uniform)
1138 LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
1139 result = LLVMBuildLoad(ctx->builder, pointer, "");
1140 if (invariant)
1141 LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
1142 return result;
1143 }
1144
1145 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1146 LLVMValueRef index)
1147 {
1148 return ac_build_load_custom(ctx, base_ptr, index, false, false, false);
1149 }
1150
1151 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
1152 LLVMValueRef base_ptr, LLVMValueRef index)
1153 {
1154 return ac_build_load_custom(ctx, base_ptr, index, false, true, false);
1155 }
1156
1157 /* This assumes that there is no unsigned integer wraparound during the address
1158 * computation, excluding all GEPs within base_ptr. */
1159 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
1160 LLVMValueRef base_ptr, LLVMValueRef index)
1161 {
1162 return ac_build_load_custom(ctx, base_ptr, index, true, true, true);
1163 }
1164
1165 /* See ac_build_load_custom() documentation. */
1166 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
1167 LLVMValueRef base_ptr, LLVMValueRef index)
1168 {
1169 return ac_build_load_custom(ctx, base_ptr, index, true, true, false);
1170 }
1171
1172 static unsigned get_load_cache_policy(struct ac_llvm_context *ctx,
1173 unsigned cache_policy)
1174 {
1175 return cache_policy |
1176 (ctx->chip_class >= GFX10 && cache_policy & ac_glc ? ac_dlc : 0);
1177 }
1178
1179 static void
1180 ac_build_buffer_store_common(struct ac_llvm_context *ctx,
1181 LLVMValueRef rsrc,
1182 LLVMValueRef data,
1183 LLVMValueRef vindex,
1184 LLVMValueRef voffset,
1185 LLVMValueRef soffset,
1186 unsigned num_channels,
1187 LLVMTypeRef return_channel_type,
1188 unsigned cache_policy,
1189 bool use_format,
1190 bool structurized)
1191 {
1192 LLVMValueRef args[6];
1193 int idx = 0;
1194 args[idx++] = data;
1195 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1196 if (structurized)
1197 args[idx++] = vindex ? vindex : ctx->i32_0;
1198 args[idx++] = voffset ? voffset : ctx->i32_0;
1199 args[idx++] = soffset ? soffset : ctx->i32_0;
1200 args[idx++] = LLVMConstInt(ctx->i32, cache_policy, 0);
1201 unsigned func = !ac_has_vec3_support(ctx->chip_class, use_format) && num_channels == 3 ? 4 : num_channels;
1202 const char *indexing_kind = structurized ? "struct" : "raw";
1203 char name[256], type_name[8];
1204
1205 LLVMTypeRef type = func > 1 ? LLVMVectorType(return_channel_type, func) : return_channel_type;
1206 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1207
1208 if (use_format) {
1209 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.format.%s",
1210 indexing_kind, type_name);
1211 } else {
1212 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.%s",
1213 indexing_kind, type_name);
1214 }
1215
1216 ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
1217 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
1218 }
1219
1220 void
1221 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
1222 LLVMValueRef rsrc,
1223 LLVMValueRef data,
1224 LLVMValueRef vindex,
1225 LLVMValueRef voffset,
1226 unsigned num_channels,
1227 unsigned cache_policy)
1228 {
1229 ac_build_buffer_store_common(ctx, rsrc, data, vindex,
1230 voffset, NULL, num_channels,
1231 ctx->f32, cache_policy,
1232 true, true);
1233 }
1234
1235 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1236 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1237 * or v4i32 (num_channels=3,4).
1238 */
1239 void
1240 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
1241 LLVMValueRef rsrc,
1242 LLVMValueRef vdata,
1243 unsigned num_channels,
1244 LLVMValueRef voffset,
1245 LLVMValueRef soffset,
1246 unsigned inst_offset,
1247 unsigned cache_policy)
1248 {
1249 /* Split 3 channel stores, because only LLVM 9+ support 3-channel
1250 * intrinsics. */
1251 if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false)) {
1252 LLVMValueRef v[3], v01;
1253
1254 for (int i = 0; i < 3; i++) {
1255 v[i] = LLVMBuildExtractElement(ctx->builder, vdata,
1256 LLVMConstInt(ctx->i32, i, 0), "");
1257 }
1258 v01 = ac_build_gather_values(ctx, v, 2);
1259
1260 ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset,
1261 soffset, inst_offset, cache_policy);
1262 ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset,
1263 soffset, inst_offset + 8,
1264 cache_policy);
1265 return;
1266 }
1267
1268 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1269 * (voffset is swizzled, but soffset isn't swizzled).
1270 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1271 */
1272 if (!(cache_policy & ac_swizzled)) {
1273 LLVMValueRef offset = soffset;
1274
1275 if (inst_offset)
1276 offset = LLVMBuildAdd(ctx->builder, offset,
1277 LLVMConstInt(ctx->i32, inst_offset, 0), "");
1278
1279 ac_build_buffer_store_common(ctx, rsrc, ac_to_float(ctx, vdata),
1280 ctx->i32_0, voffset, offset,
1281 num_channels, ctx->f32,
1282 cache_policy, false, false);
1283 return;
1284 }
1285
1286 static const unsigned dfmts[] = {
1287 V_008F0C_BUF_DATA_FORMAT_32,
1288 V_008F0C_BUF_DATA_FORMAT_32_32,
1289 V_008F0C_BUF_DATA_FORMAT_32_32_32,
1290 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1291 };
1292 unsigned dfmt = dfmts[num_channels - 1];
1293 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1294 LLVMValueRef immoffset = LLVMConstInt(ctx->i32, inst_offset, 0);
1295
1296 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1297 immoffset, num_channels, dfmt, nfmt, cache_policy);
1298 }
1299
1300 static LLVMValueRef
1301 ac_build_buffer_load_common(struct ac_llvm_context *ctx,
1302 LLVMValueRef rsrc,
1303 LLVMValueRef vindex,
1304 LLVMValueRef voffset,
1305 LLVMValueRef soffset,
1306 unsigned num_channels,
1307 LLVMTypeRef channel_type,
1308 unsigned cache_policy,
1309 bool can_speculate,
1310 bool use_format,
1311 bool structurized)
1312 {
1313 LLVMValueRef args[5];
1314 int idx = 0;
1315 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1316 if (structurized)
1317 args[idx++] = vindex ? vindex : ctx->i32_0;
1318 args[idx++] = voffset ? voffset : ctx->i32_0;
1319 args[idx++] = soffset ? soffset : ctx->i32_0;
1320 args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
1321 unsigned func = !ac_has_vec3_support(ctx->chip_class, use_format) && num_channels == 3 ? 4 : num_channels;
1322 const char *indexing_kind = structurized ? "struct" : "raw";
1323 char name[256], type_name[8];
1324
1325 LLVMTypeRef type = func > 1 ? LLVMVectorType(channel_type, func) : channel_type;
1326 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1327
1328 if (use_format) {
1329 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.format.%s",
1330 indexing_kind, type_name);
1331 } else {
1332 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.%s",
1333 indexing_kind, type_name);
1334 }
1335
1336 return ac_build_intrinsic(ctx, name, type, args, idx,
1337 ac_get_load_intr_attribs(can_speculate));
1338 }
1339
1340 LLVMValueRef
1341 ac_build_buffer_load(struct ac_llvm_context *ctx,
1342 LLVMValueRef rsrc,
1343 int num_channels,
1344 LLVMValueRef vindex,
1345 LLVMValueRef voffset,
1346 LLVMValueRef soffset,
1347 unsigned inst_offset,
1348 unsigned cache_policy,
1349 bool can_speculate,
1350 bool allow_smem)
1351 {
1352 LLVMValueRef offset = LLVMConstInt(ctx->i32, inst_offset, 0);
1353 if (voffset)
1354 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1355 if (soffset)
1356 offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
1357
1358 if (allow_smem && !(cache_policy & ac_slc) &&
1359 (!(cache_policy & ac_glc) || ctx->chip_class >= GFX8)) {
1360 assert(vindex == NULL);
1361
1362 LLVMValueRef result[8];
1363
1364 for (int i = 0; i < num_channels; i++) {
1365 if (i) {
1366 offset = LLVMBuildAdd(ctx->builder, offset,
1367 LLVMConstInt(ctx->i32, 4, 0), "");
1368 }
1369 LLVMValueRef args[3] = {
1370 rsrc,
1371 offset,
1372 LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0),
1373 };
1374 result[i] = ac_build_intrinsic(ctx,
1375 "llvm.amdgcn.s.buffer.load.f32",
1376 ctx->f32, args, 3,
1377 AC_FUNC_ATTR_READNONE);
1378 }
1379 if (num_channels == 1)
1380 return result[0];
1381
1382 if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false))
1383 result[num_channels++] = LLVMGetUndef(ctx->f32);
1384 return ac_build_gather_values(ctx, result, num_channels);
1385 }
1386
1387 return ac_build_buffer_load_common(ctx, rsrc, vindex,
1388 offset, ctx->i32_0,
1389 num_channels, ctx->f32,
1390 cache_policy,
1391 can_speculate, false, false);
1392 }
1393
1394 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
1395 LLVMValueRef rsrc,
1396 LLVMValueRef vindex,
1397 LLVMValueRef voffset,
1398 unsigned num_channels,
1399 unsigned cache_policy,
1400 bool can_speculate)
1401 {
1402 return ac_build_buffer_load_common(ctx, rsrc, vindex, voffset,
1403 ctx->i32_0, num_channels, ctx->f32,
1404 cache_policy, can_speculate,
1405 true, true);
1406 }
1407
1408 static LLVMValueRef
1409 ac_build_tbuffer_load(struct ac_llvm_context *ctx,
1410 LLVMValueRef rsrc,
1411 LLVMValueRef vindex,
1412 LLVMValueRef voffset,
1413 LLVMValueRef soffset,
1414 LLVMValueRef immoffset,
1415 unsigned num_channels,
1416 unsigned dfmt,
1417 unsigned nfmt,
1418 unsigned cache_policy,
1419 bool can_speculate,
1420 bool structurized)
1421 {
1422 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1423
1424 LLVMValueRef args[6];
1425 int idx = 0;
1426 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1427 if (structurized)
1428 args[idx++] = vindex ? vindex : ctx->i32_0;
1429 args[idx++] = voffset ? voffset : ctx->i32_0;
1430 args[idx++] = soffset ? soffset : ctx->i32_0;
1431 args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx->chip_class, dfmt, nfmt), 0);
1432 args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
1433 unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
1434 const char *indexing_kind = structurized ? "struct" : "raw";
1435 char name[256], type_name[8];
1436
1437 LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
1438 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1439
1440 snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.load.%s",
1441 indexing_kind, type_name);
1442
1443 return ac_build_intrinsic(ctx, name, type, args, idx,
1444 ac_get_load_intr_attribs(can_speculate));
1445 }
1446
1447 LLVMValueRef
1448 ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
1449 LLVMValueRef rsrc,
1450 LLVMValueRef vindex,
1451 LLVMValueRef voffset,
1452 LLVMValueRef soffset,
1453 LLVMValueRef immoffset,
1454 unsigned num_channels,
1455 unsigned dfmt,
1456 unsigned nfmt,
1457 unsigned cache_policy,
1458 bool can_speculate)
1459 {
1460 return ac_build_tbuffer_load(ctx, rsrc, vindex, voffset, soffset,
1461 immoffset, num_channels, dfmt, nfmt,
1462 cache_policy, can_speculate, true);
1463 }
1464
1465 LLVMValueRef
1466 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
1467 LLVMValueRef rsrc,
1468 LLVMValueRef voffset,
1469 LLVMValueRef soffset,
1470 LLVMValueRef immoffset,
1471 unsigned num_channels,
1472 unsigned dfmt,
1473 unsigned nfmt,
1474 unsigned cache_policy,
1475 bool can_speculate)
1476 {
1477 return ac_build_tbuffer_load(ctx, rsrc, NULL, voffset, soffset,
1478 immoffset, num_channels, dfmt, nfmt,
1479 cache_policy, can_speculate, false);
1480 }
1481
1482 LLVMValueRef
1483 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
1484 LLVMValueRef rsrc,
1485 LLVMValueRef voffset,
1486 LLVMValueRef soffset,
1487 LLVMValueRef immoffset,
1488 unsigned cache_policy)
1489 {
1490 LLVMValueRef res;
1491
1492 if (LLVM_VERSION_MAJOR >= 9) {
1493 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1494
1495 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1496 res = ac_build_buffer_load_common(ctx, rsrc, NULL,
1497 voffset, soffset,
1498 1, ctx->i16, cache_policy,
1499 false, false, false);
1500 } else {
1501 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
1502 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1503
1504 res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
1505 immoffset, 1, dfmt, nfmt, cache_policy,
1506 false);
1507
1508 res = LLVMBuildTrunc(ctx->builder, res, ctx->i16, "");
1509 }
1510
1511 return res;
1512 }
1513
1514 LLVMValueRef
1515 ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
1516 LLVMValueRef rsrc,
1517 LLVMValueRef voffset,
1518 LLVMValueRef soffset,
1519 LLVMValueRef immoffset,
1520 unsigned cache_policy)
1521 {
1522 LLVMValueRef res;
1523
1524 if (LLVM_VERSION_MAJOR >= 9) {
1525 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1526
1527 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1528 res = ac_build_buffer_load_common(ctx, rsrc, NULL,
1529 voffset, soffset,
1530 1, ctx->i8, cache_policy,
1531 false, false, false);
1532 } else {
1533 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
1534 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1535
1536 res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
1537 immoffset, 1, dfmt, nfmt, cache_policy,
1538 false);
1539
1540 res = LLVMBuildTrunc(ctx->builder, res, ctx->i8, "");
1541 }
1542
1543 return res;
1544 }
1545
1546 /**
1547 * Convert an 11- or 10-bit unsigned floating point number to an f32.
1548 *
1549 * The input exponent is expected to be biased analogous to IEEE-754, i.e. by
1550 * 2^(exp_bits-1) - 1 (as defined in OpenGL and other graphics APIs).
1551 */
1552 static LLVMValueRef
1553 ac_ufN_to_float(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned exp_bits, unsigned mant_bits)
1554 {
1555 assert(LLVMTypeOf(src) == ctx->i32);
1556
1557 LLVMValueRef tmp;
1558 LLVMValueRef mantissa;
1559 mantissa = LLVMBuildAnd(ctx->builder, src, LLVMConstInt(ctx->i32, (1 << mant_bits) - 1, false), "");
1560
1561 /* Converting normal numbers is just a shift + correcting the exponent bias */
1562 unsigned normal_shift = 23 - mant_bits;
1563 unsigned bias_shift = 127 - ((1 << (exp_bits - 1)) - 1);
1564 LLVMValueRef shifted, normal;
1565
1566 shifted = LLVMBuildShl(ctx->builder, src, LLVMConstInt(ctx->i32, normal_shift, false), "");
1567 normal = LLVMBuildAdd(ctx->builder, shifted, LLVMConstInt(ctx->i32, bias_shift << 23, false), "");
1568
1569 /* Converting nan/inf numbers is the same, but with a different exponent update */
1570 LLVMValueRef naninf;
1571 naninf = LLVMBuildOr(ctx->builder, normal, LLVMConstInt(ctx->i32, 0xff << 23, false), "");
1572
1573 /* Converting denormals is the complex case: determine the leading zeros of the
1574 * mantissa to obtain the correct shift for the mantissa and exponent correction.
1575 */
1576 LLVMValueRef denormal;
1577 LLVMValueRef params[2] = {
1578 mantissa,
1579 ctx->i1true, /* result can be undef when arg is 0 */
1580 };
1581 LLVMValueRef ctlz = ac_build_intrinsic(ctx, "llvm.ctlz.i32", ctx->i32,
1582 params, 2, AC_FUNC_ATTR_READNONE);
1583
1584 /* Shift such that the leading 1 ends up as the LSB of the exponent field. */
1585 tmp = LLVMBuildSub(ctx->builder, ctlz, LLVMConstInt(ctx->i32, 8, false), "");
1586 denormal = LLVMBuildShl(ctx->builder, mantissa, tmp, "");
1587
1588 unsigned denormal_exp = bias_shift + (32 - mant_bits) - 1;
1589 tmp = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, denormal_exp, false), ctlz, "");
1590 tmp = LLVMBuildShl(ctx->builder, tmp, LLVMConstInt(ctx->i32, 23, false), "");
1591 denormal = LLVMBuildAdd(ctx->builder, denormal, tmp, "");
1592
1593 /* Select the final result. */
1594 LLVMValueRef result;
1595
1596 tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src,
1597 LLVMConstInt(ctx->i32, ((1 << exp_bits) - 1) << mant_bits, false), "");
1598 result = LLVMBuildSelect(ctx->builder, tmp, naninf, normal, "");
1599
1600 tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src,
1601 LLVMConstInt(ctx->i32, 1 << mant_bits, false), "");
1602 result = LLVMBuildSelect(ctx->builder, tmp, result, denormal, "");
1603
1604 tmp = LLVMBuildICmp(ctx->builder, LLVMIntNE, src, ctx->i32_0, "");
1605 result = LLVMBuildSelect(ctx->builder, tmp, result, ctx->i32_0, "");
1606
1607 return ac_to_float(ctx, result);
1608 }
1609
1610 /**
1611 * Generate a fully general open coded buffer format fetch with all required
1612 * fixups suitable for vertex fetch, using non-format buffer loads.
1613 *
1614 * Some combinations of argument values have special interpretations:
1615 * - size = 8 bytes, format = fixed indicates PIPE_FORMAT_R11G11B10_FLOAT
1616 * - size = 8 bytes, format != {float,fixed} indicates a 2_10_10_10 data format
1617 *
1618 * \param log_size log(size of channel in bytes)
1619 * \param num_channels number of channels (1 to 4)
1620 * \param format AC_FETCH_FORMAT_xxx value
1621 * \param reverse whether XYZ channels are reversed
1622 * \param known_aligned whether the source is known to be aligned to hardware's
1623 * effective element size for loading the given format
1624 * (note: this means dword alignment for 8_8_8_8, 16_16, etc.)
1625 * \param rsrc buffer resource descriptor
1626 * \return the resulting vector of floats or integers bitcast to <4 x i32>
1627 */
1628 LLVMValueRef
1629 ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
1630 unsigned log_size,
1631 unsigned num_channels,
1632 unsigned format,
1633 bool reverse,
1634 bool known_aligned,
1635 LLVMValueRef rsrc,
1636 LLVMValueRef vindex,
1637 LLVMValueRef voffset,
1638 LLVMValueRef soffset,
1639 unsigned cache_policy,
1640 bool can_speculate)
1641 {
1642 LLVMValueRef tmp;
1643 unsigned load_log_size = log_size;
1644 unsigned load_num_channels = num_channels;
1645 if (log_size == 3) {
1646 load_log_size = 2;
1647 if (format == AC_FETCH_FORMAT_FLOAT) {
1648 load_num_channels = 2 * num_channels;
1649 } else {
1650 load_num_channels = 1; /* 10_11_11 or 2_10_10_10 */
1651 }
1652 }
1653
1654 int log_recombine = 0;
1655 if (ctx->chip_class == GFX6 && !known_aligned) {
1656 /* Avoid alignment restrictions by loading one byte at a time. */
1657 load_num_channels <<= load_log_size;
1658 log_recombine = load_log_size;
1659 load_log_size = 0;
1660 } else if (load_num_channels == 2 || load_num_channels == 4) {
1661 log_recombine = -util_logbase2(load_num_channels);
1662 load_num_channels = 1;
1663 load_log_size += -log_recombine;
1664 }
1665
1666 assert(load_log_size >= 2 || LLVM_VERSION_MAJOR >= 9);
1667
1668 LLVMValueRef loads[32]; /* up to 32 bytes */
1669 for (unsigned i = 0; i < load_num_channels; ++i) {
1670 tmp = LLVMBuildAdd(ctx->builder, soffset,
1671 LLVMConstInt(ctx->i32, i << load_log_size, false), "");
1672 LLVMTypeRef channel_type = load_log_size == 0 ? ctx->i8 :
1673 load_log_size == 1 ? ctx->i16 : ctx->i32;
1674 unsigned num_channels = 1 << (MAX2(load_log_size, 2) - 2);
1675 loads[i] = ac_build_buffer_load_common(
1676 ctx, rsrc, vindex, voffset, tmp,
1677 num_channels, channel_type, cache_policy,
1678 can_speculate, false, true);
1679 if (load_log_size >= 2)
1680 loads[i] = ac_to_integer(ctx, loads[i]);
1681 }
1682
1683 if (log_recombine > 0) {
1684 /* Recombine bytes if necessary (GFX6 only) */
1685 LLVMTypeRef dst_type = log_recombine == 2 ? ctx->i32 : ctx->i16;
1686
1687 for (unsigned src = 0, dst = 0; src < load_num_channels; ++dst) {
1688 LLVMValueRef accum = NULL;
1689 for (unsigned i = 0; i < (1 << log_recombine); ++i, ++src) {
1690 tmp = LLVMBuildZExt(ctx->builder, loads[src], dst_type, "");
1691 if (i == 0) {
1692 accum = tmp;
1693 } else {
1694 tmp = LLVMBuildShl(ctx->builder, tmp,
1695 LLVMConstInt(dst_type, 8 * i, false), "");
1696 accum = LLVMBuildOr(ctx->builder, accum, tmp, "");
1697 }
1698 }
1699 loads[dst] = accum;
1700 }
1701 } else if (log_recombine < 0) {
1702 /* Split vectors of dwords */
1703 if (load_log_size > 2) {
1704 assert(load_num_channels == 1);
1705 LLVMValueRef loaded = loads[0];
1706 unsigned log_split = load_log_size - 2;
1707 log_recombine += log_split;
1708 load_num_channels = 1 << log_split;
1709 load_log_size = 2;
1710 for (unsigned i = 0; i < load_num_channels; ++i) {
1711 tmp = LLVMConstInt(ctx->i32, i, false);
1712 loads[i] = LLVMBuildExtractElement(ctx->builder, loaded, tmp, "");
1713 }
1714 }
1715
1716 /* Further split dwords and shorts if required */
1717 if (log_recombine < 0) {
1718 for (unsigned src = load_num_channels,
1719 dst = load_num_channels << -log_recombine;
1720 src > 0; --src) {
1721 unsigned dst_bits = 1 << (3 + load_log_size + log_recombine);
1722 LLVMTypeRef dst_type = LLVMIntTypeInContext(ctx->context, dst_bits);
1723 LLVMValueRef loaded = loads[src - 1];
1724 LLVMTypeRef loaded_type = LLVMTypeOf(loaded);
1725 for (unsigned i = 1 << -log_recombine; i > 0; --i, --dst) {
1726 tmp = LLVMConstInt(loaded_type, dst_bits * (i - 1), false);
1727 tmp = LLVMBuildLShr(ctx->builder, loaded, tmp, "");
1728 loads[dst - 1] = LLVMBuildTrunc(ctx->builder, tmp, dst_type, "");
1729 }
1730 }
1731 }
1732 }
1733
1734 if (log_size == 3) {
1735 if (format == AC_FETCH_FORMAT_FLOAT) {
1736 for (unsigned i = 0; i < num_channels; ++i) {
1737 tmp = ac_build_gather_values(ctx, &loads[2 * i], 2);
1738 loads[i] = LLVMBuildBitCast(ctx->builder, tmp, ctx->f64, "");
1739 }
1740 } else if (format == AC_FETCH_FORMAT_FIXED) {
1741 /* 10_11_11_FLOAT */
1742 LLVMValueRef data = loads[0];
1743 LLVMValueRef i32_2047 = LLVMConstInt(ctx->i32, 2047, false);
1744 LLVMValueRef r = LLVMBuildAnd(ctx->builder, data, i32_2047, "");
1745 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 11, false), "");
1746 LLVMValueRef g = LLVMBuildAnd(ctx->builder, tmp, i32_2047, "");
1747 LLVMValueRef b = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 22, false), "");
1748
1749 loads[0] = ac_to_integer(ctx, ac_ufN_to_float(ctx, r, 5, 6));
1750 loads[1] = ac_to_integer(ctx, ac_ufN_to_float(ctx, g, 5, 6));
1751 loads[2] = ac_to_integer(ctx, ac_ufN_to_float(ctx, b, 5, 5));
1752
1753 num_channels = 3;
1754 log_size = 2;
1755 format = AC_FETCH_FORMAT_FLOAT;
1756 } else {
1757 /* 2_10_10_10 data formats */
1758 LLVMValueRef data = loads[0];
1759 LLVMTypeRef i10 = LLVMIntTypeInContext(ctx->context, 10);
1760 LLVMTypeRef i2 = LLVMIntTypeInContext(ctx->context, 2);
1761 loads[0] = LLVMBuildTrunc(ctx->builder, data, i10, "");
1762 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 10, false), "");
1763 loads[1] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
1764 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 20, false), "");
1765 loads[2] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
1766 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 30, false), "");
1767 loads[3] = LLVMBuildTrunc(ctx->builder, tmp, i2, "");
1768
1769 num_channels = 4;
1770 }
1771 }
1772
1773 if (format == AC_FETCH_FORMAT_FLOAT) {
1774 if (log_size != 2) {
1775 for (unsigned chan = 0; chan < num_channels; ++chan) {
1776 tmp = ac_to_float(ctx, loads[chan]);
1777 if (log_size == 3)
1778 tmp = LLVMBuildFPTrunc(ctx->builder, tmp, ctx->f32, "");
1779 else if (log_size == 1)
1780 tmp = LLVMBuildFPExt(ctx->builder, tmp, ctx->f32, "");
1781 loads[chan] = ac_to_integer(ctx, tmp);
1782 }
1783 }
1784 } else if (format == AC_FETCH_FORMAT_UINT) {
1785 if (log_size != 2) {
1786 for (unsigned chan = 0; chan < num_channels; ++chan)
1787 loads[chan] = LLVMBuildZExt(ctx->builder, loads[chan], ctx->i32, "");
1788 }
1789 } else if (format == AC_FETCH_FORMAT_SINT) {
1790 if (log_size != 2) {
1791 for (unsigned chan = 0; chan < num_channels; ++chan)
1792 loads[chan] = LLVMBuildSExt(ctx->builder, loads[chan], ctx->i32, "");
1793 }
1794 } else {
1795 bool unsign = format == AC_FETCH_FORMAT_UNORM ||
1796 format == AC_FETCH_FORMAT_USCALED ||
1797 format == AC_FETCH_FORMAT_UINT;
1798
1799 for (unsigned chan = 0; chan < num_channels; ++chan) {
1800 if (unsign) {
1801 tmp = LLVMBuildUIToFP(ctx->builder, loads[chan], ctx->f32, "");
1802 } else {
1803 tmp = LLVMBuildSIToFP(ctx->builder, loads[chan], ctx->f32, "");
1804 }
1805
1806 LLVMValueRef scale = NULL;
1807 if (format == AC_FETCH_FORMAT_FIXED) {
1808 assert(log_size == 2);
1809 scale = LLVMConstReal(ctx->f32, 1.0 / 0x10000);
1810 } else if (format == AC_FETCH_FORMAT_UNORM) {
1811 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
1812 scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << bits) - 1));
1813 } else if (format == AC_FETCH_FORMAT_SNORM) {
1814 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
1815 scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << (bits - 1)) - 1));
1816 }
1817 if (scale)
1818 tmp = LLVMBuildFMul(ctx->builder, tmp, scale, "");
1819
1820 if (format == AC_FETCH_FORMAT_SNORM) {
1821 /* Clamp to [-1, 1] */
1822 LLVMValueRef neg_one = LLVMConstReal(ctx->f32, -1.0);
1823 LLVMValueRef clamp =
1824 LLVMBuildFCmp(ctx->builder, LLVMRealULT, tmp, neg_one, "");
1825 tmp = LLVMBuildSelect(ctx->builder, clamp, neg_one, tmp, "");
1826 }
1827
1828 loads[chan] = ac_to_integer(ctx, tmp);
1829 }
1830 }
1831
1832 while (num_channels < 4) {
1833 if (format == AC_FETCH_FORMAT_UINT || format == AC_FETCH_FORMAT_SINT) {
1834 loads[num_channels] = num_channels == 3 ? ctx->i32_1 : ctx->i32_0;
1835 } else {
1836 loads[num_channels] = ac_to_integer(ctx, num_channels == 3 ? ctx->f32_1 : ctx->f32_0);
1837 }
1838 num_channels++;
1839 }
1840
1841 if (reverse) {
1842 tmp = loads[0];
1843 loads[0] = loads[2];
1844 loads[2] = tmp;
1845 }
1846
1847 return ac_build_gather_values(ctx, loads, 4);
1848 }
1849
1850 static void
1851 ac_build_tbuffer_store(struct ac_llvm_context *ctx,
1852 LLVMValueRef rsrc,
1853 LLVMValueRef vdata,
1854 LLVMValueRef vindex,
1855 LLVMValueRef voffset,
1856 LLVMValueRef soffset,
1857 LLVMValueRef immoffset,
1858 unsigned num_channels,
1859 unsigned dfmt,
1860 unsigned nfmt,
1861 unsigned cache_policy,
1862 bool structurized)
1863 {
1864 voffset = LLVMBuildAdd(ctx->builder, voffset ? voffset : ctx->i32_0,
1865 immoffset, "");
1866
1867 LLVMValueRef args[7];
1868 int idx = 0;
1869 args[idx++] = vdata;
1870 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1871 if (structurized)
1872 args[idx++] = vindex ? vindex : ctx->i32_0;
1873 args[idx++] = voffset ? voffset : ctx->i32_0;
1874 args[idx++] = soffset ? soffset : ctx->i32_0;
1875 args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx->chip_class, dfmt, nfmt), 0);
1876 args[idx++] = LLVMConstInt(ctx->i32, cache_policy, 0);
1877 unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
1878 const char *indexing_kind = structurized ? "struct" : "raw";
1879 char name[256], type_name[8];
1880
1881 LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
1882 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1883
1884 snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.store.%s",
1885 indexing_kind, type_name);
1886
1887 ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
1888 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
1889 }
1890
1891 void
1892 ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
1893 LLVMValueRef rsrc,
1894 LLVMValueRef vdata,
1895 LLVMValueRef vindex,
1896 LLVMValueRef voffset,
1897 LLVMValueRef soffset,
1898 LLVMValueRef immoffset,
1899 unsigned num_channels,
1900 unsigned dfmt,
1901 unsigned nfmt,
1902 unsigned cache_policy)
1903 {
1904 ac_build_tbuffer_store(ctx, rsrc, vdata, vindex, voffset, soffset,
1905 immoffset, num_channels, dfmt, nfmt, cache_policy,
1906 true);
1907 }
1908
1909 void
1910 ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
1911 LLVMValueRef rsrc,
1912 LLVMValueRef vdata,
1913 LLVMValueRef voffset,
1914 LLVMValueRef soffset,
1915 LLVMValueRef immoffset,
1916 unsigned num_channels,
1917 unsigned dfmt,
1918 unsigned nfmt,
1919 unsigned cache_policy)
1920 {
1921 ac_build_tbuffer_store(ctx, rsrc, vdata, NULL, voffset, soffset,
1922 immoffset, num_channels, dfmt, nfmt, cache_policy,
1923 false);
1924 }
1925
1926 void
1927 ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
1928 LLVMValueRef rsrc,
1929 LLVMValueRef vdata,
1930 LLVMValueRef voffset,
1931 LLVMValueRef soffset,
1932 unsigned cache_policy)
1933 {
1934 vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i16, "");
1935
1936 if (LLVM_VERSION_MAJOR >= 9) {
1937 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1938 ac_build_buffer_store_common(ctx, rsrc, vdata, NULL,
1939 voffset, soffset, 1,
1940 ctx->i16, cache_policy,
1941 false, false);
1942 } else {
1943 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
1944 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1945
1946 vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
1947
1948 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1949 ctx->i32_0, 1, dfmt, nfmt, cache_policy);
1950 }
1951 }
1952
1953 void
1954 ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
1955 LLVMValueRef rsrc,
1956 LLVMValueRef vdata,
1957 LLVMValueRef voffset,
1958 LLVMValueRef soffset,
1959 unsigned cache_policy)
1960 {
1961 vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i8, "");
1962
1963 if (LLVM_VERSION_MAJOR >= 9) {
1964 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1965 ac_build_buffer_store_common(ctx, rsrc, vdata, NULL,
1966 voffset, soffset, 1,
1967 ctx->i8, cache_policy,
1968 false, false);
1969 } else {
1970 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
1971 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1972
1973 vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
1974
1975 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1976 ctx->i32_0, 1, dfmt, nfmt, cache_policy);
1977 }
1978 }
1979 /**
1980 * Set range metadata on an instruction. This can only be used on load and
1981 * call instructions. If you know an instruction can only produce the values
1982 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1983 * \p lo is the minimum value inclusive.
1984 * \p hi is the maximum value exclusive.
1985 */
1986 static void set_range_metadata(struct ac_llvm_context *ctx,
1987 LLVMValueRef value, unsigned lo, unsigned hi)
1988 {
1989 LLVMValueRef range_md, md_args[2];
1990 LLVMTypeRef type = LLVMTypeOf(value);
1991 LLVMContextRef context = LLVMGetTypeContext(type);
1992
1993 md_args[0] = LLVMConstInt(type, lo, false);
1994 md_args[1] = LLVMConstInt(type, hi, false);
1995 range_md = LLVMMDNodeInContext(context, md_args, 2);
1996 LLVMSetMetadata(value, ctx->range_md_kind, range_md);
1997 }
1998
1999 LLVMValueRef
2000 ac_get_thread_id(struct ac_llvm_context *ctx)
2001 {
2002 LLVMValueRef tid;
2003
2004 LLVMValueRef tid_args[2];
2005 tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
2006 tid_args[1] = ctx->i32_0;
2007 tid_args[1] = ac_build_intrinsic(ctx,
2008 "llvm.amdgcn.mbcnt.lo", ctx->i32,
2009 tid_args, 2, AC_FUNC_ATTR_READNONE);
2010
2011 if (ctx->wave_size == 32) {
2012 tid = tid_args[1];
2013 } else {
2014 tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
2015 ctx->i32, tid_args,
2016 2, AC_FUNC_ATTR_READNONE);
2017 }
2018 set_range_metadata(ctx, tid, 0, ctx->wave_size);
2019 return tid;
2020 }
2021
2022 /*
2023 * AMD GCN implements derivatives using the local data store (LDS)
2024 * All writes to the LDS happen in all executing threads at
2025 * the same time. TID is the Thread ID for the current
2026 * thread and is a value between 0 and 63, representing
2027 * the thread's position in the wavefront.
2028 *
2029 * For the pixel shader threads are grouped into quads of four pixels.
2030 * The TIDs of the pixels of a quad are:
2031 *
2032 * +------+------+
2033 * |4n + 0|4n + 1|
2034 * +------+------+
2035 * |4n + 2|4n + 3|
2036 * +------+------+
2037 *
2038 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
2039 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
2040 * the current pixel's column, and masking with 0xfffffffe yields the TID
2041 * of the left pixel of the current pixel's row.
2042 *
2043 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
2044 * adding 2 yields the TID of the pixel below the top pixel.
2045 */
2046 LLVMValueRef
2047 ac_build_ddxy(struct ac_llvm_context *ctx,
2048 uint32_t mask,
2049 int idx,
2050 LLVMValueRef val)
2051 {
2052 unsigned tl_lanes[4], trbl_lanes[4];
2053 char name[32], type[8];
2054 LLVMValueRef tl, trbl;
2055 LLVMTypeRef result_type;
2056 LLVMValueRef result;
2057
2058 result_type = ac_to_float_type(ctx, LLVMTypeOf(val));
2059
2060 if (result_type == ctx->f16)
2061 val = LLVMBuildZExt(ctx->builder, val, ctx->i32, "");
2062
2063 for (unsigned i = 0; i < 4; ++i) {
2064 tl_lanes[i] = i & mask;
2065 trbl_lanes[i] = (i & mask) + idx;
2066 }
2067
2068 tl = ac_build_quad_swizzle(ctx, val,
2069 tl_lanes[0], tl_lanes[1],
2070 tl_lanes[2], tl_lanes[3]);
2071 trbl = ac_build_quad_swizzle(ctx, val,
2072 trbl_lanes[0], trbl_lanes[1],
2073 trbl_lanes[2], trbl_lanes[3]);
2074
2075 if (result_type == ctx->f16) {
2076 tl = LLVMBuildTrunc(ctx->builder, tl, ctx->i16, "");
2077 trbl = LLVMBuildTrunc(ctx->builder, trbl, ctx->i16, "");
2078 }
2079
2080 tl = LLVMBuildBitCast(ctx->builder, tl, result_type, "");
2081 trbl = LLVMBuildBitCast(ctx->builder, trbl, result_type, "");
2082 result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
2083
2084 ac_build_type_name_for_intr(result_type, type, sizeof(type));
2085 snprintf(name, sizeof(name), "llvm.amdgcn.wqm.%s", type);
2086
2087 return ac_build_intrinsic(ctx, name, result_type, &result, 1, 0);
2088 }
2089
2090 void
2091 ac_build_sendmsg(struct ac_llvm_context *ctx,
2092 uint32_t msg,
2093 LLVMValueRef wave_id)
2094 {
2095 LLVMValueRef args[2];
2096 args[0] = LLVMConstInt(ctx->i32, msg, false);
2097 args[1] = wave_id;
2098 ac_build_intrinsic(ctx, "llvm.amdgcn.s.sendmsg", ctx->voidt, args, 2, 0);
2099 }
2100
2101 LLVMValueRef
2102 ac_build_imsb(struct ac_llvm_context *ctx,
2103 LLVMValueRef arg,
2104 LLVMTypeRef dst_type)
2105 {
2106 LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.amdgcn.sffbh.i32",
2107 dst_type, &arg, 1,
2108 AC_FUNC_ATTR_READNONE);
2109
2110 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
2111 * the index from LSB. Invert it by doing "31 - msb". */
2112 msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
2113 msb, "");
2114
2115 LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
2116 LLVMValueRef cond = LLVMBuildOr(ctx->builder,
2117 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
2118 arg, ctx->i32_0, ""),
2119 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
2120 arg, all_ones, ""), "");
2121
2122 return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
2123 }
2124
2125 LLVMValueRef
2126 ac_build_umsb(struct ac_llvm_context *ctx,
2127 LLVMValueRef arg,
2128 LLVMTypeRef dst_type)
2129 {
2130 const char *intrin_name;
2131 LLVMTypeRef type;
2132 LLVMValueRef highest_bit;
2133 LLVMValueRef zero;
2134 unsigned bitsize;
2135
2136 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(arg));
2137 switch (bitsize) {
2138 case 64:
2139 intrin_name = "llvm.ctlz.i64";
2140 type = ctx->i64;
2141 highest_bit = LLVMConstInt(ctx->i64, 63, false);
2142 zero = ctx->i64_0;
2143 break;
2144 case 32:
2145 intrin_name = "llvm.ctlz.i32";
2146 type = ctx->i32;
2147 highest_bit = LLVMConstInt(ctx->i32, 31, false);
2148 zero = ctx->i32_0;
2149 break;
2150 case 16:
2151 intrin_name = "llvm.ctlz.i16";
2152 type = ctx->i16;
2153 highest_bit = LLVMConstInt(ctx->i16, 15, false);
2154 zero = ctx->i16_0;
2155 break;
2156 case 8:
2157 intrin_name = "llvm.ctlz.i8";
2158 type = ctx->i8;
2159 highest_bit = LLVMConstInt(ctx->i8, 7, false);
2160 zero = ctx->i8_0;
2161 break;
2162 default:
2163 unreachable(!"invalid bitsize");
2164 break;
2165 }
2166
2167 LLVMValueRef params[2] = {
2168 arg,
2169 ctx->i1true,
2170 };
2171
2172 LLVMValueRef msb = ac_build_intrinsic(ctx, intrin_name, type,
2173 params, 2,
2174 AC_FUNC_ATTR_READNONE);
2175
2176 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
2177 * the index from LSB. Invert it by doing "31 - msb". */
2178 msb = LLVMBuildSub(ctx->builder, highest_bit, msb, "");
2179
2180 if (bitsize == 64) {
2181 msb = LLVMBuildTrunc(ctx->builder, msb, ctx->i32, "");
2182 } else if (bitsize < 32) {
2183 msb = LLVMBuildSExt(ctx->builder, msb, ctx->i32, "");
2184 }
2185
2186 /* check for zero */
2187 return LLVMBuildSelect(ctx->builder,
2188 LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg, zero, ""),
2189 LLVMConstInt(ctx->i32, -1, true), msb, "");
2190 }
2191
2192 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
2193 LLVMValueRef b)
2194 {
2195 char name[64], type[64];
2196
2197 ac_build_type_name_for_intr(LLVMTypeOf(a), type, sizeof(type));
2198 snprintf(name, sizeof(name), "llvm.minnum.%s", type);
2199 LLVMValueRef args[2] = {a, b};
2200 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
2201 AC_FUNC_ATTR_READNONE);
2202 }
2203
2204 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
2205 LLVMValueRef b)
2206 {
2207 char name[64], type[64];
2208
2209 ac_build_type_name_for_intr(LLVMTypeOf(a), type, sizeof(type));
2210 snprintf(name, sizeof(name), "llvm.maxnum.%s", type);
2211 LLVMValueRef args[2] = {a, b};
2212 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
2213 AC_FUNC_ATTR_READNONE);
2214 }
2215
2216 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
2217 LLVMValueRef b)
2218 {
2219 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSLE, a, b, "");
2220 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2221 }
2222
2223 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
2224 LLVMValueRef b)
2225 {
2226 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, a, b, "");
2227 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2228 }
2229
2230 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a,
2231 LLVMValueRef b)
2232 {
2233 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntULE, a, b, "");
2234 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2235 }
2236
2237 LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a,
2238 LLVMValueRef b)
2239 {
2240 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, a, b, "");
2241 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2242 }
2243
2244 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value)
2245 {
2246 LLVMTypeRef t = LLVMTypeOf(value);
2247 return ac_build_fmin(ctx, ac_build_fmax(ctx, value, LLVMConstReal(t, 0.0)),
2248 LLVMConstReal(t, 1.0));
2249 }
2250
2251 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a)
2252 {
2253 LLVMValueRef args[9];
2254
2255 args[0] = LLVMConstInt(ctx->i32, a->target, 0);
2256 args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
2257
2258 if (a->compr) {
2259 args[2] = LLVMBuildBitCast(ctx->builder, a->out[0],
2260 ctx->v2i16, "");
2261 args[3] = LLVMBuildBitCast(ctx->builder, a->out[1],
2262 ctx->v2i16, "");
2263 args[4] = LLVMConstInt(ctx->i1, a->done, 0);
2264 args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
2265
2266 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16",
2267 ctx->voidt, args, 6, 0);
2268 } else {
2269 args[2] = a->out[0];
2270 args[3] = a->out[1];
2271 args[4] = a->out[2];
2272 args[5] = a->out[3];
2273 args[6] = LLVMConstInt(ctx->i1, a->done, 0);
2274 args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
2275
2276 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.f32",
2277 ctx->voidt, args, 8, 0);
2278 }
2279 }
2280
2281 void ac_build_export_null(struct ac_llvm_context *ctx)
2282 {
2283 struct ac_export_args args;
2284
2285 args.enabled_channels = 0x0; /* enabled channels */
2286 args.valid_mask = 1; /* whether the EXEC mask is valid */
2287 args.done = 1; /* DONE bit */
2288 args.target = V_008DFC_SQ_EXP_NULL;
2289 args.compr = 0; /* COMPR flag (0 = 32-bit export) */
2290 args.out[0] = LLVMGetUndef(ctx->f32); /* R */
2291 args.out[1] = LLVMGetUndef(ctx->f32); /* G */
2292 args.out[2] = LLVMGetUndef(ctx->f32); /* B */
2293 args.out[3] = LLVMGetUndef(ctx->f32); /* A */
2294
2295 ac_build_export(ctx, &args);
2296 }
2297
2298 static unsigned ac_num_coords(enum ac_image_dim dim)
2299 {
2300 switch (dim) {
2301 case ac_image_1d:
2302 return 1;
2303 case ac_image_2d:
2304 case ac_image_1darray:
2305 return 2;
2306 case ac_image_3d:
2307 case ac_image_cube:
2308 case ac_image_2darray:
2309 case ac_image_2dmsaa:
2310 return 3;
2311 case ac_image_2darraymsaa:
2312 return 4;
2313 default:
2314 unreachable("ac_num_coords: bad dim");
2315 }
2316 }
2317
2318 static unsigned ac_num_derivs(enum ac_image_dim dim)
2319 {
2320 switch (dim) {
2321 case ac_image_1d:
2322 case ac_image_1darray:
2323 return 2;
2324 case ac_image_2d:
2325 case ac_image_2darray:
2326 case ac_image_cube:
2327 return 4;
2328 case ac_image_3d:
2329 return 6;
2330 case ac_image_2dmsaa:
2331 case ac_image_2darraymsaa:
2332 default:
2333 unreachable("derivatives not supported");
2334 }
2335 }
2336
2337 static const char *get_atomic_name(enum ac_atomic_op op)
2338 {
2339 switch (op) {
2340 case ac_atomic_swap: return "swap";
2341 case ac_atomic_add: return "add";
2342 case ac_atomic_sub: return "sub";
2343 case ac_atomic_smin: return "smin";
2344 case ac_atomic_umin: return "umin";
2345 case ac_atomic_smax: return "smax";
2346 case ac_atomic_umax: return "umax";
2347 case ac_atomic_and: return "and";
2348 case ac_atomic_or: return "or";
2349 case ac_atomic_xor: return "xor";
2350 case ac_atomic_inc_wrap: return "inc";
2351 case ac_atomic_dec_wrap: return "dec";
2352 }
2353 unreachable("bad atomic op");
2354 }
2355
2356 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
2357 struct ac_image_args *a)
2358 {
2359 const char *overload[3] = { "", "", "" };
2360 unsigned num_overloads = 0;
2361 LLVMValueRef args[18];
2362 unsigned num_args = 0;
2363 enum ac_image_dim dim = a->dim;
2364
2365 assert(!a->lod || a->lod == ctx->i32_0 || a->lod == ctx->f32_0 ||
2366 !a->level_zero);
2367 assert((a->opcode != ac_image_get_resinfo && a->opcode != ac_image_load_mip &&
2368 a->opcode != ac_image_store_mip) ||
2369 a->lod);
2370 assert(a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
2371 (!a->compare && !a->offset));
2372 assert((a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
2373 a->opcode == ac_image_get_lod) ||
2374 !a->bias);
2375 assert((a->bias ? 1 : 0) +
2376 (a->lod ? 1 : 0) +
2377 (a->level_zero ? 1 : 0) +
2378 (a->derivs[0] ? 1 : 0) <= 1);
2379 assert((a->min_lod ? 1 : 0) +
2380 (a->lod ? 1 : 0) +
2381 (a->level_zero ? 1 : 0) <= 1);
2382
2383 if (a->opcode == ac_image_get_lod) {
2384 switch (dim) {
2385 case ac_image_1darray:
2386 dim = ac_image_1d;
2387 break;
2388 case ac_image_2darray:
2389 case ac_image_cube:
2390 dim = ac_image_2d;
2391 break;
2392 default:
2393 break;
2394 }
2395 }
2396
2397 bool sample = a->opcode == ac_image_sample ||
2398 a->opcode == ac_image_gather4 ||
2399 a->opcode == ac_image_get_lod;
2400 bool atomic = a->opcode == ac_image_atomic ||
2401 a->opcode == ac_image_atomic_cmpswap;
2402 bool load = a->opcode == ac_image_sample ||
2403 a->opcode == ac_image_gather4 ||
2404 a->opcode == ac_image_load ||
2405 a->opcode == ac_image_load_mip;
2406 LLVMTypeRef coord_type = sample ? ctx->f32 : ctx->i32;
2407
2408 if (atomic || a->opcode == ac_image_store || a->opcode == ac_image_store_mip) {
2409 args[num_args++] = a->data[0];
2410 if (a->opcode == ac_image_atomic_cmpswap)
2411 args[num_args++] = a->data[1];
2412 }
2413
2414 if (!atomic)
2415 args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, false);
2416
2417 if (a->offset)
2418 args[num_args++] = ac_to_integer(ctx, a->offset);
2419 if (a->bias) {
2420 args[num_args++] = ac_to_float(ctx, a->bias);
2421 overload[num_overloads++] = ".f32";
2422 }
2423 if (a->compare)
2424 args[num_args++] = ac_to_float(ctx, a->compare);
2425 if (a->derivs[0]) {
2426 unsigned count = ac_num_derivs(dim);
2427 for (unsigned i = 0; i < count; ++i)
2428 args[num_args++] = ac_to_float(ctx, a->derivs[i]);
2429 overload[num_overloads++] = ".f32";
2430 }
2431 unsigned num_coords =
2432 a->opcode != ac_image_get_resinfo ? ac_num_coords(dim) : 0;
2433 for (unsigned i = 0; i < num_coords; ++i)
2434 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->coords[i], coord_type, "");
2435 if (a->lod)
2436 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->lod, coord_type, "");
2437 if (a->min_lod)
2438 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->min_lod, coord_type, "");
2439
2440 overload[num_overloads++] = sample ? ".f32" : ".i32";
2441
2442 args[num_args++] = a->resource;
2443 if (sample) {
2444 args[num_args++] = a->sampler;
2445 args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, false);
2446 }
2447
2448 args[num_args++] = ctx->i32_0; /* texfailctrl */
2449 args[num_args++] = LLVMConstInt(ctx->i32,
2450 load ? get_load_cache_policy(ctx, a->cache_policy) :
2451 a->cache_policy, false);
2452
2453 const char *name;
2454 const char *atomic_subop = "";
2455 switch (a->opcode) {
2456 case ac_image_sample: name = "sample"; break;
2457 case ac_image_gather4: name = "gather4"; break;
2458 case ac_image_load: name = "load"; break;
2459 case ac_image_load_mip: name = "load.mip"; break;
2460 case ac_image_store: name = "store"; break;
2461 case ac_image_store_mip: name = "store.mip"; break;
2462 case ac_image_atomic:
2463 name = "atomic.";
2464 atomic_subop = get_atomic_name(a->atomic);
2465 break;
2466 case ac_image_atomic_cmpswap:
2467 name = "atomic.";
2468 atomic_subop = "cmpswap";
2469 break;
2470 case ac_image_get_lod: name = "getlod"; break;
2471 case ac_image_get_resinfo: name = "getresinfo"; break;
2472 default: unreachable("invalid image opcode");
2473 }
2474
2475 const char *dimname;
2476 switch (dim) {
2477 case ac_image_1d: dimname = "1d"; break;
2478 case ac_image_2d: dimname = "2d"; break;
2479 case ac_image_3d: dimname = "3d"; break;
2480 case ac_image_cube: dimname = "cube"; break;
2481 case ac_image_1darray: dimname = "1darray"; break;
2482 case ac_image_2darray: dimname = "2darray"; break;
2483 case ac_image_2dmsaa: dimname = "2dmsaa"; break;
2484 case ac_image_2darraymsaa: dimname = "2darraymsaa"; break;
2485 default: unreachable("invalid dim");
2486 }
2487
2488 bool lod_suffix =
2489 a->lod && (a->opcode == ac_image_sample || a->opcode == ac_image_gather4);
2490 char intr_name[96];
2491 snprintf(intr_name, sizeof(intr_name),
2492 "llvm.amdgcn.image.%s%s" /* base name */
2493 "%s%s%s%s" /* sample/gather modifiers */
2494 ".%s.%s%s%s%s", /* dimension and type overloads */
2495 name, atomic_subop,
2496 a->compare ? ".c" : "",
2497 a->bias ? ".b" :
2498 lod_suffix ? ".l" :
2499 a->derivs[0] ? ".d" :
2500 a->level_zero ? ".lz" : "",
2501 a->min_lod ? ".cl" : "",
2502 a->offset ? ".o" : "",
2503 dimname,
2504 atomic ? "i32" : "v4f32",
2505 overload[0], overload[1], overload[2]);
2506
2507 LLVMTypeRef retty;
2508 if (atomic)
2509 retty = ctx->i32;
2510 else if (a->opcode == ac_image_store || a->opcode == ac_image_store_mip)
2511 retty = ctx->voidt;
2512 else
2513 retty = ctx->v4f32;
2514
2515 LLVMValueRef result =
2516 ac_build_intrinsic(ctx, intr_name, retty, args, num_args,
2517 a->attributes);
2518 if (!sample && retty == ctx->v4f32) {
2519 result = LLVMBuildBitCast(ctx->builder, result,
2520 ctx->v4i32, "");
2521 }
2522 return result;
2523 }
2524
2525 LLVMValueRef ac_build_image_get_sample_count(struct ac_llvm_context *ctx,
2526 LLVMValueRef rsrc)
2527 {
2528 LLVMValueRef samples;
2529
2530 /* Read the samples from the descriptor directly.
2531 * Hardware doesn't have any instruction for this.
2532 */
2533 samples = LLVMBuildExtractElement(ctx->builder, rsrc,
2534 LLVMConstInt(ctx->i32, 3, 0), "");
2535 samples = LLVMBuildLShr(ctx->builder, samples,
2536 LLVMConstInt(ctx->i32, 16, 0), "");
2537 samples = LLVMBuildAnd(ctx->builder, samples,
2538 LLVMConstInt(ctx->i32, 0xf, 0), "");
2539 samples = LLVMBuildShl(ctx->builder, ctx->i32_1,
2540 samples, "");
2541 return samples;
2542 }
2543
2544 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
2545 LLVMValueRef args[2])
2546 {
2547 return ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz", ctx->v2f16,
2548 args, 2, AC_FUNC_ATTR_READNONE);
2549 }
2550
2551 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
2552 LLVMValueRef args[2])
2553 {
2554 LLVMValueRef res =
2555 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16",
2556 ctx->v2i16, args, 2,
2557 AC_FUNC_ATTR_READNONE);
2558 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2559 }
2560
2561 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
2562 LLVMValueRef args[2])
2563 {
2564 LLVMValueRef res =
2565 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16",
2566 ctx->v2i16, args, 2,
2567 AC_FUNC_ATTR_READNONE);
2568 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2569 }
2570
2571 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2572 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
2573 LLVMValueRef args[2], unsigned bits, bool hi)
2574 {
2575 assert(bits == 8 || bits == 10 || bits == 16);
2576
2577 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2578 bits == 8 ? 127 : bits == 10 ? 511 : 32767, 0);
2579 LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
2580 bits == 8 ? -128 : bits == 10 ? -512 : -32768, 0);
2581 LLVMValueRef max_alpha =
2582 bits != 10 ? max_rgb : ctx->i32_1;
2583 LLVMValueRef min_alpha =
2584 bits != 10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
2585
2586 /* Clamp. */
2587 if (bits != 16) {
2588 for (int i = 0; i < 2; i++) {
2589 bool alpha = hi && i == 1;
2590 args[i] = ac_build_imin(ctx, args[i],
2591 alpha ? max_alpha : max_rgb);
2592 args[i] = ac_build_imax(ctx, args[i],
2593 alpha ? min_alpha : min_rgb);
2594 }
2595 }
2596
2597 LLVMValueRef res =
2598 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16",
2599 ctx->v2i16, args, 2,
2600 AC_FUNC_ATTR_READNONE);
2601 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2602 }
2603
2604 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2605 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
2606 LLVMValueRef args[2], unsigned bits, bool hi)
2607 {
2608 assert(bits == 8 || bits == 10 || bits == 16);
2609
2610 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2611 bits == 8 ? 255 : bits == 10 ? 1023 : 65535, 0);
2612 LLVMValueRef max_alpha =
2613 bits != 10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
2614
2615 /* Clamp. */
2616 if (bits != 16) {
2617 for (int i = 0; i < 2; i++) {
2618 bool alpha = hi && i == 1;
2619 args[i] = ac_build_umin(ctx, args[i],
2620 alpha ? max_alpha : max_rgb);
2621 }
2622 }
2623
2624 LLVMValueRef res =
2625 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16",
2626 ctx->v2i16, args, 2,
2627 AC_FUNC_ATTR_READNONE);
2628 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2629 }
2630
2631 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1)
2632 {
2633 return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1,
2634 &i1, 1, AC_FUNC_ATTR_READNONE);
2635 }
2636
2637 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1)
2638 {
2639 ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt,
2640 &i1, 1, 0);
2641 }
2642
2643 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
2644 LLVMValueRef offset, LLVMValueRef width,
2645 bool is_signed)
2646 {
2647 LLVMValueRef args[] = {
2648 input,
2649 offset,
2650 width,
2651 };
2652
2653 return ac_build_intrinsic(ctx, is_signed ? "llvm.amdgcn.sbfe.i32" :
2654 "llvm.amdgcn.ubfe.i32",
2655 ctx->i32, args, 3, AC_FUNC_ATTR_READNONE);
2656
2657 }
2658
2659 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2660 LLVMValueRef s1, LLVMValueRef s2)
2661 {
2662 return LLVMBuildAdd(ctx->builder,
2663 LLVMBuildMul(ctx->builder, s0, s1, ""), s2, "");
2664 }
2665
2666 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2667 LLVMValueRef s1, LLVMValueRef s2)
2668 {
2669 /* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
2670 if (ctx->chip_class >= GFX10) {
2671 return ac_build_intrinsic(ctx, "llvm.fma.f32", ctx->f32,
2672 (LLVMValueRef []) {s0, s1, s2}, 3,
2673 AC_FUNC_ATTR_READNONE);
2674 }
2675
2676 return LLVMBuildFAdd(ctx->builder,
2677 LLVMBuildFMul(ctx->builder, s0, s1, ""), s2, "");
2678 }
2679
2680 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags)
2681 {
2682 if (!wait_flags)
2683 return;
2684
2685 unsigned lgkmcnt = 63;
2686 unsigned vmcnt = ctx->chip_class >= GFX9 ? 63 : 15;
2687 unsigned vscnt = 63;
2688
2689 if (wait_flags & AC_WAIT_LGKM)
2690 lgkmcnt = 0;
2691 if (wait_flags & AC_WAIT_VLOAD)
2692 vmcnt = 0;
2693
2694 if (wait_flags & AC_WAIT_VSTORE) {
2695 if (ctx->chip_class >= GFX10)
2696 vscnt = 0;
2697 else
2698 vmcnt = 0;
2699 }
2700
2701 /* There is no intrinsic for vscnt(0), so use a fence. */
2702 if ((wait_flags & AC_WAIT_LGKM &&
2703 wait_flags & AC_WAIT_VLOAD &&
2704 wait_flags & AC_WAIT_VSTORE) ||
2705 vscnt == 0) {
2706 LLVMBuildFence(ctx->builder, LLVMAtomicOrderingRelease, false, "");
2707 return;
2708 }
2709
2710 unsigned simm16 = (lgkmcnt << 8) |
2711 (7 << 4) | /* expcnt */
2712 (vmcnt & 0xf) |
2713 ((vmcnt >> 4) << 14);
2714
2715 LLVMValueRef args[1] = {
2716 LLVMConstInt(ctx->i32, simm16, false),
2717 };
2718 ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt",
2719 ctx->voidt, args, 1, 0);
2720 }
2721
2722 LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
2723 LLVMValueRef src1, LLVMValueRef src2,
2724 unsigned bitsize)
2725 {
2726 LLVMValueRef result;
2727
2728 if (bitsize == 64 || (bitsize == 16 && ctx->chip_class <= GFX8)) {
2729 /* Lower 64-bit fmed because LLVM doesn't expose an intrinsic,
2730 * or lower 16-bit fmed because it's only supported on GFX9+.
2731 */
2732 LLVMValueRef min1, min2, max1;
2733
2734 min1 = ac_build_fmin(ctx, src0, src1);
2735 max1 = ac_build_fmax(ctx, src0, src1);
2736 min2 = ac_build_fmin(ctx, max1, src2);
2737
2738 result = ac_build_fmax(ctx, min2, min1);
2739 } else {
2740 LLVMTypeRef type;
2741 char *intr;
2742
2743 if (bitsize == 16) {
2744 intr = "llvm.amdgcn.fmed3.f16";
2745 type = ctx->f16;
2746 } else {
2747 assert(bitsize == 32);
2748 intr = "llvm.amdgcn.fmed3.f32";
2749 type = ctx->f32;
2750 }
2751
2752 LLVMValueRef params[] = {
2753 src0,
2754 src1,
2755 src2,
2756 };
2757
2758 result = ac_build_intrinsic(ctx, intr, type, params, 3,
2759 AC_FUNC_ATTR_READNONE);
2760 }
2761
2762 if (ctx->chip_class < GFX9 && bitsize == 32) {
2763 /* Only pre-GFX9 chips do not flush denorms. */
2764 result = ac_build_canonicalize(ctx, result, bitsize);
2765 }
2766
2767 return result;
2768 }
2769
2770 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
2771 unsigned bitsize)
2772 {
2773 LLVMTypeRef type;
2774 char *intr;
2775
2776 if (bitsize == 16) {
2777 intr = "llvm.amdgcn.fract.f16";
2778 type = ctx->f16;
2779 } else if (bitsize == 32) {
2780 intr = "llvm.amdgcn.fract.f32";
2781 type = ctx->f32;
2782 } else {
2783 intr = "llvm.amdgcn.fract.f64";
2784 type = ctx->f64;
2785 }
2786
2787 LLVMValueRef params[] = {
2788 src0,
2789 };
2790 return ac_build_intrinsic(ctx, intr, type, params, 1,
2791 AC_FUNC_ATTR_READNONE);
2792 }
2793
2794 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2795 unsigned bitsize)
2796 {
2797 LLVMTypeRef type = LLVMIntTypeInContext(ctx->context, bitsize);
2798 LLVMValueRef zero = LLVMConstInt(type, 0, false);
2799 LLVMValueRef one = LLVMConstInt(type, 1, false);
2800
2801 LLVMValueRef cmp, val;
2802 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, src0, zero, "");
2803 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2804 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGE, val, zero, "");
2805 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstInt(type, -1, true), "");
2806 return val;
2807 }
2808
2809 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2810 unsigned bitsize)
2811 {
2812 LLVMValueRef cmp, val, zero, one;
2813 LLVMTypeRef type;
2814
2815 if (bitsize == 16) {
2816 type = ctx->f16;
2817 zero = ctx->f16_0;
2818 one = ctx->f16_1;
2819 } else if (bitsize == 32) {
2820 type = ctx->f32;
2821 zero = ctx->f32_0;
2822 one = ctx->f32_1;
2823 } else {
2824 type = ctx->f64;
2825 zero = ctx->f64_0;
2826 one = ctx->f64_1;
2827 }
2828
2829 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src0, zero, "");
2830 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2831 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGE, val, zero, "");
2832 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstReal(type, -1.0), "");
2833 return val;
2834 }
2835
2836 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0)
2837 {
2838 LLVMValueRef result;
2839 unsigned bitsize;
2840
2841 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2842
2843 switch (bitsize) {
2844 case 128:
2845 result = ac_build_intrinsic(ctx, "llvm.ctpop.i128", ctx->i128,
2846 (LLVMValueRef []) { src0 }, 1,
2847 AC_FUNC_ATTR_READNONE);
2848 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
2849 break;
2850 case 64:
2851 result = ac_build_intrinsic(ctx, "llvm.ctpop.i64", ctx->i64,
2852 (LLVMValueRef []) { src0 }, 1,
2853 AC_FUNC_ATTR_READNONE);
2854
2855 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
2856 break;
2857 case 32:
2858 result = ac_build_intrinsic(ctx, "llvm.ctpop.i32", ctx->i32,
2859 (LLVMValueRef []) { src0 }, 1,
2860 AC_FUNC_ATTR_READNONE);
2861 break;
2862 case 16:
2863 result = ac_build_intrinsic(ctx, "llvm.ctpop.i16", ctx->i16,
2864 (LLVMValueRef []) { src0 }, 1,
2865 AC_FUNC_ATTR_READNONE);
2866
2867 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2868 break;
2869 case 8:
2870 result = ac_build_intrinsic(ctx, "llvm.ctpop.i8", ctx->i8,
2871 (LLVMValueRef []) { src0 }, 1,
2872 AC_FUNC_ATTR_READNONE);
2873
2874 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2875 break;
2876 default:
2877 unreachable(!"invalid bitsize");
2878 break;
2879 }
2880
2881 return result;
2882 }
2883
2884 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
2885 LLVMValueRef src0)
2886 {
2887 LLVMValueRef result;
2888 unsigned bitsize;
2889
2890 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2891
2892 switch (bitsize) {
2893 case 64:
2894 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i64", ctx->i64,
2895 (LLVMValueRef []) { src0 }, 1,
2896 AC_FUNC_ATTR_READNONE);
2897
2898 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
2899 break;
2900 case 32:
2901 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i32", ctx->i32,
2902 (LLVMValueRef []) { src0 }, 1,
2903 AC_FUNC_ATTR_READNONE);
2904 break;
2905 case 16:
2906 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i16", ctx->i16,
2907 (LLVMValueRef []) { src0 }, 1,
2908 AC_FUNC_ATTR_READNONE);
2909
2910 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2911 break;
2912 case 8:
2913 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i8", ctx->i8,
2914 (LLVMValueRef []) { src0 }, 1,
2915 AC_FUNC_ATTR_READNONE);
2916
2917 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2918 break;
2919 default:
2920 unreachable(!"invalid bitsize");
2921 break;
2922 }
2923
2924 return result;
2925 }
2926
2927 #define AC_EXP_TARGET 0
2928 #define AC_EXP_ENABLED_CHANNELS 1
2929 #define AC_EXP_OUT0 2
2930
2931 enum ac_ir_type {
2932 AC_IR_UNDEF,
2933 AC_IR_CONST,
2934 AC_IR_VALUE,
2935 };
2936
2937 struct ac_vs_exp_chan
2938 {
2939 LLVMValueRef value;
2940 float const_float;
2941 enum ac_ir_type type;
2942 };
2943
2944 struct ac_vs_exp_inst {
2945 unsigned offset;
2946 LLVMValueRef inst;
2947 struct ac_vs_exp_chan chan[4];
2948 };
2949
2950 struct ac_vs_exports {
2951 unsigned num;
2952 struct ac_vs_exp_inst exp[VARYING_SLOT_MAX];
2953 };
2954
2955 /* Return true if the PARAM export has been eliminated. */
2956 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset,
2957 uint32_t num_outputs,
2958 struct ac_vs_exp_inst *exp)
2959 {
2960 unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
2961 bool is_zero[4] = {}, is_one[4] = {};
2962
2963 for (i = 0; i < 4; i++) {
2964 /* It's a constant expression. Undef outputs are eliminated too. */
2965 if (exp->chan[i].type == AC_IR_UNDEF) {
2966 is_zero[i] = true;
2967 is_one[i] = true;
2968 } else if (exp->chan[i].type == AC_IR_CONST) {
2969 if (exp->chan[i].const_float == 0)
2970 is_zero[i] = true;
2971 else if (exp->chan[i].const_float == 1)
2972 is_one[i] = true;
2973 else
2974 return false; /* other constant */
2975 } else
2976 return false;
2977 }
2978
2979 /* Only certain combinations of 0 and 1 can be eliminated. */
2980 if (is_zero[0] && is_zero[1] && is_zero[2])
2981 default_val = is_zero[3] ? 0 : 1;
2982 else if (is_one[0] && is_one[1] && is_one[2])
2983 default_val = is_zero[3] ? 2 : 3;
2984 else
2985 return false;
2986
2987 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
2988 LLVMInstructionEraseFromParent(exp->inst);
2989
2990 /* Change OFFSET to DEFAULT_VAL. */
2991 for (i = 0; i < num_outputs; i++) {
2992 if (vs_output_param_offset[i] == exp->offset) {
2993 vs_output_param_offset[i] =
2994 AC_EXP_PARAM_DEFAULT_VAL_0000 + default_val;
2995 break;
2996 }
2997 }
2998 return true;
2999 }
3000
3001 static bool ac_eliminate_duplicated_output(struct ac_llvm_context *ctx,
3002 uint8_t *vs_output_param_offset,
3003 uint32_t num_outputs,
3004 struct ac_vs_exports *processed,
3005 struct ac_vs_exp_inst *exp)
3006 {
3007 unsigned p, copy_back_channels = 0;
3008
3009 /* See if the output is already in the list of processed outputs.
3010 * The LLVMValueRef comparison relies on SSA.
3011 */
3012 for (p = 0; p < processed->num; p++) {
3013 bool different = false;
3014
3015 for (unsigned j = 0; j < 4; j++) {
3016 struct ac_vs_exp_chan *c1 = &processed->exp[p].chan[j];
3017 struct ac_vs_exp_chan *c2 = &exp->chan[j];
3018
3019 /* Treat undef as a match. */
3020 if (c2->type == AC_IR_UNDEF)
3021 continue;
3022
3023 /* If c1 is undef but c2 isn't, we can copy c2 to c1
3024 * and consider the instruction duplicated.
3025 */
3026 if (c1->type == AC_IR_UNDEF) {
3027 copy_back_channels |= 1 << j;
3028 continue;
3029 }
3030
3031 /* Test whether the channels are not equal. */
3032 if (c1->type != c2->type ||
3033 (c1->type == AC_IR_CONST &&
3034 c1->const_float != c2->const_float) ||
3035 (c1->type == AC_IR_VALUE &&
3036 c1->value != c2->value)) {
3037 different = true;
3038 break;
3039 }
3040 }
3041 if (!different)
3042 break;
3043
3044 copy_back_channels = 0;
3045 }
3046 if (p == processed->num)
3047 return false;
3048
3049 /* If a match was found, but the matching export has undef where the new
3050 * one has a normal value, copy the normal value to the undef channel.
3051 */
3052 struct ac_vs_exp_inst *match = &processed->exp[p];
3053
3054 /* Get current enabled channels mask. */
3055 LLVMValueRef arg = LLVMGetOperand(match->inst, AC_EXP_ENABLED_CHANNELS);
3056 unsigned enabled_channels = LLVMConstIntGetZExtValue(arg);
3057
3058 while (copy_back_channels) {
3059 unsigned chan = u_bit_scan(&copy_back_channels);
3060
3061 assert(match->chan[chan].type == AC_IR_UNDEF);
3062 LLVMSetOperand(match->inst, AC_EXP_OUT0 + chan,
3063 exp->chan[chan].value);
3064 match->chan[chan] = exp->chan[chan];
3065
3066 /* Update number of enabled channels because the original mask
3067 * is not always 0xf.
3068 */
3069 enabled_channels |= (1 << chan);
3070 LLVMSetOperand(match->inst, AC_EXP_ENABLED_CHANNELS,
3071 LLVMConstInt(ctx->i32, enabled_channels, 0));
3072 }
3073
3074 /* The PARAM export is duplicated. Kill it. */
3075 LLVMInstructionEraseFromParent(exp->inst);
3076
3077 /* Change OFFSET to the matching export. */
3078 for (unsigned i = 0; i < num_outputs; i++) {
3079 if (vs_output_param_offset[i] == exp->offset) {
3080 vs_output_param_offset[i] = match->offset;
3081 break;
3082 }
3083 }
3084 return true;
3085 }
3086
3087 void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
3088 LLVMValueRef main_fn,
3089 uint8_t *vs_output_param_offset,
3090 uint32_t num_outputs,
3091 uint32_t skip_output_mask,
3092 uint8_t *num_param_exports)
3093 {
3094 LLVMBasicBlockRef bb;
3095 bool removed_any = false;
3096 struct ac_vs_exports exports;
3097
3098 exports.num = 0;
3099
3100 /* Process all LLVM instructions. */
3101 bb = LLVMGetFirstBasicBlock(main_fn);
3102 while (bb) {
3103 LLVMValueRef inst = LLVMGetFirstInstruction(bb);
3104
3105 while (inst) {
3106 LLVMValueRef cur = inst;
3107 inst = LLVMGetNextInstruction(inst);
3108 struct ac_vs_exp_inst exp;
3109
3110 if (LLVMGetInstructionOpcode(cur) != LLVMCall)
3111 continue;
3112
3113 LLVMValueRef callee = ac_llvm_get_called_value(cur);
3114
3115 if (!ac_llvm_is_function(callee))
3116 continue;
3117
3118 const char *name = LLVMGetValueName(callee);
3119 unsigned num_args = LLVMCountParams(callee);
3120
3121 /* Check if this is an export instruction. */
3122 if ((num_args != 9 && num_args != 8) ||
3123 (strcmp(name, "llvm.SI.export") &&
3124 strcmp(name, "llvm.amdgcn.exp.f32")))
3125 continue;
3126
3127 LLVMValueRef arg = LLVMGetOperand(cur, AC_EXP_TARGET);
3128 unsigned target = LLVMConstIntGetZExtValue(arg);
3129
3130 if (target < V_008DFC_SQ_EXP_PARAM)
3131 continue;
3132
3133 target -= V_008DFC_SQ_EXP_PARAM;
3134
3135 /* Parse the instruction. */
3136 memset(&exp, 0, sizeof(exp));
3137 exp.offset = target;
3138 exp.inst = cur;
3139
3140 for (unsigned i = 0; i < 4; i++) {
3141 LLVMValueRef v = LLVMGetOperand(cur, AC_EXP_OUT0 + i);
3142
3143 exp.chan[i].value = v;
3144
3145 if (LLVMIsUndef(v)) {
3146 exp.chan[i].type = AC_IR_UNDEF;
3147 } else if (LLVMIsAConstantFP(v)) {
3148 LLVMBool loses_info;
3149 exp.chan[i].type = AC_IR_CONST;
3150 exp.chan[i].const_float =
3151 LLVMConstRealGetDouble(v, &loses_info);
3152 } else {
3153 exp.chan[i].type = AC_IR_VALUE;
3154 }
3155 }
3156
3157 /* Eliminate constant and duplicated PARAM exports. */
3158 if (!((1u << target) & skip_output_mask) &&
3159 (ac_eliminate_const_output(vs_output_param_offset,
3160 num_outputs, &exp) ||
3161 ac_eliminate_duplicated_output(ctx,
3162 vs_output_param_offset,
3163 num_outputs, &exports,
3164 &exp))) {
3165 removed_any = true;
3166 } else {
3167 exports.exp[exports.num++] = exp;
3168 }
3169 }
3170 bb = LLVMGetNextBasicBlock(bb);
3171 }
3172
3173 /* Remove holes in export memory due to removed PARAM exports.
3174 * This is done by renumbering all PARAM exports.
3175 */
3176 if (removed_any) {
3177 uint8_t old_offset[VARYING_SLOT_MAX];
3178 unsigned out, i;
3179
3180 /* Make a copy of the offsets. We need the old version while
3181 * we are modifying some of them. */
3182 memcpy(old_offset, vs_output_param_offset,
3183 sizeof(old_offset));
3184
3185 for (i = 0; i < exports.num; i++) {
3186 unsigned offset = exports.exp[i].offset;
3187
3188 /* Update vs_output_param_offset. Multiple outputs can
3189 * have the same offset.
3190 */
3191 for (out = 0; out < num_outputs; out++) {
3192 if (old_offset[out] == offset)
3193 vs_output_param_offset[out] = i;
3194 }
3195
3196 /* Change the PARAM offset in the instruction. */
3197 LLVMSetOperand(exports.exp[i].inst, AC_EXP_TARGET,
3198 LLVMConstInt(ctx->i32,
3199 V_008DFC_SQ_EXP_PARAM + i, 0));
3200 }
3201 *num_param_exports = exports.num;
3202 }
3203 }
3204
3205 void ac_init_exec_full_mask(struct ac_llvm_context *ctx)
3206 {
3207 LLVMValueRef full_mask = LLVMConstInt(ctx->i64, ~0ull, 0);
3208 ac_build_intrinsic(ctx,
3209 "llvm.amdgcn.init.exec", ctx->voidt,
3210 &full_mask, 1, AC_FUNC_ATTR_CONVERGENT);
3211 }
3212
3213 void ac_declare_lds_as_pointer(struct ac_llvm_context *ctx)
3214 {
3215 unsigned lds_size = ctx->chip_class >= GFX7 ? 65536 : 32768;
3216 ctx->lds = LLVMBuildIntToPtr(ctx->builder, ctx->i32_0,
3217 LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), AC_ADDR_SPACE_LDS),
3218 "lds");
3219 }
3220
3221 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
3222 LLVMValueRef dw_addr)
3223 {
3224 return LLVMBuildLoad(ctx->builder, ac_build_gep0(ctx, ctx->lds, dw_addr), "");
3225 }
3226
3227 void ac_lds_store(struct ac_llvm_context *ctx,
3228 LLVMValueRef dw_addr,
3229 LLVMValueRef value)
3230 {
3231 value = ac_to_integer(ctx, value);
3232 ac_build_indexed_store(ctx, ctx->lds,
3233 dw_addr, value);
3234 }
3235
3236 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
3237 LLVMTypeRef dst_type,
3238 LLVMValueRef src0)
3239 {
3240 unsigned src0_bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
3241 const char *intrin_name;
3242 LLVMTypeRef type;
3243 LLVMValueRef zero;
3244
3245 switch (src0_bitsize) {
3246 case 64:
3247 intrin_name = "llvm.cttz.i64";
3248 type = ctx->i64;
3249 zero = ctx->i64_0;
3250 break;
3251 case 32:
3252 intrin_name = "llvm.cttz.i32";
3253 type = ctx->i32;
3254 zero = ctx->i32_0;
3255 break;
3256 case 16:
3257 intrin_name = "llvm.cttz.i16";
3258 type = ctx->i16;
3259 zero = ctx->i16_0;
3260 break;
3261 case 8:
3262 intrin_name = "llvm.cttz.i8";
3263 type = ctx->i8;
3264 zero = ctx->i8_0;
3265 break;
3266 default:
3267 unreachable(!"invalid bitsize");
3268 }
3269
3270 LLVMValueRef params[2] = {
3271 src0,
3272
3273 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
3274 * add special code to check for x=0. The reason is that
3275 * the LLVM behavior for x=0 is different from what we
3276 * need here. However, LLVM also assumes that ffs(x) is
3277 * in [0, 31], but GLSL expects that ffs(0) = -1, so
3278 * a conditional assignment to handle 0 is still required.
3279 *
3280 * The hardware already implements the correct behavior.
3281 */
3282 ctx->i1true,
3283 };
3284
3285 LLVMValueRef lsb = ac_build_intrinsic(ctx, intrin_name, type,
3286 params, 2,
3287 AC_FUNC_ATTR_READNONE);
3288
3289 if (src0_bitsize == 64) {
3290 lsb = LLVMBuildTrunc(ctx->builder, lsb, ctx->i32, "");
3291 } else if (src0_bitsize < 32) {
3292 lsb = LLVMBuildSExt(ctx->builder, lsb, ctx->i32, "");
3293 }
3294
3295 /* TODO: We need an intrinsic to skip this conditional. */
3296 /* Check for zero: */
3297 return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder,
3298 LLVMIntEQ, src0,
3299 zero, ""),
3300 LLVMConstInt(ctx->i32, -1, 0), lsb, "");
3301 }
3302
3303 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type)
3304 {
3305 return LLVMPointerType(elem_type, AC_ADDR_SPACE_CONST);
3306 }
3307
3308 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type)
3309 {
3310 return LLVMPointerType(elem_type, AC_ADDR_SPACE_CONST_32BIT);
3311 }
3312
3313 static struct ac_llvm_flow *
3314 get_current_flow(struct ac_llvm_context *ctx)
3315 {
3316 if (ctx->flow->depth > 0)
3317 return &ctx->flow->stack[ctx->flow->depth - 1];
3318 return NULL;
3319 }
3320
3321 static struct ac_llvm_flow *
3322 get_innermost_loop(struct ac_llvm_context *ctx)
3323 {
3324 for (unsigned i = ctx->flow->depth; i > 0; --i) {
3325 if (ctx->flow->stack[i - 1].loop_entry_block)
3326 return &ctx->flow->stack[i - 1];
3327 }
3328 return NULL;
3329 }
3330
3331 static struct ac_llvm_flow *
3332 push_flow(struct ac_llvm_context *ctx)
3333 {
3334 struct ac_llvm_flow *flow;
3335
3336 if (ctx->flow->depth >= ctx->flow->depth_max) {
3337 unsigned new_max = MAX2(ctx->flow->depth << 1,
3338 AC_LLVM_INITIAL_CF_DEPTH);
3339
3340 ctx->flow->stack = realloc(ctx->flow->stack, new_max * sizeof(*ctx->flow->stack));
3341 ctx->flow->depth_max = new_max;
3342 }
3343
3344 flow = &ctx->flow->stack[ctx->flow->depth];
3345 ctx->flow->depth++;
3346
3347 flow->next_block = NULL;
3348 flow->loop_entry_block = NULL;
3349 return flow;
3350 }
3351
3352 static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base,
3353 int label_id)
3354 {
3355 char buf[32];
3356 snprintf(buf, sizeof(buf), "%s%d", base, label_id);
3357 LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
3358 }
3359
3360 /* Append a basic block at the level of the parent flow.
3361 */
3362 static LLVMBasicBlockRef append_basic_block(struct ac_llvm_context *ctx,
3363 const char *name)
3364 {
3365 assert(ctx->flow->depth >= 1);
3366
3367 if (ctx->flow->depth >= 2) {
3368 struct ac_llvm_flow *flow = &ctx->flow->stack[ctx->flow->depth - 2];
3369
3370 return LLVMInsertBasicBlockInContext(ctx->context,
3371 flow->next_block, name);
3372 }
3373
3374 LLVMValueRef main_fn =
3375 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->builder));
3376 return LLVMAppendBasicBlockInContext(ctx->context, main_fn, name);
3377 }
3378
3379 /* Emit a branch to the given default target for the current block if
3380 * applicable -- that is, if the current block does not already contain a
3381 * branch from a break or continue.
3382 */
3383 static void emit_default_branch(LLVMBuilderRef builder,
3384 LLVMBasicBlockRef target)
3385 {
3386 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
3387 LLVMBuildBr(builder, target);
3388 }
3389
3390 void ac_build_bgnloop(struct ac_llvm_context *ctx, int label_id)
3391 {
3392 struct ac_llvm_flow *flow = push_flow(ctx);
3393 flow->loop_entry_block = append_basic_block(ctx, "LOOP");
3394 flow->next_block = append_basic_block(ctx, "ENDLOOP");
3395 set_basicblock_name(flow->loop_entry_block, "loop", label_id);
3396 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
3397 LLVMPositionBuilderAtEnd(ctx->builder, flow->loop_entry_block);
3398 }
3399
3400 void ac_build_break(struct ac_llvm_context *ctx)
3401 {
3402 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
3403 LLVMBuildBr(ctx->builder, flow->next_block);
3404 }
3405
3406 void ac_build_continue(struct ac_llvm_context *ctx)
3407 {
3408 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
3409 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
3410 }
3411
3412 void ac_build_else(struct ac_llvm_context *ctx, int label_id)
3413 {
3414 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
3415 LLVMBasicBlockRef endif_block;
3416
3417 assert(!current_branch->loop_entry_block);
3418
3419 endif_block = append_basic_block(ctx, "ENDIF");
3420 emit_default_branch(ctx->builder, endif_block);
3421
3422 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
3423 set_basicblock_name(current_branch->next_block, "else", label_id);
3424
3425 current_branch->next_block = endif_block;
3426 }
3427
3428 void ac_build_endif(struct ac_llvm_context *ctx, int label_id)
3429 {
3430 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
3431
3432 assert(!current_branch->loop_entry_block);
3433
3434 emit_default_branch(ctx->builder, current_branch->next_block);
3435 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
3436 set_basicblock_name(current_branch->next_block, "endif", label_id);
3437
3438 ctx->flow->depth--;
3439 }
3440
3441 void ac_build_endloop(struct ac_llvm_context *ctx, int label_id)
3442 {
3443 struct ac_llvm_flow *current_loop = get_current_flow(ctx);
3444
3445 assert(current_loop->loop_entry_block);
3446
3447 emit_default_branch(ctx->builder, current_loop->loop_entry_block);
3448
3449 LLVMPositionBuilderAtEnd(ctx->builder, current_loop->next_block);
3450 set_basicblock_name(current_loop->next_block, "endloop", label_id);
3451 ctx->flow->depth--;
3452 }
3453
3454 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id)
3455 {
3456 struct ac_llvm_flow *flow = push_flow(ctx);
3457 LLVMBasicBlockRef if_block;
3458
3459 if_block = append_basic_block(ctx, "IF");
3460 flow->next_block = append_basic_block(ctx, "ELSE");
3461 set_basicblock_name(if_block, "if", label_id);
3462 LLVMBuildCondBr(ctx->builder, cond, if_block, flow->next_block);
3463 LLVMPositionBuilderAtEnd(ctx->builder, if_block);
3464 }
3465
3466 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
3467 int label_id)
3468 {
3469 LLVMValueRef cond = LLVMBuildFCmp(ctx->builder, LLVMRealUNE,
3470 value, ctx->f32_0, "");
3471 ac_build_ifcc(ctx, cond, label_id);
3472 }
3473
3474 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
3475 int label_id)
3476 {
3477 LLVMValueRef cond = LLVMBuildICmp(ctx->builder, LLVMIntNE,
3478 ac_to_integer(ctx, value),
3479 ctx->i32_0, "");
3480 ac_build_ifcc(ctx, cond, label_id);
3481 }
3482
3483 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
3484 const char *name)
3485 {
3486 LLVMBuilderRef builder = ac->builder;
3487 LLVMBasicBlockRef current_block = LLVMGetInsertBlock(builder);
3488 LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
3489 LLVMBasicBlockRef first_block = LLVMGetEntryBasicBlock(function);
3490 LLVMValueRef first_instr = LLVMGetFirstInstruction(first_block);
3491 LLVMBuilderRef first_builder = LLVMCreateBuilderInContext(ac->context);
3492 LLVMValueRef res;
3493
3494 if (first_instr) {
3495 LLVMPositionBuilderBefore(first_builder, first_instr);
3496 } else {
3497 LLVMPositionBuilderAtEnd(first_builder, first_block);
3498 }
3499
3500 res = LLVMBuildAlloca(first_builder, type, name);
3501 LLVMDisposeBuilder(first_builder);
3502 return res;
3503 }
3504
3505 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac,
3506 LLVMTypeRef type, const char *name)
3507 {
3508 LLVMValueRef ptr = ac_build_alloca_undef(ac, type, name);
3509 LLVMBuildStore(ac->builder, LLVMConstNull(type), ptr);
3510 return ptr;
3511 }
3512
3513 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
3514 LLVMTypeRef type)
3515 {
3516 int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
3517 return LLVMBuildBitCast(ctx->builder, ptr,
3518 LLVMPointerType(type, addr_space), "");
3519 }
3520
3521 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
3522 unsigned count)
3523 {
3524 unsigned num_components = ac_get_llvm_num_components(value);
3525 if (count == num_components)
3526 return value;
3527
3528 LLVMValueRef masks[MAX2(count, 2)];
3529 masks[0] = ctx->i32_0;
3530 masks[1] = ctx->i32_1;
3531 for (unsigned i = 2; i < count; i++)
3532 masks[i] = LLVMConstInt(ctx->i32, i, false);
3533
3534 if (count == 1)
3535 return LLVMBuildExtractElement(ctx->builder, value, masks[0],
3536 "");
3537
3538 LLVMValueRef swizzle = LLVMConstVector(masks, count);
3539 return LLVMBuildShuffleVector(ctx->builder, value, value, swizzle, "");
3540 }
3541
3542 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
3543 unsigned rshift, unsigned bitwidth)
3544 {
3545 LLVMValueRef value = param;
3546 if (rshift)
3547 value = LLVMBuildLShr(ctx->builder, value,
3548 LLVMConstInt(ctx->i32, rshift, false), "");
3549
3550 if (rshift + bitwidth < 32) {
3551 unsigned mask = (1 << bitwidth) - 1;
3552 value = LLVMBuildAnd(ctx->builder, value,
3553 LLVMConstInt(ctx->i32, mask, false), "");
3554 }
3555 return value;
3556 }
3557
3558 /* Adjust the sample index according to FMASK.
3559 *
3560 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3561 * which is the identity mapping. Each nibble says which physical sample
3562 * should be fetched to get that sample.
3563 *
3564 * For example, 0x11111100 means there are only 2 samples stored and
3565 * the second sample covers 3/4 of the pixel. When reading samples 0
3566 * and 1, return physical sample 0 (determined by the first two 0s
3567 * in FMASK), otherwise return physical sample 1.
3568 *
3569 * The sample index should be adjusted as follows:
3570 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
3571 */
3572 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
3573 LLVMValueRef *addr, bool is_array_tex)
3574 {
3575 struct ac_image_args fmask_load = {};
3576 fmask_load.opcode = ac_image_load;
3577 fmask_load.resource = fmask;
3578 fmask_load.dmask = 0xf;
3579 fmask_load.dim = is_array_tex ? ac_image_2darray : ac_image_2d;
3580 fmask_load.attributes = AC_FUNC_ATTR_READNONE;
3581
3582 fmask_load.coords[0] = addr[0];
3583 fmask_load.coords[1] = addr[1];
3584 if (is_array_tex)
3585 fmask_load.coords[2] = addr[2];
3586
3587 LLVMValueRef fmask_value = ac_build_image_opcode(ac, &fmask_load);
3588 fmask_value = LLVMBuildExtractElement(ac->builder, fmask_value,
3589 ac->i32_0, "");
3590
3591 /* Apply the formula. */
3592 unsigned sample_chan = is_array_tex ? 3 : 2;
3593 LLVMValueRef final_sample;
3594 final_sample = LLVMBuildMul(ac->builder, addr[sample_chan],
3595 LLVMConstInt(ac->i32, 4, 0), "");
3596 final_sample = LLVMBuildLShr(ac->builder, fmask_value, final_sample, "");
3597 /* Mask the sample index by 0x7, because 0x8 means an unknown value
3598 * with EQAA, so those will map to 0. */
3599 final_sample = LLVMBuildAnd(ac->builder, final_sample,
3600 LLVMConstInt(ac->i32, 0x7, 0), "");
3601
3602 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3603 * resource descriptor is 0 (invalid).
3604 */
3605 LLVMValueRef tmp;
3606 tmp = LLVMBuildBitCast(ac->builder, fmask, ac->v8i32, "");
3607 tmp = LLVMBuildExtractElement(ac->builder, tmp, ac->i32_1, "");
3608 tmp = LLVMBuildICmp(ac->builder, LLVMIntNE, tmp, ac->i32_0, "");
3609
3610 /* Replace the MSAA sample index. */
3611 addr[sample_chan] = LLVMBuildSelect(ac->builder, tmp, final_sample,
3612 addr[sample_chan], "");
3613 }
3614
3615 static LLVMValueRef
3616 _ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src,
3617 LLVMValueRef lane, bool with_opt_barrier)
3618 {
3619 LLVMTypeRef type = LLVMTypeOf(src);
3620 LLVMValueRef result;
3621
3622 if (with_opt_barrier)
3623 ac_build_optimization_barrier(ctx, &src);
3624
3625 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
3626 if (lane)
3627 lane = LLVMBuildZExt(ctx->builder, lane, ctx->i32, "");
3628
3629 result = ac_build_intrinsic(ctx,
3630 lane == NULL ? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
3631 ctx->i32, (LLVMValueRef []) { src, lane },
3632 lane == NULL ? 1 : 2,
3633 AC_FUNC_ATTR_READNONE |
3634 AC_FUNC_ATTR_CONVERGENT);
3635
3636 return LLVMBuildTrunc(ctx->builder, result, type, "");
3637 }
3638
3639 static LLVMValueRef
3640 ac_build_readlane_common(struct ac_llvm_context *ctx,
3641 LLVMValueRef src, LLVMValueRef lane,
3642 bool with_opt_barrier)
3643 {
3644 LLVMTypeRef src_type = LLVMTypeOf(src);
3645 src = ac_to_integer(ctx, src);
3646 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3647 LLVMValueRef ret;
3648
3649 if (bits > 32) {
3650 assert(bits % 32 == 0);
3651 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3652 LLVMValueRef src_vector =
3653 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3654 ret = LLVMGetUndef(vec_type);
3655 for (unsigned i = 0; i < bits / 32; i++) {
3656 LLVMValueRef ret_comp;
3657
3658 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3659 LLVMConstInt(ctx->i32, i, 0), "");
3660
3661 ret_comp = _ac_build_readlane(ctx, src, lane,
3662 with_opt_barrier);
3663
3664 ret = LLVMBuildInsertElement(ctx->builder, ret, ret_comp,
3665 LLVMConstInt(ctx->i32, i, 0), "");
3666 }
3667 } else {
3668 ret = _ac_build_readlane(ctx, src, lane, with_opt_barrier);
3669 }
3670
3671 if (LLVMGetTypeKind(src_type) == LLVMPointerTypeKind)
3672 return LLVMBuildIntToPtr(ctx->builder, ret, src_type, "");
3673 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3674 }
3675
3676 /**
3677 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
3678 *
3679 * The optimization barrier is not needed if the value is the same in all lanes
3680 * or if this is called in the outermost block.
3681 *
3682 * @param ctx
3683 * @param src
3684 * @param lane - id of the lane or NULL for the first active lane
3685 * @return value of the lane
3686 */
3687 LLVMValueRef ac_build_readlane_no_opt_barrier(struct ac_llvm_context *ctx,
3688 LLVMValueRef src, LLVMValueRef lane)
3689 {
3690 return ac_build_readlane_common(ctx, src, lane, false);
3691 }
3692
3693
3694 LLVMValueRef
3695 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
3696 {
3697 return ac_build_readlane_common(ctx, src, lane, true);
3698 }
3699
3700 LLVMValueRef
3701 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane)
3702 {
3703 return ac_build_intrinsic(ctx, "llvm.amdgcn.writelane", ctx->i32,
3704 (LLVMValueRef []) {value, lane, src}, 3,
3705 AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3706 }
3707
3708 LLVMValueRef
3709 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask)
3710 {
3711 if (ctx->wave_size == 32) {
3712 return ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
3713 (LLVMValueRef []) { mask, ctx->i32_0 },
3714 2, AC_FUNC_ATTR_READNONE);
3715 }
3716 LLVMValueRef mask_vec = LLVMBuildBitCast(ctx->builder, mask, ctx->v2i32, "");
3717 LLVMValueRef mask_lo = LLVMBuildExtractElement(ctx->builder, mask_vec,
3718 ctx->i32_0, "");
3719 LLVMValueRef mask_hi = LLVMBuildExtractElement(ctx->builder, mask_vec,
3720 ctx->i32_1, "");
3721 LLVMValueRef val =
3722 ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
3723 (LLVMValueRef []) { mask_lo, ctx->i32_0 },
3724 2, AC_FUNC_ATTR_READNONE);
3725 val = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi", ctx->i32,
3726 (LLVMValueRef []) { mask_hi, val },
3727 2, AC_FUNC_ATTR_READNONE);
3728 return val;
3729 }
3730
3731 enum dpp_ctrl {
3732 _dpp_quad_perm = 0x000,
3733 _dpp_row_sl = 0x100,
3734 _dpp_row_sr = 0x110,
3735 _dpp_row_rr = 0x120,
3736 dpp_wf_sl1 = 0x130,
3737 dpp_wf_rl1 = 0x134,
3738 dpp_wf_sr1 = 0x138,
3739 dpp_wf_rr1 = 0x13C,
3740 dpp_row_mirror = 0x140,
3741 dpp_row_half_mirror = 0x141,
3742 dpp_row_bcast15 = 0x142,
3743 dpp_row_bcast31 = 0x143
3744 };
3745
3746 static inline enum dpp_ctrl
3747 dpp_quad_perm(unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
3748 {
3749 assert(lane0 < 4 && lane1 < 4 && lane2 < 4 && lane3 < 4);
3750 return _dpp_quad_perm | lane0 | (lane1 << 2) | (lane2 << 4) | (lane3 << 6);
3751 }
3752
3753 static inline enum dpp_ctrl
3754 dpp_row_sl(unsigned amount)
3755 {
3756 assert(amount > 0 && amount < 16);
3757 return _dpp_row_sl | amount;
3758 }
3759
3760 static inline enum dpp_ctrl
3761 dpp_row_sr(unsigned amount)
3762 {
3763 assert(amount > 0 && amount < 16);
3764 return _dpp_row_sr | amount;
3765 }
3766
3767 static LLVMValueRef
3768 _ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3769 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3770 bool bound_ctrl)
3771 {
3772 LLVMTypeRef type = LLVMTypeOf(src);
3773 LLVMValueRef res;
3774
3775 old = LLVMBuildZExt(ctx->builder, old, ctx->i32, "");
3776 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
3777
3778 res = ac_build_intrinsic(ctx, "llvm.amdgcn.update.dpp.i32", ctx->i32,
3779 (LLVMValueRef[]) {
3780 old, src,
3781 LLVMConstInt(ctx->i32, dpp_ctrl, 0),
3782 LLVMConstInt(ctx->i32, row_mask, 0),
3783 LLVMConstInt(ctx->i32, bank_mask, 0),
3784 LLVMConstInt(ctx->i1, bound_ctrl, 0) },
3785 6, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3786
3787 return LLVMBuildTrunc(ctx->builder, res, type, "");
3788 }
3789
3790 static LLVMValueRef
3791 ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3792 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3793 bool bound_ctrl)
3794 {
3795 LLVMTypeRef src_type = LLVMTypeOf(src);
3796 src = ac_to_integer(ctx, src);
3797 old = ac_to_integer(ctx, old);
3798 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3799 LLVMValueRef ret;
3800 if (bits > 32) {
3801 assert(bits % 32 == 0);
3802 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3803 LLVMValueRef src_vector =
3804 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3805 LLVMValueRef old_vector =
3806 LLVMBuildBitCast(ctx->builder, old, vec_type, "");
3807 ret = LLVMGetUndef(vec_type);
3808 for (unsigned i = 0; i < bits / 32; i++) {
3809 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3810 LLVMConstInt(ctx->i32, i,
3811 0), "");
3812 old = LLVMBuildExtractElement(ctx->builder, old_vector,
3813 LLVMConstInt(ctx->i32, i,
3814 0), "");
3815 LLVMValueRef ret_comp = _ac_build_dpp(ctx, old, src,
3816 dpp_ctrl,
3817 row_mask,
3818 bank_mask,
3819 bound_ctrl);
3820 ret = LLVMBuildInsertElement(ctx->builder, ret,
3821 ret_comp,
3822 LLVMConstInt(ctx->i32, i,
3823 0), "");
3824 }
3825 } else {
3826 ret = _ac_build_dpp(ctx, old, src, dpp_ctrl, row_mask,
3827 bank_mask, bound_ctrl);
3828 }
3829 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3830 }
3831
3832 static LLVMValueRef
3833 _ac_build_permlane16(struct ac_llvm_context *ctx, LLVMValueRef src, uint64_t sel,
3834 bool exchange_rows, bool bound_ctrl)
3835 {
3836 LLVMTypeRef type = LLVMTypeOf(src);
3837 LLVMValueRef result;
3838
3839 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
3840
3841 LLVMValueRef args[6] = {
3842 src,
3843 src,
3844 LLVMConstInt(ctx->i32, sel, false),
3845 LLVMConstInt(ctx->i32, sel >> 32, false),
3846 ctx->i1true, /* fi */
3847 bound_ctrl ? ctx->i1true : ctx->i1false,
3848 };
3849
3850 result = ac_build_intrinsic(ctx, exchange_rows ? "llvm.amdgcn.permlanex16"
3851 : "llvm.amdgcn.permlane16",
3852 ctx->i32, args, 6,
3853 AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3854
3855 return LLVMBuildTrunc(ctx->builder, result, type, "");
3856 }
3857
3858 static LLVMValueRef
3859 ac_build_permlane16(struct ac_llvm_context *ctx, LLVMValueRef src, uint64_t sel,
3860 bool exchange_rows, bool bound_ctrl)
3861 {
3862 LLVMTypeRef src_type = LLVMTypeOf(src);
3863 src = ac_to_integer(ctx, src);
3864 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3865 LLVMValueRef ret;
3866 if (bits > 32) {
3867 assert(bits % 32 == 0);
3868 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3869 LLVMValueRef src_vector =
3870 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3871 ret = LLVMGetUndef(vec_type);
3872 for (unsigned i = 0; i < bits / 32; i++) {
3873 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3874 LLVMConstInt(ctx->i32, i,
3875 0), "");
3876 LLVMValueRef ret_comp =
3877 _ac_build_permlane16(ctx, src, sel,
3878 exchange_rows,
3879 bound_ctrl);
3880 ret = LLVMBuildInsertElement(ctx->builder, ret,
3881 ret_comp,
3882 LLVMConstInt(ctx->i32, i,
3883 0), "");
3884 }
3885 } else {
3886 ret = _ac_build_permlane16(ctx, src, sel, exchange_rows,
3887 bound_ctrl);
3888 }
3889 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3890 }
3891
3892 static inline unsigned
3893 ds_pattern_bitmode(unsigned and_mask, unsigned or_mask, unsigned xor_mask)
3894 {
3895 assert(and_mask < 32 && or_mask < 32 && xor_mask < 32);
3896 return and_mask | (or_mask << 5) | (xor_mask << 10);
3897 }
3898
3899 static LLVMValueRef
3900 _ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
3901 {
3902 LLVMTypeRef src_type = LLVMTypeOf(src);
3903 LLVMValueRef ret;
3904
3905 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
3906
3907 ret = ac_build_intrinsic(ctx, "llvm.amdgcn.ds.swizzle", ctx->i32,
3908 (LLVMValueRef []) {
3909 src, LLVMConstInt(ctx->i32, mask, 0) },
3910 2, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3911
3912 return LLVMBuildTrunc(ctx->builder, ret, src_type, "");
3913 }
3914
3915 LLVMValueRef
3916 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
3917 {
3918 LLVMTypeRef src_type = LLVMTypeOf(src);
3919 src = ac_to_integer(ctx, src);
3920 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3921 LLVMValueRef ret;
3922 if (bits > 32) {
3923 assert(bits % 32 == 0);
3924 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3925 LLVMValueRef src_vector =
3926 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3927 ret = LLVMGetUndef(vec_type);
3928 for (unsigned i = 0; i < bits / 32; i++) {
3929 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3930 LLVMConstInt(ctx->i32, i,
3931 0), "");
3932 LLVMValueRef ret_comp = _ac_build_ds_swizzle(ctx, src,
3933 mask);
3934 ret = LLVMBuildInsertElement(ctx->builder, ret,
3935 ret_comp,
3936 LLVMConstInt(ctx->i32, i,
3937 0), "");
3938 }
3939 } else {
3940 ret = _ac_build_ds_swizzle(ctx, src, mask);
3941 }
3942 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3943 }
3944
3945 static LLVMValueRef
3946 ac_build_wwm(struct ac_llvm_context *ctx, LLVMValueRef src)
3947 {
3948 LLVMTypeRef src_type = LLVMTypeOf(src);
3949 unsigned bitsize = ac_get_elem_bits(ctx, src_type);
3950 char name[32], type[8];
3951 LLVMValueRef ret;
3952
3953 src = ac_to_integer(ctx, src);
3954
3955 if (bitsize < 32)
3956 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
3957
3958 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
3959 snprintf(name, sizeof(name), "llvm.amdgcn.wwm.%s", type);
3960 ret = ac_build_intrinsic(ctx, name, LLVMTypeOf(src),
3961 (LLVMValueRef []) { src }, 1,
3962 AC_FUNC_ATTR_READNONE);
3963
3964 if (bitsize < 32)
3965 ret = LLVMBuildTrunc(ctx->builder, ret,
3966 ac_to_integer_type(ctx, src_type), "");
3967
3968 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3969 }
3970
3971 static LLVMValueRef
3972 ac_build_set_inactive(struct ac_llvm_context *ctx, LLVMValueRef src,
3973 LLVMValueRef inactive)
3974 {
3975 char name[33], type[8];
3976 LLVMTypeRef src_type = LLVMTypeOf(src);
3977 unsigned bitsize = ac_get_elem_bits(ctx, src_type);
3978 src = ac_to_integer(ctx, src);
3979 inactive = ac_to_integer(ctx, inactive);
3980
3981 if (bitsize < 32) {
3982 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
3983 inactive = LLVMBuildZExt(ctx->builder, inactive, ctx->i32, "");
3984 }
3985
3986 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
3987 snprintf(name, sizeof(name), "llvm.amdgcn.set.inactive.%s", type);
3988 LLVMValueRef ret =
3989 ac_build_intrinsic(ctx, name,
3990 LLVMTypeOf(src), (LLVMValueRef []) {
3991 src, inactive }, 2,
3992 AC_FUNC_ATTR_READNONE |
3993 AC_FUNC_ATTR_CONVERGENT);
3994 if (bitsize < 32)
3995 ret = LLVMBuildTrunc(ctx->builder, ret, src_type, "");
3996
3997 return ret;
3998 }
3999
4000 static LLVMValueRef
4001 get_reduction_identity(struct ac_llvm_context *ctx, nir_op op, unsigned type_size)
4002 {
4003 if (type_size == 1) {
4004 switch (op) {
4005 case nir_op_iadd: return ctx->i8_0;
4006 case nir_op_imul: return ctx->i8_1;
4007 case nir_op_imin: return LLVMConstInt(ctx->i8, INT8_MAX, 0);
4008 case nir_op_umin: return LLVMConstInt(ctx->i8, UINT8_MAX, 0);
4009 case nir_op_imax: return LLVMConstInt(ctx->i8, INT8_MIN, 0);
4010 case nir_op_umax: return ctx->i8_0;
4011 case nir_op_iand: return LLVMConstInt(ctx->i8, -1, 0);
4012 case nir_op_ior: return ctx->i8_0;
4013 case nir_op_ixor: return ctx->i8_0;
4014 default:
4015 unreachable("bad reduction intrinsic");
4016 }
4017 } else if (type_size == 2) {
4018 switch (op) {
4019 case nir_op_iadd: return ctx->i16_0;
4020 case nir_op_fadd: return ctx->f16_0;
4021 case nir_op_imul: return ctx->i16_1;
4022 case nir_op_fmul: return ctx->f16_1;
4023 case nir_op_imin: return LLVMConstInt(ctx->i16, INT16_MAX, 0);
4024 case nir_op_umin: return LLVMConstInt(ctx->i16, UINT16_MAX, 0);
4025 case nir_op_fmin: return LLVMConstReal(ctx->f16, INFINITY);
4026 case nir_op_imax: return LLVMConstInt(ctx->i16, INT16_MIN, 0);
4027 case nir_op_umax: return ctx->i16_0;
4028 case nir_op_fmax: return LLVMConstReal(ctx->f16, -INFINITY);
4029 case nir_op_iand: return LLVMConstInt(ctx->i16, -1, 0);
4030 case nir_op_ior: return ctx->i16_0;
4031 case nir_op_ixor: return ctx->i16_0;
4032 default:
4033 unreachable("bad reduction intrinsic");
4034 }
4035 } else if (type_size == 4) {
4036 switch (op) {
4037 case nir_op_iadd: return ctx->i32_0;
4038 case nir_op_fadd: return ctx->f32_0;
4039 case nir_op_imul: return ctx->i32_1;
4040 case nir_op_fmul: return ctx->f32_1;
4041 case nir_op_imin: return LLVMConstInt(ctx->i32, INT32_MAX, 0);
4042 case nir_op_umin: return LLVMConstInt(ctx->i32, UINT32_MAX, 0);
4043 case nir_op_fmin: return LLVMConstReal(ctx->f32, INFINITY);
4044 case nir_op_imax: return LLVMConstInt(ctx->i32, INT32_MIN, 0);
4045 case nir_op_umax: return ctx->i32_0;
4046 case nir_op_fmax: return LLVMConstReal(ctx->f32, -INFINITY);
4047 case nir_op_iand: return LLVMConstInt(ctx->i32, -1, 0);
4048 case nir_op_ior: return ctx->i32_0;
4049 case nir_op_ixor: return ctx->i32_0;
4050 default:
4051 unreachable("bad reduction intrinsic");
4052 }
4053 } else { /* type_size == 64bit */
4054 switch (op) {
4055 case nir_op_iadd: return ctx->i64_0;
4056 case nir_op_fadd: return ctx->f64_0;
4057 case nir_op_imul: return ctx->i64_1;
4058 case nir_op_fmul: return ctx->f64_1;
4059 case nir_op_imin: return LLVMConstInt(ctx->i64, INT64_MAX, 0);
4060 case nir_op_umin: return LLVMConstInt(ctx->i64, UINT64_MAX, 0);
4061 case nir_op_fmin: return LLVMConstReal(ctx->f64, INFINITY);
4062 case nir_op_imax: return LLVMConstInt(ctx->i64, INT64_MIN, 0);
4063 case nir_op_umax: return ctx->i64_0;
4064 case nir_op_fmax: return LLVMConstReal(ctx->f64, -INFINITY);
4065 case nir_op_iand: return LLVMConstInt(ctx->i64, -1, 0);
4066 case nir_op_ior: return ctx->i64_0;
4067 case nir_op_ixor: return ctx->i64_0;
4068 default:
4069 unreachable("bad reduction intrinsic");
4070 }
4071 }
4072 }
4073
4074 static LLVMValueRef
4075 ac_build_alu_op(struct ac_llvm_context *ctx, LLVMValueRef lhs, LLVMValueRef rhs, nir_op op)
4076 {
4077 bool _64bit = ac_get_type_size(LLVMTypeOf(lhs)) == 8;
4078 bool _32bit = ac_get_type_size(LLVMTypeOf(lhs)) == 4;
4079 switch (op) {
4080 case nir_op_iadd: return LLVMBuildAdd(ctx->builder, lhs, rhs, "");
4081 case nir_op_fadd: return LLVMBuildFAdd(ctx->builder, lhs, rhs, "");
4082 case nir_op_imul: return LLVMBuildMul(ctx->builder, lhs, rhs, "");
4083 case nir_op_fmul: return LLVMBuildFMul(ctx->builder, lhs, rhs, "");
4084 case nir_op_imin: return LLVMBuildSelect(ctx->builder,
4085 LLVMBuildICmp(ctx->builder, LLVMIntSLT, lhs, rhs, ""),
4086 lhs, rhs, "");
4087 case nir_op_umin: return LLVMBuildSelect(ctx->builder,
4088 LLVMBuildICmp(ctx->builder, LLVMIntULT, lhs, rhs, ""),
4089 lhs, rhs, "");
4090 case nir_op_fmin: return ac_build_intrinsic(ctx,
4091 _64bit ? "llvm.minnum.f64" : _32bit ? "llvm.minnum.f32" : "llvm.minnum.f16",
4092 _64bit ? ctx->f64 : _32bit ? ctx->f32 : ctx->f16,
4093 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
4094 case nir_op_imax: return LLVMBuildSelect(ctx->builder,
4095 LLVMBuildICmp(ctx->builder, LLVMIntSGT, lhs, rhs, ""),
4096 lhs, rhs, "");
4097 case nir_op_umax: return LLVMBuildSelect(ctx->builder,
4098 LLVMBuildICmp(ctx->builder, LLVMIntUGT, lhs, rhs, ""),
4099 lhs, rhs, "");
4100 case nir_op_fmax: return ac_build_intrinsic(ctx,
4101 _64bit ? "llvm.maxnum.f64" : _32bit ? "llvm.maxnum.f32" : "llvm.maxnum.f16",
4102 _64bit ? ctx->f64 : _32bit ? ctx->f32 : ctx->f16,
4103 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
4104 case nir_op_iand: return LLVMBuildAnd(ctx->builder, lhs, rhs, "");
4105 case nir_op_ior: return LLVMBuildOr(ctx->builder, lhs, rhs, "");
4106 case nir_op_ixor: return LLVMBuildXor(ctx->builder, lhs, rhs, "");
4107 default:
4108 unreachable("bad reduction intrinsic");
4109 }
4110 }
4111
4112 /**
4113 * \param src The value to shift.
4114 * \param identity The value to use the first lane.
4115 * \param maxprefix specifies that the result only needs to be correct for a
4116 * prefix of this many threads
4117 * \return src, shifted 1 lane up, and identity shifted into lane 0.
4118 */
4119 static LLVMValueRef
4120 ac_wavefront_shift_right_1(struct ac_llvm_context *ctx, LLVMValueRef src,
4121 LLVMValueRef identity, unsigned maxprefix)
4122 {
4123 if (ctx->chip_class >= GFX10) {
4124 /* wavefront shift_right by 1 on GFX10 (emulate dpp_wf_sr1) */
4125 LLVMValueRef active, tmp1, tmp2;
4126 LLVMValueRef tid = ac_get_thread_id(ctx);
4127
4128 tmp1 = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
4129
4130 tmp2 = ac_build_permlane16(ctx, src, (uint64_t)~0, true, false);
4131
4132 if (maxprefix > 32) {
4133 active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid,
4134 LLVMConstInt(ctx->i32, 32, false), "");
4135
4136 tmp2 = LLVMBuildSelect(ctx->builder, active,
4137 ac_build_readlane(ctx, src,
4138 LLVMConstInt(ctx->i32, 31, false)),
4139 tmp2, "");
4140
4141 active = LLVMBuildOr(ctx->builder, active,
4142 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
4143 LLVMBuildAnd(ctx->builder, tid,
4144 LLVMConstInt(ctx->i32, 0x1f, false), ""),
4145 LLVMConstInt(ctx->i32, 0x10, false), ""), "");
4146 return LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
4147 } else if (maxprefix > 16) {
4148 active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid,
4149 LLVMConstInt(ctx->i32, 16, false), "");
4150
4151 return LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
4152 }
4153 } else if (ctx->chip_class >= GFX8) {
4154 return ac_build_dpp(ctx, identity, src, dpp_wf_sr1, 0xf, 0xf, false);
4155 }
4156
4157 /* wavefront shift_right by 1 on SI/CI */
4158 LLVMValueRef active, tmp1, tmp2;
4159 LLVMValueRef tid = ac_get_thread_id(ctx);
4160 tmp1 = ac_build_ds_swizzle(ctx, src, (1 << 15) | dpp_quad_perm(0, 0, 1, 2));
4161 tmp2 = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x18, 0x03, 0x00));
4162 active = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
4163 LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 0x7, 0), ""),
4164 LLVMConstInt(ctx->i32, 0x4, 0), "");
4165 tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
4166 tmp2 = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x10, 0x07, 0x00));
4167 active = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
4168 LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 0xf, 0), ""),
4169 LLVMConstInt(ctx->i32, 0x8, 0), "");
4170 tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
4171 tmp2 = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x00, 0x0f, 0x00));
4172 active = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
4173 LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 0x1f, 0), ""),
4174 LLVMConstInt(ctx->i32, 0x10, 0), "");
4175 tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
4176 tmp2 = ac_build_readlane(ctx, src, LLVMConstInt(ctx->i32, 31, 0));
4177 active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid, LLVMConstInt(ctx->i32, 32, 0), "");
4178 tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
4179 active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid, LLVMConstInt(ctx->i32, 0, 0), "");
4180 return LLVMBuildSelect(ctx->builder, active, identity, tmp1, "");
4181 }
4182
4183 /**
4184 * \param maxprefix specifies that the result only needs to be correct for a
4185 * prefix of this many threads
4186 */
4187 static LLVMValueRef
4188 ac_build_scan(struct ac_llvm_context *ctx, nir_op op, LLVMValueRef src, LLVMValueRef identity,
4189 unsigned maxprefix, bool inclusive)
4190 {
4191 LLVMValueRef result, tmp;
4192
4193 if (!inclusive)
4194 src = ac_wavefront_shift_right_1(ctx, src, identity, maxprefix);
4195
4196 result = src;
4197
4198 if (ctx->chip_class <= GFX7) {
4199 assert(maxprefix == 64);
4200 LLVMValueRef tid = ac_get_thread_id(ctx);
4201 LLVMValueRef active;
4202 tmp = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x1e, 0x00, 0x00));
4203 active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
4204 LLVMBuildAnd(ctx->builder, tid, ctx->i32_1, ""),
4205 ctx->i32_0, "");
4206 tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
4207 result = ac_build_alu_op(ctx, result, tmp, op);
4208 tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1c, 0x01, 0x00));
4209 active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
4210 LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 2, 0), ""),
4211 ctx->i32_0, "");
4212 tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
4213 result = ac_build_alu_op(ctx, result, tmp, op);
4214 tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x18, 0x03, 0x00));
4215 active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
4216 LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 4, 0), ""),
4217 ctx->i32_0, "");
4218 tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
4219 result = ac_build_alu_op(ctx, result, tmp, op);
4220 tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x10, 0x07, 0x00));
4221 active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
4222 LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 8, 0), ""),
4223 ctx->i32_0, "");
4224 tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
4225 result = ac_build_alu_op(ctx, result, tmp, op);
4226 tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x00, 0x0f, 0x00));
4227 active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
4228 LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 16, 0), ""),
4229 ctx->i32_0, "");
4230 tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
4231 result = ac_build_alu_op(ctx, result, tmp, op);
4232 tmp = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, 0));
4233 active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
4234 LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 32, 0), ""),
4235 ctx->i32_0, "");
4236 tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
4237 result = ac_build_alu_op(ctx, result, tmp, op);
4238 return result;
4239 }
4240
4241 if (maxprefix <= 1)
4242 return result;
4243 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
4244 result = ac_build_alu_op(ctx, result, tmp, op);
4245 if (maxprefix <= 2)
4246 return result;
4247 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(2), 0xf, 0xf, false);
4248 result = ac_build_alu_op(ctx, result, tmp, op);
4249 if (maxprefix <= 3)
4250 return result;
4251 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(3), 0xf, 0xf, false);
4252 result = ac_build_alu_op(ctx, result, tmp, op);
4253 if (maxprefix <= 4)
4254 return result;
4255 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(4), 0xf, 0xe, false);
4256 result = ac_build_alu_op(ctx, result, tmp, op);
4257 if (maxprefix <= 8)
4258 return result;
4259 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(8), 0xf, 0xc, false);
4260 result = ac_build_alu_op(ctx, result, tmp, op);
4261 if (maxprefix <= 16)
4262 return result;
4263
4264 if (ctx->chip_class >= GFX10) {
4265 LLVMValueRef tid = ac_get_thread_id(ctx);
4266 LLVMValueRef active;
4267
4268 tmp = ac_build_permlane16(ctx, result, ~(uint64_t)0, true, false);
4269
4270 active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
4271 LLVMBuildAnd(ctx->builder, tid,
4272 LLVMConstInt(ctx->i32, 16, false), ""),
4273 ctx->i32_0, "");
4274
4275 tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
4276
4277 result = ac_build_alu_op(ctx, result, tmp, op);
4278
4279 if (maxprefix <= 32)
4280 return result;
4281
4282 tmp = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, false));
4283
4284 active = LLVMBuildICmp(ctx->builder, LLVMIntUGE, tid,
4285 LLVMConstInt(ctx->i32, 32, false), "");
4286
4287 tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
4288
4289 result = ac_build_alu_op(ctx, result, tmp, op);
4290 return result;
4291 }
4292
4293 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
4294 result = ac_build_alu_op(ctx, result, tmp, op);
4295 if (maxprefix <= 32)
4296 return result;
4297 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
4298 result = ac_build_alu_op(ctx, result, tmp, op);
4299 return result;
4300 }
4301
4302 LLVMValueRef
4303 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
4304 {
4305 LLVMValueRef result;
4306
4307 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
4308 LLVMBuilderRef builder = ctx->builder;
4309 src = LLVMBuildZExt(builder, src, ctx->i32, "");
4310 result = ac_build_ballot(ctx, src);
4311 result = ac_build_mbcnt(ctx, result);
4312 result = LLVMBuildAdd(builder, result, src, "");
4313 return result;
4314 }
4315
4316 ac_build_optimization_barrier(ctx, &src);
4317
4318 LLVMValueRef identity =
4319 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
4320 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
4321 LLVMTypeOf(identity), "");
4322 result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, true);
4323
4324 return ac_build_wwm(ctx, result);
4325 }
4326
4327 LLVMValueRef
4328 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
4329 {
4330 LLVMValueRef result;
4331
4332 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
4333 LLVMBuilderRef builder = ctx->builder;
4334 src = LLVMBuildZExt(builder, src, ctx->i32, "");
4335 result = ac_build_ballot(ctx, src);
4336 result = ac_build_mbcnt(ctx, result);
4337 return result;
4338 }
4339
4340 ac_build_optimization_barrier(ctx, &src);
4341
4342 LLVMValueRef identity =
4343 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
4344 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
4345 LLVMTypeOf(identity), "");
4346 result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, false);
4347
4348 return ac_build_wwm(ctx, result);
4349 }
4350
4351 LLVMValueRef
4352 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size)
4353 {
4354 if (cluster_size == 1) return src;
4355 ac_build_optimization_barrier(ctx, &src);
4356 LLVMValueRef result, swap;
4357 LLVMValueRef identity = get_reduction_identity(ctx, op,
4358 ac_get_type_size(LLVMTypeOf(src)));
4359 result = LLVMBuildBitCast(ctx->builder,
4360 ac_build_set_inactive(ctx, src, identity),
4361 LLVMTypeOf(identity), "");
4362 swap = ac_build_quad_swizzle(ctx, result, 1, 0, 3, 2);
4363 result = ac_build_alu_op(ctx, result, swap, op);
4364 if (cluster_size == 2) return ac_build_wwm(ctx, result);
4365
4366 swap = ac_build_quad_swizzle(ctx, result, 2, 3, 0, 1);
4367 result = ac_build_alu_op(ctx, result, swap, op);
4368 if (cluster_size == 4) return ac_build_wwm(ctx, result);
4369
4370 if (ctx->chip_class >= GFX8)
4371 swap = ac_build_dpp(ctx, identity, result, dpp_row_half_mirror, 0xf, 0xf, false);
4372 else
4373 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x04));
4374 result = ac_build_alu_op(ctx, result, swap, op);
4375 if (cluster_size == 8) return ac_build_wwm(ctx, result);
4376
4377 if (ctx->chip_class >= GFX8)
4378 swap = ac_build_dpp(ctx, identity, result, dpp_row_mirror, 0xf, 0xf, false);
4379 else
4380 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x08));
4381 result = ac_build_alu_op(ctx, result, swap, op);
4382 if (cluster_size == 16) return ac_build_wwm(ctx, result);
4383
4384 if (ctx->chip_class >= GFX10)
4385 swap = ac_build_permlane16(ctx, result, 0, true, false);
4386 else if (ctx->chip_class >= GFX8 && cluster_size != 32)
4387 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
4388 else
4389 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x10));
4390 result = ac_build_alu_op(ctx, result, swap, op);
4391 if (cluster_size == 32) return ac_build_wwm(ctx, result);
4392
4393 if (ctx->chip_class >= GFX8) {
4394 if (ctx->wave_size == 64) {
4395 if (ctx->chip_class >= GFX10)
4396 swap = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, false));
4397 else
4398 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
4399 result = ac_build_alu_op(ctx, result, swap, op);
4400 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 63, 0));
4401 }
4402
4403 return ac_build_wwm(ctx, result);
4404 } else {
4405 swap = ac_build_readlane(ctx, result, ctx->i32_0);
4406 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 32, 0));
4407 result = ac_build_alu_op(ctx, result, swap, op);
4408 return ac_build_wwm(ctx, result);
4409 }
4410 }
4411
4412 /**
4413 * "Top half" of a scan that reduces per-wave values across an entire
4414 * workgroup.
4415 *
4416 * The source value must be present in the highest lane of the wave, and the
4417 * highest lane must be live.
4418 */
4419 void
4420 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4421 {
4422 if (ws->maxwaves <= 1)
4423 return;
4424
4425 const LLVMValueRef last_lane = LLVMConstInt(ctx->i32, ctx->wave_size - 1, false);
4426 LLVMBuilderRef builder = ctx->builder;
4427 LLVMValueRef tid = ac_get_thread_id(ctx);
4428 LLVMValueRef tmp;
4429
4430 tmp = LLVMBuildICmp(builder, LLVMIntEQ, tid, last_lane, "");
4431 ac_build_ifcc(ctx, tmp, 1000);
4432 LLVMBuildStore(builder, ws->src, LLVMBuildGEP(builder, ws->scratch, &ws->waveidx, 1, ""));
4433 ac_build_endif(ctx, 1000);
4434 }
4435
4436 /**
4437 * "Bottom half" of a scan that reduces per-wave values across an entire
4438 * workgroup.
4439 *
4440 * The caller must place a barrier between the top and bottom halves.
4441 */
4442 void
4443 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4444 {
4445 const LLVMTypeRef type = LLVMTypeOf(ws->src);
4446 const LLVMValueRef identity =
4447 get_reduction_identity(ctx, ws->op, ac_get_type_size(type));
4448
4449 if (ws->maxwaves <= 1) {
4450 ws->result_reduce = ws->src;
4451 ws->result_inclusive = ws->src;
4452 ws->result_exclusive = identity;
4453 return;
4454 }
4455 assert(ws->maxwaves <= 32);
4456
4457 LLVMBuilderRef builder = ctx->builder;
4458 LLVMValueRef tid = ac_get_thread_id(ctx);
4459 LLVMBasicBlockRef bbs[2];
4460 LLVMValueRef phivalues_scan[2];
4461 LLVMValueRef tmp, tmp2;
4462
4463 bbs[0] = LLVMGetInsertBlock(builder);
4464 phivalues_scan[0] = LLVMGetUndef(type);
4465
4466 if (ws->enable_reduce)
4467 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->numwaves, "");
4468 else if (ws->enable_inclusive)
4469 tmp = LLVMBuildICmp(builder, LLVMIntULE, tid, ws->waveidx, "");
4470 else
4471 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->waveidx, "");
4472 ac_build_ifcc(ctx, tmp, 1001);
4473 {
4474 tmp = LLVMBuildLoad(builder, LLVMBuildGEP(builder, ws->scratch, &tid, 1, ""), "");
4475
4476 ac_build_optimization_barrier(ctx, &tmp);
4477
4478 bbs[1] = LLVMGetInsertBlock(builder);
4479 phivalues_scan[1] = ac_build_scan(ctx, ws->op, tmp, identity, ws->maxwaves, true);
4480 }
4481 ac_build_endif(ctx, 1001);
4482
4483 const LLVMValueRef scan = ac_build_phi(ctx, type, 2, phivalues_scan, bbs);
4484
4485 if (ws->enable_reduce) {
4486 tmp = LLVMBuildSub(builder, ws->numwaves, ctx->i32_1, "");
4487 ws->result_reduce = ac_build_readlane(ctx, scan, tmp);
4488 }
4489 if (ws->enable_inclusive)
4490 ws->result_inclusive = ac_build_readlane(ctx, scan, ws->waveidx);
4491 if (ws->enable_exclusive) {
4492 tmp = LLVMBuildSub(builder, ws->waveidx, ctx->i32_1, "");
4493 tmp = ac_build_readlane(ctx, scan, tmp);
4494 tmp2 = LLVMBuildICmp(builder, LLVMIntEQ, ws->waveidx, ctx->i32_0, "");
4495 ws->result_exclusive = LLVMBuildSelect(builder, tmp2, identity, tmp, "");
4496 }
4497 }
4498
4499 /**
4500 * Inclusive scan of a per-wave value across an entire workgroup.
4501 *
4502 * This implies an s_barrier instruction.
4503 *
4504 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
4505 * of the workgroup are live. (This requirement cannot easily be relaxed in a
4506 * useful manner because of the barrier in the algorithm.)
4507 */
4508 void
4509 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4510 {
4511 ac_build_wg_wavescan_top(ctx, ws);
4512 ac_build_s_barrier(ctx);
4513 ac_build_wg_wavescan_bottom(ctx, ws);
4514 }
4515
4516 /**
4517 * "Top half" of a scan that reduces per-thread values across an entire
4518 * workgroup.
4519 *
4520 * All lanes must be active when this code runs.
4521 */
4522 void
4523 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4524 {
4525 if (ws->enable_exclusive) {
4526 ws->extra = ac_build_exclusive_scan(ctx, ws->src, ws->op);
4527 if (LLVMTypeOf(ws->src) == ctx->i1 && ws->op == nir_op_iadd)
4528 ws->src = LLVMBuildZExt(ctx->builder, ws->src, ctx->i32, "");
4529 ws->src = ac_build_alu_op(ctx, ws->extra, ws->src, ws->op);
4530 } else {
4531 ws->src = ac_build_inclusive_scan(ctx, ws->src, ws->op);
4532 }
4533
4534 bool enable_inclusive = ws->enable_inclusive;
4535 bool enable_exclusive = ws->enable_exclusive;
4536 ws->enable_inclusive = false;
4537 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
4538 ac_build_wg_wavescan_top(ctx, ws);
4539 ws->enable_inclusive = enable_inclusive;
4540 ws->enable_exclusive = enable_exclusive;
4541 }
4542
4543 /**
4544 * "Bottom half" of a scan that reduces per-thread values across an entire
4545 * workgroup.
4546 *
4547 * The caller must place a barrier between the top and bottom halves.
4548 */
4549 void
4550 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4551 {
4552 bool enable_inclusive = ws->enable_inclusive;
4553 bool enable_exclusive = ws->enable_exclusive;
4554 ws->enable_inclusive = false;
4555 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
4556 ac_build_wg_wavescan_bottom(ctx, ws);
4557 ws->enable_inclusive = enable_inclusive;
4558 ws->enable_exclusive = enable_exclusive;
4559
4560 /* ws->result_reduce is already the correct value */
4561 if (ws->enable_inclusive)
4562 ws->result_inclusive = ac_build_alu_op(ctx, ws->result_inclusive, ws->src, ws->op);
4563 if (ws->enable_exclusive)
4564 ws->result_exclusive = ac_build_alu_op(ctx, ws->result_exclusive, ws->extra, ws->op);
4565 }
4566
4567 /**
4568 * A scan that reduces per-thread values across an entire workgroup.
4569 *
4570 * The caller must ensure that all lanes are active when this code runs
4571 * (WWM is insufficient!), because there is an implied barrier.
4572 */
4573 void
4574 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4575 {
4576 ac_build_wg_scan_top(ctx, ws);
4577 ac_build_s_barrier(ctx);
4578 ac_build_wg_scan_bottom(ctx, ws);
4579 }
4580
4581 LLVMValueRef
4582 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
4583 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
4584 {
4585 unsigned mask = dpp_quad_perm(lane0, lane1, lane2, lane3);
4586 if (ctx->chip_class >= GFX8) {
4587 return ac_build_dpp(ctx, src, src, mask, 0xf, 0xf, false);
4588 } else {
4589 return ac_build_ds_swizzle(ctx, src, (1 << 15) | mask);
4590 }
4591 }
4592
4593 LLVMValueRef
4594 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index)
4595 {
4596 LLVMTypeRef type = LLVMTypeOf(src);
4597 LLVMValueRef result;
4598
4599 index = LLVMBuildMul(ctx->builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
4600 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
4601
4602 result = ac_build_intrinsic(ctx, "llvm.amdgcn.ds.bpermute", ctx->i32,
4603 (LLVMValueRef []) {index, src}, 2,
4604 AC_FUNC_ATTR_READNONE |
4605 AC_FUNC_ATTR_CONVERGENT);
4606 return LLVMBuildTrunc(ctx->builder, result, type, "");
4607 }
4608
4609 LLVMValueRef
4610 ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
4611 unsigned bitsize)
4612 {
4613 LLVMTypeRef type;
4614 char *intr;
4615
4616 if (bitsize == 16) {
4617 intr = "llvm.amdgcn.frexp.exp.i16.f16";
4618 type = ctx->i16;
4619 } else if (bitsize == 32) {
4620 intr = "llvm.amdgcn.frexp.exp.i32.f32";
4621 type = ctx->i32;
4622 } else {
4623 intr = "llvm.amdgcn.frexp.exp.i32.f64";
4624 type = ctx->i32;
4625 }
4626
4627 LLVMValueRef params[] = {
4628 src0,
4629 };
4630 return ac_build_intrinsic(ctx, intr, type, params, 1,
4631 AC_FUNC_ATTR_READNONE);
4632 }
4633 LLVMValueRef
4634 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
4635 unsigned bitsize)
4636 {
4637 LLVMTypeRef type;
4638 char *intr;
4639
4640 if (bitsize == 16) {
4641 intr = "llvm.amdgcn.frexp.mant.f16";
4642 type = ctx->f16;
4643 } else if (bitsize == 32) {
4644 intr = "llvm.amdgcn.frexp.mant.f32";
4645 type = ctx->f32;
4646 } else {
4647 intr = "llvm.amdgcn.frexp.mant.f64";
4648 type = ctx->f64;
4649 }
4650
4651 LLVMValueRef params[] = {
4652 src0,
4653 };
4654 return ac_build_intrinsic(ctx, intr, type, params, 1,
4655 AC_FUNC_ATTR_READNONE);
4656 }
4657
4658 LLVMValueRef
4659 ac_build_canonicalize(struct ac_llvm_context *ctx, LLVMValueRef src0,
4660 unsigned bitsize)
4661 {
4662 LLVMTypeRef type;
4663 char *intr;
4664
4665 if (bitsize == 16) {
4666 intr = "llvm.canonicalize.f16";
4667 type = ctx->f16;
4668 } else if (bitsize == 32) {
4669 intr = "llvm.canonicalize.f32";
4670 type = ctx->f32;
4671 } else {
4672 intr = "llvm.canonicalize.f64";
4673 type = ctx->f64;
4674 }
4675
4676 LLVMValueRef params[] = {
4677 src0,
4678 };
4679 return ac_build_intrinsic(ctx, intr, type, params, 1,
4680 AC_FUNC_ATTR_READNONE);
4681 }
4682
4683 /*
4684 * this takes an I,J coordinate pair,
4685 * and works out the X and Y derivatives.
4686 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
4687 */
4688 LLVMValueRef
4689 ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij)
4690 {
4691 LLVMValueRef result[4], a;
4692 unsigned i;
4693
4694 for (i = 0; i < 2; i++) {
4695 a = LLVMBuildExtractElement(ctx->builder, interp_ij,
4696 LLVMConstInt(ctx->i32, i, false), "");
4697 result[i] = ac_build_ddxy(ctx, AC_TID_MASK_TOP_LEFT, 1, a);
4698 result[2+i] = ac_build_ddxy(ctx, AC_TID_MASK_TOP_LEFT, 2, a);
4699 }
4700 return ac_build_gather_values(ctx, result, 4);
4701 }
4702
4703 LLVMValueRef
4704 ac_build_load_helper_invocation(struct ac_llvm_context *ctx)
4705 {
4706 LLVMValueRef result = ac_build_intrinsic(ctx, "llvm.amdgcn.ps.live",
4707 ctx->i1, NULL, 0,
4708 AC_FUNC_ATTR_READNONE);
4709 result = LLVMBuildNot(ctx->builder, result, "");
4710 return LLVMBuildSExt(ctx->builder, result, ctx->i32, "");
4711 }
4712
4713 LLVMValueRef
4714 ac_build_is_helper_invocation(struct ac_llvm_context *ctx)
4715 {
4716 if (!ctx->postponed_kill)
4717 return ac_build_load_helper_invocation(ctx);
4718
4719 /* !(exact && postponed) */
4720 LLVMValueRef exact = ac_build_intrinsic(ctx, "llvm.amdgcn.ps.live",
4721 ctx->i1, NULL, 0,
4722 AC_FUNC_ATTR_READNONE);
4723
4724 LLVMValueRef postponed = LLVMBuildLoad(ctx->builder, ctx->postponed_kill, "");
4725 LLVMValueRef result = LLVMBuildAnd(ctx->builder, exact, postponed, "");
4726
4727 return LLVMBuildSelect(ctx->builder, result, ctx->i32_0,
4728 LLVMConstInt(ctx->i32, 0xFFFFFFFF, false), "");
4729 }
4730
4731 LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
4732 LLVMValueRef *args, unsigned num_args)
4733 {
4734 LLVMValueRef ret = LLVMBuildCall(ctx->builder, func, args, num_args, "");
4735 LLVMSetInstructionCallConv(ret, LLVMGetFunctionCallConv(func));
4736 return ret;
4737 }
4738
4739 void
4740 ac_export_mrt_z(struct ac_llvm_context *ctx, LLVMValueRef depth,
4741 LLVMValueRef stencil, LLVMValueRef samplemask,
4742 struct ac_export_args *args)
4743 {
4744 unsigned mask = 0;
4745 unsigned format = ac_get_spi_shader_z_format(depth != NULL,
4746 stencil != NULL,
4747 samplemask != NULL);
4748
4749 assert(depth || stencil || samplemask);
4750
4751 memset(args, 0, sizeof(*args));
4752
4753 args->valid_mask = 1; /* whether the EXEC mask is valid */
4754 args->done = 1; /* DONE bit */
4755
4756 /* Specify the target we are exporting */
4757 args->target = V_008DFC_SQ_EXP_MRTZ;
4758
4759 args->compr = 0; /* COMP flag */
4760 args->out[0] = LLVMGetUndef(ctx->f32); /* R, depth */
4761 args->out[1] = LLVMGetUndef(ctx->f32); /* G, stencil test val[0:7], stencil op val[8:15] */
4762 args->out[2] = LLVMGetUndef(ctx->f32); /* B, sample mask */
4763 args->out[3] = LLVMGetUndef(ctx->f32); /* A, alpha to mask */
4764
4765 if (format == V_028710_SPI_SHADER_UINT16_ABGR) {
4766 assert(!depth);
4767 args->compr = 1; /* COMPR flag */
4768
4769 if (stencil) {
4770 /* Stencil should be in X[23:16]. */
4771 stencil = ac_to_integer(ctx, stencil);
4772 stencil = LLVMBuildShl(ctx->builder, stencil,
4773 LLVMConstInt(ctx->i32, 16, 0), "");
4774 args->out[0] = ac_to_float(ctx, stencil);
4775 mask |= 0x3;
4776 }
4777 if (samplemask) {
4778 /* SampleMask should be in Y[15:0]. */
4779 args->out[1] = samplemask;
4780 mask |= 0xc;
4781 }
4782 } else {
4783 if (depth) {
4784 args->out[0] = depth;
4785 mask |= 0x1;
4786 }
4787 if (stencil) {
4788 args->out[1] = stencil;
4789 mask |= 0x2;
4790 }
4791 if (samplemask) {
4792 args->out[2] = samplemask;
4793 mask |= 0x4;
4794 }
4795 }
4796
4797 /* GFX6 (except OLAND and HAINAN) has a bug that it only looks
4798 * at the X writemask component. */
4799 if (ctx->chip_class == GFX6 &&
4800 ctx->family != CHIP_OLAND &&
4801 ctx->family != CHIP_HAINAN)
4802 mask |= 0x1;
4803
4804 /* Specify which components to enable */
4805 args->enabled_channels = mask;
4806 }
4807
4808 /* Send GS Alloc Req message from the first wave of the group to SPI.
4809 * Message payload is:
4810 * - bits 0..10: vertices in group
4811 * - bits 12..22: primitives in group
4812 */
4813 void ac_build_sendmsg_gs_alloc_req(struct ac_llvm_context *ctx, LLVMValueRef wave_id,
4814 LLVMValueRef vtx_cnt, LLVMValueRef prim_cnt)
4815 {
4816 LLVMBuilderRef builder = ctx->builder;
4817 LLVMValueRef tmp;
4818 bool export_dummy_prim = false;
4819
4820 /* HW workaround for a GPU hang with 100% culling.
4821 * We always have to export at least 1 primitive.
4822 * Export a degenerate triangle using vertex 0 for all 3 vertices.
4823 */
4824 if (prim_cnt == ctx->i32_0 && ctx->chip_class == GFX10) {
4825 assert(vtx_cnt == ctx->i32_0);
4826 prim_cnt = ctx->i32_1;
4827 vtx_cnt = ctx->i32_1;
4828 export_dummy_prim = true;
4829 }
4830
4831 ac_build_ifcc(ctx, LLVMBuildICmp(builder, LLVMIntEQ, wave_id, ctx->i32_0, ""), 5020);
4832
4833 tmp = LLVMBuildShl(builder, prim_cnt, LLVMConstInt(ctx->i32, 12, false),"");
4834 tmp = LLVMBuildOr(builder, tmp, vtx_cnt, "");
4835 ac_build_sendmsg(ctx, AC_SENDMSG_GS_ALLOC_REQ, tmp);
4836
4837 if (export_dummy_prim) {
4838 struct ac_ngg_prim prim = {};
4839 /* The vertex indices are 0,0,0. */
4840 prim.passthrough = ctx->i32_0;
4841
4842 struct ac_export_args pos = {};
4843 pos.out[0] = pos.out[1] = pos.out[2] = pos.out[3] = ctx->f32_0;
4844 pos.target = V_008DFC_SQ_EXP_POS;
4845 pos.enabled_channels = 0xf;
4846 pos.done = true;
4847
4848 ac_build_ifcc(ctx, LLVMBuildICmp(builder, LLVMIntEQ, ac_get_thread_id(ctx),
4849 ctx->i32_0, ""), 5021);
4850 ac_build_export_prim(ctx, &prim);
4851 ac_build_export(ctx, &pos);
4852 ac_build_endif(ctx, 5021);
4853 }
4854
4855 ac_build_endif(ctx, 5020);
4856 }
4857
4858 LLVMValueRef ac_pack_prim_export(struct ac_llvm_context *ctx,
4859 const struct ac_ngg_prim *prim)
4860 {
4861 /* The prim export format is:
4862 * - bits 0..8: index 0
4863 * - bit 9: edge flag 0
4864 * - bits 10..18: index 1
4865 * - bit 19: edge flag 1
4866 * - bits 20..28: index 2
4867 * - bit 29: edge flag 2
4868 * - bit 31: null primitive (skip)
4869 */
4870 LLVMBuilderRef builder = ctx->builder;
4871 LLVMValueRef tmp = LLVMBuildZExt(builder, prim->isnull, ctx->i32, "");
4872 LLVMValueRef result = LLVMBuildShl(builder, tmp, LLVMConstInt(ctx->i32, 31, false), "");
4873
4874 for (unsigned i = 0; i < prim->num_vertices; ++i) {
4875 tmp = LLVMBuildShl(builder, prim->index[i],
4876 LLVMConstInt(ctx->i32, 10 * i, false), "");
4877 result = LLVMBuildOr(builder, result, tmp, "");
4878 tmp = LLVMBuildZExt(builder, prim->edgeflag[i], ctx->i32, "");
4879 tmp = LLVMBuildShl(builder, tmp,
4880 LLVMConstInt(ctx->i32, 10 * i + 9, false), "");
4881 result = LLVMBuildOr(builder, result, tmp, "");
4882 }
4883 return result;
4884 }
4885
4886 void ac_build_export_prim(struct ac_llvm_context *ctx,
4887 const struct ac_ngg_prim *prim)
4888 {
4889 struct ac_export_args args;
4890
4891 if (prim->passthrough) {
4892 args.out[0] = prim->passthrough;
4893 } else {
4894 args.out[0] = ac_pack_prim_export(ctx, prim);
4895 }
4896
4897 args.out[0] = LLVMBuildBitCast(ctx->builder, args.out[0], ctx->f32, "");
4898 args.out[1] = LLVMGetUndef(ctx->f32);
4899 args.out[2] = LLVMGetUndef(ctx->f32);
4900 args.out[3] = LLVMGetUndef(ctx->f32);
4901
4902 args.target = V_008DFC_SQ_EXP_PRIM;
4903 args.enabled_channels = 1;
4904 args.done = true;
4905 args.valid_mask = false;
4906 args.compr = false;
4907
4908 ac_build_export(ctx, &args);
4909 }
4910
4911 static LLVMTypeRef
4912 arg_llvm_type(enum ac_arg_type type, unsigned size, struct ac_llvm_context *ctx)
4913 {
4914 if (type == AC_ARG_FLOAT) {
4915 return size == 1 ? ctx->f32 : LLVMVectorType(ctx->f32, size);
4916 } else if (type == AC_ARG_INT) {
4917 return size == 1 ? ctx->i32 : LLVMVectorType(ctx->i32, size);
4918 } else {
4919 LLVMTypeRef ptr_type;
4920 switch (type) {
4921 case AC_ARG_CONST_PTR:
4922 ptr_type = ctx->i8;
4923 break;
4924 case AC_ARG_CONST_FLOAT_PTR:
4925 ptr_type = ctx->f32;
4926 break;
4927 case AC_ARG_CONST_PTR_PTR:
4928 ptr_type = ac_array_in_const32_addr_space(ctx->i8);
4929 break;
4930 case AC_ARG_CONST_DESC_PTR:
4931 ptr_type = ctx->v4i32;
4932 break;
4933 case AC_ARG_CONST_IMAGE_PTR:
4934 ptr_type = ctx->v8i32;
4935 break;
4936 default:
4937 unreachable("unknown arg type");
4938 }
4939 if (size == 1) {
4940 return ac_array_in_const32_addr_space(ptr_type);
4941 } else {
4942 assert(size == 2);
4943 return ac_array_in_const_addr_space(ptr_type);
4944 }
4945 }
4946 }
4947
4948 LLVMValueRef
4949 ac_build_main(const struct ac_shader_args *args,
4950 struct ac_llvm_context *ctx,
4951 enum ac_llvm_calling_convention convention,
4952 const char *name, LLVMTypeRef ret_type,
4953 LLVMModuleRef module)
4954 {
4955 LLVMTypeRef arg_types[AC_MAX_ARGS];
4956
4957 for (unsigned i = 0; i < args->arg_count; i++) {
4958 arg_types[i] = arg_llvm_type(args->args[i].type,
4959 args->args[i].size, ctx);
4960 }
4961
4962 LLVMTypeRef main_function_type =
4963 LLVMFunctionType(ret_type, arg_types, args->arg_count, 0);
4964
4965 LLVMValueRef main_function =
4966 LLVMAddFunction(module, name, main_function_type);
4967 LLVMBasicBlockRef main_function_body =
4968 LLVMAppendBasicBlockInContext(ctx->context, main_function, "main_body");
4969 LLVMPositionBuilderAtEnd(ctx->builder, main_function_body);
4970
4971 LLVMSetFunctionCallConv(main_function, convention);
4972 for (unsigned i = 0; i < args->arg_count; ++i) {
4973 LLVMValueRef P = LLVMGetParam(main_function, i);
4974
4975 if (args->args[i].file != AC_ARG_SGPR)
4976 continue;
4977
4978 ac_add_function_attr(ctx->context, main_function, i + 1, AC_FUNC_ATTR_INREG);
4979
4980 if (LLVMGetTypeKind(LLVMTypeOf(P)) == LLVMPointerTypeKind) {
4981 ac_add_function_attr(ctx->context, main_function, i + 1, AC_FUNC_ATTR_NOALIAS);
4982 ac_add_attr_dereferenceable(P, UINT64_MAX);
4983 }
4984 }
4985
4986 ctx->main_function = main_function;
4987
4988 if (LLVM_VERSION_MAJOR >= 11) {
4989 /* Enable denormals for FP16 and FP64: */
4990 LLVMAddTargetDependentFunctionAttr(main_function, "denormal-fp-math",
4991 "ieee,ieee");
4992 /* Disable denormals for FP32: */
4993 LLVMAddTargetDependentFunctionAttr(main_function, "denormal-fp-math-f32",
4994 "preserve-sign,preserve-sign");
4995 }
4996 return main_function;
4997 }
4998
4999 void ac_build_s_endpgm(struct ac_llvm_context *ctx)
5000 {
5001 LLVMTypeRef calltype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
5002 LLVMValueRef code = LLVMConstInlineAsm(calltype, "s_endpgm", "", true, false);
5003 LLVMBuildCall(ctx->builder, code, NULL, 0, "");
5004 }
5005
5006 LLVMValueRef ac_prefix_bitcount(struct ac_llvm_context *ctx,
5007 LLVMValueRef mask, LLVMValueRef index)
5008 {
5009 LLVMBuilderRef builder = ctx->builder;
5010 LLVMTypeRef type = LLVMTypeOf(mask);
5011
5012 LLVMValueRef bit = LLVMBuildShl(builder, LLVMConstInt(type, 1, 0),
5013 LLVMBuildZExt(builder, index, type, ""), "");
5014 LLVMValueRef prefix_bits = LLVMBuildSub(builder, bit, LLVMConstInt(type, 1, 0), "");
5015 LLVMValueRef prefix_mask = LLVMBuildAnd(builder, mask, prefix_bits, "");
5016 return ac_build_bit_count(ctx, prefix_mask);
5017 }
5018
5019 /* Compute the prefix sum of the "mask" bit array with 128 elements (bits). */
5020 LLVMValueRef ac_prefix_bitcount_2x64(struct ac_llvm_context *ctx,
5021 LLVMValueRef mask[2], LLVMValueRef index)
5022 {
5023 LLVMBuilderRef builder = ctx->builder;
5024 #if 0
5025 /* Reference version using i128. */
5026 LLVMValueRef input_mask =
5027 LLVMBuildBitCast(builder, ac_build_gather_values(ctx, mask, 2), ctx->i128, "");
5028
5029 return ac_prefix_bitcount(ctx, input_mask, index);
5030 #else
5031 /* Optimized version using 2 64-bit masks. */
5032 LLVMValueRef is_hi, is_0, c64, c128, all_bits;
5033 LLVMValueRef prefix_mask[2], shift[2], mask_bcnt0, prefix_bcnt[2];
5034
5035 /* Compute the 128-bit prefix mask. */
5036 c64 = LLVMConstInt(ctx->i32, 64, 0);
5037 c128 = LLVMConstInt(ctx->i32, 128, 0);
5038 all_bits = LLVMConstInt(ctx->i64, UINT64_MAX, 0);
5039 /* The first index that can have non-zero high bits in the prefix mask is 65. */
5040 is_hi = LLVMBuildICmp(builder, LLVMIntUGT, index, c64, "");
5041 is_0 = LLVMBuildICmp(builder, LLVMIntEQ, index, ctx->i32_0, "");
5042 mask_bcnt0 = ac_build_bit_count(ctx, mask[0]);
5043
5044 for (unsigned i = 0; i < 2; i++) {
5045 shift[i] = LLVMBuildSub(builder, i ? c128 : c64, index, "");
5046 /* For i==0, index==0, the right shift by 64 doesn't give the desired result,
5047 * so we handle it by the is_0 select.
5048 * For i==1, index==64, same story, so we handle it by the last is_hi select.
5049 * For i==0, index==64, we shift by 0, which is what we want.
5050 */
5051 prefix_mask[i] = LLVMBuildLShr(builder, all_bits,
5052 LLVMBuildZExt(builder, shift[i], ctx->i64, ""), "");
5053 prefix_mask[i] = LLVMBuildAnd(builder, mask[i], prefix_mask[i], "");
5054 prefix_bcnt[i] = ac_build_bit_count(ctx, prefix_mask[i]);
5055 }
5056
5057 prefix_bcnt[0] = LLVMBuildSelect(builder, is_0, ctx->i32_0, prefix_bcnt[0], "");
5058 prefix_bcnt[0] = LLVMBuildSelect(builder, is_hi, mask_bcnt0, prefix_bcnt[0], "");
5059 prefix_bcnt[1] = LLVMBuildSelect(builder, is_hi, prefix_bcnt[1], ctx->i32_0, "");
5060
5061 return LLVMBuildAdd(builder, prefix_bcnt[0], prefix_bcnt[1], "");
5062 #endif
5063 }
5064
5065 /**
5066 * Convert triangle strip indices to triangle indices. This is used to decompose
5067 * triangle strips into triangles.
5068 */
5069 void ac_build_triangle_strip_indices_to_triangle(struct ac_llvm_context *ctx,
5070 LLVMValueRef is_odd,
5071 LLVMValueRef flatshade_first,
5072 LLVMValueRef index[3])
5073 {
5074 LLVMBuilderRef builder = ctx->builder;
5075 LLVMValueRef out[3];
5076
5077 /* We need to change the vertex order for odd triangles to get correct
5078 * front/back facing by swapping 2 vertex indices, but we also have to
5079 * keep the provoking vertex in the same place.
5080 *
5081 * If the first vertex is provoking, swap index 1 and 2.
5082 * If the last vertex is provoking, swap index 0 and 1.
5083 */
5084 out[0] = LLVMBuildSelect(builder, flatshade_first,
5085 index[0],
5086 LLVMBuildSelect(builder, is_odd,
5087 index[1], index[0], ""), "");
5088 out[1] = LLVMBuildSelect(builder, flatshade_first,
5089 LLVMBuildSelect(builder, is_odd,
5090 index[2], index[1], ""),
5091 LLVMBuildSelect(builder, is_odd,
5092 index[0], index[1], ""), "");
5093 out[2] = LLVMBuildSelect(builder, flatshade_first,
5094 LLVMBuildSelect(builder, is_odd,
5095 index[1], index[2], ""),
5096 index[2], "");
5097 memcpy(index, out, sizeof(out));
5098 }