projects / mesa.git / blob
? search:


de9091693c1916fd377e30e777b1ef5cce6422f8
[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 /* For doubles, we need precise division to pass GLCTS. */
709 if (ctx->float_mode == AC_FLOAT_MODE_DEFAULT_OPENGL &&
710 type_size == 8)
711 return LLVMBuildFDiv(ctx->builder, num, den, "");
712
713 if (type_size == 2)
714 name = "llvm.amdgcn.rcp.f16";
715 else if (type_size == 4)
716 name = "llvm.amdgcn.rcp.f32";
717 else
718 name = "llvm.amdgcn.rcp.f64";
719
720 LLVMValueRef rcp = ac_build_intrinsic(ctx, name, LLVMTypeOf(den),
721 &den, 1, AC_FUNC_ATTR_READNONE);
722
723 return LLVMBuildFMul(ctx->builder, num, rcp, "");
724 }
725
726 /* See fast_idiv_by_const.h. */
727 /* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
728 LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
729 LLVMValueRef num,
730 LLVMValueRef multiplier,
731 LLVMValueRef pre_shift,
732 LLVMValueRef post_shift,
733 LLVMValueRef increment)
734 {
735 LLVMBuilderRef builder = ctx->builder;
736
737 num = LLVMBuildLShr(builder, num, pre_shift, "");
738 num = LLVMBuildMul(builder,
739 LLVMBuildZExt(builder, num, ctx->i64, ""),
740 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
741 num = LLVMBuildAdd(builder, num,
742 LLVMBuildZExt(builder, increment, ctx->i64, ""), "");
743 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
744 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
745 return LLVMBuildLShr(builder, num, post_shift, "");
746 }
747
748 /* See fast_idiv_by_const.h. */
749 /* If num != UINT_MAX, this more efficient version can be used. */
750 /* Set: increment = util_fast_udiv_info::increment; */
751 LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
752 LLVMValueRef num,
753 LLVMValueRef multiplier,
754 LLVMValueRef pre_shift,
755 LLVMValueRef post_shift,
756 LLVMValueRef increment)
757 {
758 LLVMBuilderRef builder = ctx->builder;
759
760 num = LLVMBuildLShr(builder, num, pre_shift, "");
761 num = LLVMBuildNUWAdd(builder, num, increment, "");
762 num = LLVMBuildMul(builder,
763 LLVMBuildZExt(builder, num, ctx->i64, ""),
764 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
765 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
766 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
767 return LLVMBuildLShr(builder, num, post_shift, "");
768 }
769
770 /* See fast_idiv_by_const.h. */
771 /* Both operands must fit in 31 bits and the divisor must not be 1. */
772 LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
773 LLVMValueRef num,
774 LLVMValueRef multiplier,
775 LLVMValueRef post_shift)
776 {
777 LLVMBuilderRef builder = ctx->builder;
778
779 num = LLVMBuildMul(builder,
780 LLVMBuildZExt(builder, num, ctx->i64, ""),
781 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
782 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
783 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
784 return LLVMBuildLShr(builder, num, post_shift, "");
785 }
786
787 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
788 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
789 * already multiplied by two. id is the cube face number.
790 */
791 struct cube_selection_coords {
792 LLVMValueRef stc[2];
793 LLVMValueRef ma;
794 LLVMValueRef id;
795 };
796
797 static void
798 build_cube_intrinsic(struct ac_llvm_context *ctx,
799 LLVMValueRef in[3],
800 struct cube_selection_coords *out)
801 {
802 LLVMTypeRef f32 = ctx->f32;
803
804 out->stc[1] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubetc",
805 f32, in, 3, AC_FUNC_ATTR_READNONE);
806 out->stc[0] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubesc",
807 f32, in, 3, AC_FUNC_ATTR_READNONE);
808 out->ma = ac_build_intrinsic(ctx, "llvm.amdgcn.cubema",
809 f32, in, 3, AC_FUNC_ATTR_READNONE);
810 out->id = ac_build_intrinsic(ctx, "llvm.amdgcn.cubeid",
811 f32, in, 3, AC_FUNC_ATTR_READNONE);
812 }
813
814 /**
815 * Build a manual selection sequence for cube face sc/tc coordinates and
816 * major axis vector (multiplied by 2 for consistency) for the given
817 * vec3 \p coords, for the face implied by \p selcoords.
818 *
819 * For the major axis, we always adjust the sign to be in the direction of
820 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
821 * the selcoords major axis.
822 */
823 static void build_cube_select(struct ac_llvm_context *ctx,
824 const struct cube_selection_coords *selcoords,
825 const LLVMValueRef *coords,
826 LLVMValueRef *out_st,
827 LLVMValueRef *out_ma)
828 {
829 LLVMBuilderRef builder = ctx->builder;
830 LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
831 LLVMValueRef is_ma_positive;
832 LLVMValueRef sgn_ma;
833 LLVMValueRef is_ma_z, is_not_ma_z;
834 LLVMValueRef is_ma_y;
835 LLVMValueRef is_ma_x;
836 LLVMValueRef sgn;
837 LLVMValueRef tmp;
838
839 is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
840 selcoords->ma, LLVMConstReal(f32, 0.0), "");
841 sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
842 LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
843
844 is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
845 is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
846 is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
847 LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
848 is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
849
850 /* Select sc */
851 tmp = LLVMBuildSelect(builder, is_ma_x, coords[2], coords[0], "");
852 sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
853 LLVMBuildSelect(builder, is_ma_z, sgn_ma,
854 LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
855 out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
856
857 /* Select tc */
858 tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
859 sgn = LLVMBuildSelect(builder, is_ma_y, sgn_ma,
860 LLVMConstReal(f32, -1.0), "");
861 out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
862
863 /* Select ma */
864 tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
865 LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
866 tmp = ac_build_intrinsic(ctx, "llvm.fabs.f32",
867 ctx->f32, &tmp, 1, AC_FUNC_ATTR_READNONE);
868 *out_ma = LLVMBuildFMul(builder, tmp, LLVMConstReal(f32, 2.0), "");
869 }
870
871 void
872 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
873 bool is_deriv, bool is_array, bool is_lod,
874 LLVMValueRef *coords_arg,
875 LLVMValueRef *derivs_arg)
876 {
877
878 LLVMBuilderRef builder = ctx->builder;
879 struct cube_selection_coords selcoords;
880 LLVMValueRef coords[3];
881 LLVMValueRef invma;
882
883 if (is_array && !is_lod) {
884 LLVMValueRef tmp = ac_build_round(ctx, coords_arg[3]);
885
886 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
887 *
888 * "For Array forms, the array layer used will be
889 *
890 * max(0, min(d−1, floor(layer+0.5)))
891 *
892 * where d is the depth of the texture array and layer
893 * comes from the component indicated in the tables below.
894 * Workaroudn for an issue where the layer is taken from a
895 * helper invocation which happens to fall on a different
896 * layer due to extrapolation."
897 *
898 * GFX8 and earlier attempt to implement this in hardware by
899 * clamping the value of coords[2] = (8 * layer) + face.
900 * Unfortunately, this means that the we end up with the wrong
901 * face when clamping occurs.
902 *
903 * Clamp the layer earlier to work around the issue.
904 */
905 if (ctx->chip_class <= GFX8) {
906 LLVMValueRef ge0;
907 ge0 = LLVMBuildFCmp(builder, LLVMRealOGE, tmp, ctx->f32_0, "");
908 tmp = LLVMBuildSelect(builder, ge0, tmp, ctx->f32_0, "");
909 }
910
911 coords_arg[3] = tmp;
912 }
913
914 build_cube_intrinsic(ctx, coords_arg, &selcoords);
915
916 invma = ac_build_intrinsic(ctx, "llvm.fabs.f32",
917 ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
918 invma = ac_build_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
919
920 for (int i = 0; i < 2; ++i)
921 coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
922
923 coords[2] = selcoords.id;
924
925 if (is_deriv && derivs_arg) {
926 LLVMValueRef derivs[4];
927 int axis;
928
929 /* Convert cube derivatives to 2D derivatives. */
930 for (axis = 0; axis < 2; axis++) {
931 LLVMValueRef deriv_st[2];
932 LLVMValueRef deriv_ma;
933
934 /* Transform the derivative alongside the texture
935 * coordinate. Mathematically, the correct formula is
936 * as follows. Assume we're projecting onto the +Z face
937 * and denote by dx/dh the derivative of the (original)
938 * X texture coordinate with respect to horizontal
939 * window coordinates. The projection onto the +Z face
940 * plane is:
941 *
942 * f(x,z) = x/z
943 *
944 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
945 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
946 *
947 * This motivatives the implementation below.
948 *
949 * Whether this actually gives the expected results for
950 * apps that might feed in derivatives obtained via
951 * finite differences is anyone's guess. The OpenGL spec
952 * seems awfully quiet about how textureGrad for cube
953 * maps should be handled.
954 */
955 build_cube_select(ctx, &selcoords, &derivs_arg[axis * 3],
956 deriv_st, &deriv_ma);
957
958 deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
959
960 for (int i = 0; i < 2; ++i)
961 derivs[axis * 2 + i] =
962 LLVMBuildFSub(builder,
963 LLVMBuildFMul(builder, deriv_st[i], invma, ""),
964 LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
965 }
966
967 memcpy(derivs_arg, derivs, sizeof(derivs));
968 }
969
970 /* Shift the texture coordinate. This must be applied after the
971 * derivative calculation.
972 */
973 for (int i = 0; i < 2; ++i)
974 coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
975
976 if (is_array) {
977 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
978 /* coords_arg.w component - array_index for cube arrays */
979 coords[2] = ac_build_fmad(ctx, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), coords[2]);
980 }
981
982 memcpy(coords_arg, coords, sizeof(coords));
983 }
984
985
986 LLVMValueRef
987 ac_build_fs_interp(struct ac_llvm_context *ctx,
988 LLVMValueRef llvm_chan,
989 LLVMValueRef attr_number,
990 LLVMValueRef params,
991 LLVMValueRef i,
992 LLVMValueRef j)
993 {
994 LLVMValueRef args[5];
995 LLVMValueRef p1;
996
997 args[0] = i;
998 args[1] = llvm_chan;
999 args[2] = attr_number;
1000 args[3] = params;
1001
1002 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1",
1003 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
1004
1005 args[0] = p1;
1006 args[1] = j;
1007 args[2] = llvm_chan;
1008 args[3] = attr_number;
1009 args[4] = params;
1010
1011 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2",
1012 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
1013 }
1014
1015 LLVMValueRef
1016 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
1017 LLVMValueRef llvm_chan,
1018 LLVMValueRef attr_number,
1019 LLVMValueRef params,
1020 LLVMValueRef i,
1021 LLVMValueRef j)
1022 {
1023 LLVMValueRef args[6];
1024 LLVMValueRef p1;
1025
1026 args[0] = i;
1027 args[1] = llvm_chan;
1028 args[2] = attr_number;
1029 args[3] = ctx->i1false;
1030 args[4] = params;
1031
1032 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1.f16",
1033 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
1034
1035 args[0] = p1;
1036 args[1] = j;
1037 args[2] = llvm_chan;
1038 args[3] = attr_number;
1039 args[4] = ctx->i1false;
1040 args[5] = params;
1041
1042 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2.f16",
1043 ctx->f16, args, 6, AC_FUNC_ATTR_READNONE);
1044 }
1045
1046 LLVMValueRef
1047 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
1048 LLVMValueRef parameter,
1049 LLVMValueRef llvm_chan,
1050 LLVMValueRef attr_number,
1051 LLVMValueRef params)
1052 {
1053 LLVMValueRef args[4];
1054
1055 args[0] = parameter;
1056 args[1] = llvm_chan;
1057 args[2] = attr_number;
1058 args[3] = params;
1059
1060 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.mov",
1061 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
1062 }
1063
1064 LLVMValueRef
1065 ac_build_gep_ptr(struct ac_llvm_context *ctx,
1066 LLVMValueRef base_ptr,
1067 LLVMValueRef index)
1068 {
1069 return LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
1070 }
1071
1072 LLVMValueRef
1073 ac_build_gep0(struct ac_llvm_context *ctx,
1074 LLVMValueRef base_ptr,
1075 LLVMValueRef index)
1076 {
1077 LLVMValueRef indices[2] = {
1078 ctx->i32_0,
1079 index,
1080 };
1081 return LLVMBuildGEP(ctx->builder, base_ptr, indices, 2, "");
1082 }
1083
1084 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
1085 LLVMValueRef index)
1086 {
1087 return LLVMBuildPointerCast(ctx->builder,
1088 LLVMBuildGEP(ctx->builder, ptr, &index, 1, ""),
1089 LLVMTypeOf(ptr), "");
1090 }
1091
1092 void
1093 ac_build_indexed_store(struct ac_llvm_context *ctx,
1094 LLVMValueRef base_ptr, LLVMValueRef index,
1095 LLVMValueRef value)
1096 {
1097 LLVMBuildStore(ctx->builder, value,
1098 ac_build_gep0(ctx, base_ptr, index));
1099 }
1100
1101 /**
1102 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
1103 * It's equivalent to doing a load from &base_ptr[index].
1104 *
1105 * \param base_ptr Where the array starts.
1106 * \param index The element index into the array.
1107 * \param uniform Whether the base_ptr and index can be assumed to be
1108 * dynamically uniform (i.e. load to an SGPR)
1109 * \param invariant Whether the load is invariant (no other opcodes affect it)
1110 * \param no_unsigned_wraparound
1111 * For all possible re-associations and re-distributions of an expression
1112 * "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
1113 * without inbounds in base_ptr), this parameter is true if "addr + offset"
1114 * does not result in an unsigned integer wraparound. This is used for
1115 * optimal code generation of 32-bit pointer arithmetic.
1116 *
1117 * For example, a 32-bit immediate offset that causes a 32-bit unsigned
1118 * integer wraparound can't be an imm offset in s_load_dword, because
1119 * the instruction performs "addr + offset" in 64 bits.
1120 *
1121 * Expected usage for bindless textures by chaining GEPs:
1122 * // possible unsigned wraparound, don't use InBounds:
1123 * ptr1 = LLVMBuildGEP(base_ptr, index);
1124 * image = load(ptr1); // becomes "s_load ptr1, 0"
1125 *
1126 * ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1127 * sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1128 */
1129 static LLVMValueRef
1130 ac_build_load_custom(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1131 LLVMValueRef index, bool uniform, bool invariant,
1132 bool no_unsigned_wraparound)
1133 {
1134 LLVMValueRef pointer, result;
1135
1136 if (no_unsigned_wraparound &&
1137 LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr)) == AC_ADDR_SPACE_CONST_32BIT)
1138 pointer = LLVMBuildInBoundsGEP(ctx->builder, base_ptr, &index, 1, "");
1139 else
1140 pointer = LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
1141
1142 if (uniform)
1143 LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
1144 result = LLVMBuildLoad(ctx->builder, pointer, "");
1145 if (invariant)
1146 LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
1147 return result;
1148 }
1149
1150 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1151 LLVMValueRef index)
1152 {
1153 return ac_build_load_custom(ctx, base_ptr, index, false, false, false);
1154 }
1155
1156 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
1157 LLVMValueRef base_ptr, LLVMValueRef index)
1158 {
1159 return ac_build_load_custom(ctx, base_ptr, index, false, true, false);
1160 }
1161
1162 /* This assumes that there is no unsigned integer wraparound during the address
1163 * computation, excluding all GEPs within base_ptr. */
1164 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
1165 LLVMValueRef base_ptr, LLVMValueRef index)
1166 {
1167 return ac_build_load_custom(ctx, base_ptr, index, true, true, true);
1168 }
1169
1170 /* See ac_build_load_custom() documentation. */
1171 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
1172 LLVMValueRef base_ptr, LLVMValueRef index)
1173 {
1174 return ac_build_load_custom(ctx, base_ptr, index, true, true, false);
1175 }
1176
1177 static unsigned get_load_cache_policy(struct ac_llvm_context *ctx,
1178 unsigned cache_policy)
1179 {
1180 return cache_policy |
1181 (ctx->chip_class >= GFX10 && cache_policy & ac_glc ? ac_dlc : 0);
1182 }
1183
1184 static void
1185 ac_build_buffer_store_common(struct ac_llvm_context *ctx,
1186 LLVMValueRef rsrc,
1187 LLVMValueRef data,
1188 LLVMValueRef vindex,
1189 LLVMValueRef voffset,
1190 LLVMValueRef soffset,
1191 unsigned cache_policy,
1192 bool use_format,
1193 bool structurized)
1194 {
1195 LLVMValueRef args[6];
1196 int idx = 0;
1197 args[idx++] = data;
1198 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1199 if (structurized)
1200 args[idx++] = vindex ? vindex : ctx->i32_0;
1201 args[idx++] = voffset ? voffset : ctx->i32_0;
1202 args[idx++] = soffset ? soffset : ctx->i32_0;
1203 args[idx++] = LLVMConstInt(ctx->i32, cache_policy, 0);
1204 const char *indexing_kind = structurized ? "struct" : "raw";
1205 char name[256], type_name[8];
1206
1207 ac_build_type_name_for_intr(LLVMTypeOf(data), type_name, sizeof(type_name));
1208
1209 if (use_format) {
1210 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.format.%s",
1211 indexing_kind, type_name);
1212 } else {
1213 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.%s",
1214 indexing_kind, type_name);
1215 }
1216
1217 ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
1218 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
1219 }
1220
1221 void
1222 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
1223 LLVMValueRef rsrc,
1224 LLVMValueRef data,
1225 LLVMValueRef vindex,
1226 LLVMValueRef voffset,
1227 unsigned cache_policy)
1228 {
1229 ac_build_buffer_store_common(ctx, rsrc, data, vindex, voffset, NULL,
1230 cache_policy, true, true);
1231 }
1232
1233 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1234 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1235 * or v4i32 (num_channels=3,4).
1236 */
1237 void
1238 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
1239 LLVMValueRef rsrc,
1240 LLVMValueRef vdata,
1241 unsigned num_channels,
1242 LLVMValueRef voffset,
1243 LLVMValueRef soffset,
1244 unsigned inst_offset,
1245 unsigned cache_policy)
1246 {
1247 /* Split 3 channel stores, because only LLVM 9+ support 3-channel
1248 * intrinsics. */
1249 if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false)) {
1250 LLVMValueRef v[3], v01;
1251
1252 for (int i = 0; i < 3; i++) {
1253 v[i] = LLVMBuildExtractElement(ctx->builder, vdata,
1254 LLVMConstInt(ctx->i32, i, 0), "");
1255 }
1256 v01 = ac_build_gather_values(ctx, v, 2);
1257
1258 ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset,
1259 soffset, inst_offset, cache_policy);
1260 ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset,
1261 soffset, inst_offset + 8,
1262 cache_policy);
1263 return;
1264 }
1265
1266 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1267 * (voffset is swizzled, but soffset isn't swizzled).
1268 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1269 */
1270 if (!(cache_policy & ac_swizzled)) {
1271 LLVMValueRef offset = soffset;
1272
1273 if (inst_offset)
1274 offset = LLVMBuildAdd(ctx->builder, offset,
1275 LLVMConstInt(ctx->i32, inst_offset, 0), "");
1276
1277 ac_build_buffer_store_common(ctx, rsrc, ac_to_float(ctx, vdata),
1278 ctx->i32_0, voffset, offset,
1279 cache_policy, false, false);
1280 return;
1281 }
1282
1283 static const unsigned dfmts[] = {
1284 V_008F0C_BUF_DATA_FORMAT_32,
1285 V_008F0C_BUF_DATA_FORMAT_32_32,
1286 V_008F0C_BUF_DATA_FORMAT_32_32_32,
1287 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1288 };
1289 unsigned dfmt = dfmts[num_channels - 1];
1290 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1291 LLVMValueRef immoffset = LLVMConstInt(ctx->i32, inst_offset, 0);
1292
1293 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1294 immoffset, num_channels, dfmt, nfmt, cache_policy);
1295 }
1296
1297 static LLVMValueRef
1298 ac_build_buffer_load_common(struct ac_llvm_context *ctx,
1299 LLVMValueRef rsrc,
1300 LLVMValueRef vindex,
1301 LLVMValueRef voffset,
1302 LLVMValueRef soffset,
1303 unsigned num_channels,
1304 LLVMTypeRef channel_type,
1305 unsigned cache_policy,
1306 bool can_speculate,
1307 bool use_format,
1308 bool structurized)
1309 {
1310 LLVMValueRef args[5];
1311 int idx = 0;
1312 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1313 if (structurized)
1314 args[idx++] = vindex ? vindex : ctx->i32_0;
1315 args[idx++] = voffset ? voffset : ctx->i32_0;
1316 args[idx++] = soffset ? soffset : ctx->i32_0;
1317 args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
1318 unsigned func = !ac_has_vec3_support(ctx->chip_class, use_format) && num_channels == 3 ? 4 : num_channels;
1319 const char *indexing_kind = structurized ? "struct" : "raw";
1320 char name[256], type_name[8];
1321
1322 /* D16 is only supported on gfx8+ */
1323 assert(!use_format ||
1324 (channel_type != ctx->f16 && channel_type != ctx->i16) ||
1325 ctx->chip_class >= GFX8);
1326
1327 LLVMTypeRef type = func > 1 ? LLVMVectorType(channel_type, func) : channel_type;
1328 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1329
1330 if (use_format) {
1331 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.format.%s",
1332 indexing_kind, type_name);
1333 } else {
1334 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.%s",
1335 indexing_kind, type_name);
1336 }
1337
1338 return ac_build_intrinsic(ctx, name, type, args, idx,
1339 ac_get_load_intr_attribs(can_speculate));
1340 }
1341
1342 LLVMValueRef
1343 ac_build_buffer_load(struct ac_llvm_context *ctx,
1344 LLVMValueRef rsrc,
1345 int num_channels,
1346 LLVMValueRef vindex,
1347 LLVMValueRef voffset,
1348 LLVMValueRef soffset,
1349 unsigned inst_offset,
1350 unsigned cache_policy,
1351 bool can_speculate,
1352 bool allow_smem)
1353 {
1354 LLVMValueRef offset = LLVMConstInt(ctx->i32, inst_offset, 0);
1355 if (voffset)
1356 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1357 if (soffset)
1358 offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
1359
1360 if (allow_smem && !(cache_policy & ac_slc) &&
1361 (!(cache_policy & ac_glc) || ctx->chip_class >= GFX8)) {
1362 assert(vindex == NULL);
1363
1364 LLVMValueRef result[8];
1365
1366 for (int i = 0; i < num_channels; i++) {
1367 if (i) {
1368 offset = LLVMBuildAdd(ctx->builder, offset,
1369 LLVMConstInt(ctx->i32, 4, 0), "");
1370 }
1371 LLVMValueRef args[3] = {
1372 rsrc,
1373 offset,
1374 LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0),
1375 };
1376 result[i] = ac_build_intrinsic(ctx,
1377 "llvm.amdgcn.s.buffer.load.f32",
1378 ctx->f32, args, 3,
1379 AC_FUNC_ATTR_READNONE);
1380 }
1381 if (num_channels == 1)
1382 return result[0];
1383
1384 if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false))
1385 result[num_channels++] = LLVMGetUndef(ctx->f32);
1386 return ac_build_gather_values(ctx, result, num_channels);
1387 }
1388
1389 return ac_build_buffer_load_common(ctx, rsrc, vindex,
1390 offset, ctx->i32_0,
1391 num_channels, ctx->f32,
1392 cache_policy,
1393 can_speculate, false, false);
1394 }
1395
1396 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
1397 LLVMValueRef rsrc,
1398 LLVMValueRef vindex,
1399 LLVMValueRef voffset,
1400 unsigned num_channels,
1401 unsigned cache_policy,
1402 bool can_speculate,
1403 bool d16)
1404 {
1405 return ac_build_buffer_load_common(ctx, rsrc, vindex, voffset,
1406 ctx->i32_0, num_channels,
1407 d16 ? ctx->f16 : ctx->f32,
1408 cache_policy, can_speculate,
1409 true, true);
1410 }
1411
1412 static LLVMValueRef
1413 ac_build_tbuffer_load(struct ac_llvm_context *ctx,
1414 LLVMValueRef rsrc,
1415 LLVMValueRef vindex,
1416 LLVMValueRef voffset,
1417 LLVMValueRef soffset,
1418 LLVMValueRef immoffset,
1419 unsigned num_channels,
1420 unsigned dfmt,
1421 unsigned nfmt,
1422 unsigned cache_policy,
1423 bool can_speculate,
1424 bool structurized)
1425 {
1426 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1427
1428 LLVMValueRef args[6];
1429 int idx = 0;
1430 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1431 if (structurized)
1432 args[idx++] = vindex ? vindex : ctx->i32_0;
1433 args[idx++] = voffset ? voffset : ctx->i32_0;
1434 args[idx++] = soffset ? soffset : ctx->i32_0;
1435 args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx->chip_class, dfmt, nfmt), 0);
1436 args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
1437 unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
1438 const char *indexing_kind = structurized ? "struct" : "raw";
1439 char name[256], type_name[8];
1440
1441 LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
1442 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1443
1444 snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.load.%s",
1445 indexing_kind, type_name);
1446
1447 return ac_build_intrinsic(ctx, name, type, args, idx,
1448 ac_get_load_intr_attribs(can_speculate));
1449 }
1450
1451 LLVMValueRef
1452 ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
1453 LLVMValueRef rsrc,
1454 LLVMValueRef vindex,
1455 LLVMValueRef voffset,
1456 LLVMValueRef soffset,
1457 LLVMValueRef immoffset,
1458 unsigned num_channels,
1459 unsigned dfmt,
1460 unsigned nfmt,
1461 unsigned cache_policy,
1462 bool can_speculate)
1463 {
1464 return ac_build_tbuffer_load(ctx, rsrc, vindex, voffset, soffset,
1465 immoffset, num_channels, dfmt, nfmt,
1466 cache_policy, can_speculate, true);
1467 }
1468
1469 LLVMValueRef
1470 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
1471 LLVMValueRef rsrc,
1472 LLVMValueRef voffset,
1473 LLVMValueRef soffset,
1474 LLVMValueRef immoffset,
1475 unsigned num_channels,
1476 unsigned dfmt,
1477 unsigned nfmt,
1478 unsigned cache_policy,
1479 bool can_speculate)
1480 {
1481 return ac_build_tbuffer_load(ctx, rsrc, NULL, voffset, soffset,
1482 immoffset, num_channels, dfmt, nfmt,
1483 cache_policy, can_speculate, false);
1484 }
1485
1486 LLVMValueRef
1487 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
1488 LLVMValueRef rsrc,
1489 LLVMValueRef voffset,
1490 LLVMValueRef soffset,
1491 LLVMValueRef immoffset,
1492 unsigned cache_policy)
1493 {
1494 LLVMValueRef res;
1495
1496 if (LLVM_VERSION_MAJOR >= 9) {
1497 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1498
1499 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1500 res = ac_build_buffer_load_common(ctx, rsrc, NULL,
1501 voffset, soffset,
1502 1, ctx->i16, cache_policy,
1503 false, false, false);
1504 } else {
1505 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
1506 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1507
1508 res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
1509 immoffset, 1, dfmt, nfmt, cache_policy,
1510 false);
1511
1512 res = LLVMBuildTrunc(ctx->builder, res, ctx->i16, "");
1513 }
1514
1515 return res;
1516 }
1517
1518 LLVMValueRef
1519 ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
1520 LLVMValueRef rsrc,
1521 LLVMValueRef voffset,
1522 LLVMValueRef soffset,
1523 LLVMValueRef immoffset,
1524 unsigned cache_policy)
1525 {
1526 LLVMValueRef res;
1527
1528 if (LLVM_VERSION_MAJOR >= 9) {
1529 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1530
1531 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1532 res = ac_build_buffer_load_common(ctx, rsrc, NULL,
1533 voffset, soffset,
1534 1, ctx->i8, cache_policy,
1535 false, false, false);
1536 } else {
1537 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
1538 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1539
1540 res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
1541 immoffset, 1, dfmt, nfmt, cache_policy,
1542 false);
1543
1544 res = LLVMBuildTrunc(ctx->builder, res, ctx->i8, "");
1545 }
1546
1547 return res;
1548 }
1549
1550 /**
1551 * Convert an 11- or 10-bit unsigned floating point number to an f32.
1552 *
1553 * The input exponent is expected to be biased analogous to IEEE-754, i.e. by
1554 * 2^(exp_bits-1) - 1 (as defined in OpenGL and other graphics APIs).
1555 */
1556 static LLVMValueRef
1557 ac_ufN_to_float(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned exp_bits, unsigned mant_bits)
1558 {
1559 assert(LLVMTypeOf(src) == ctx->i32);
1560
1561 LLVMValueRef tmp;
1562 LLVMValueRef mantissa;
1563 mantissa = LLVMBuildAnd(ctx->builder, src, LLVMConstInt(ctx->i32, (1 << mant_bits) - 1, false), "");
1564
1565 /* Converting normal numbers is just a shift + correcting the exponent bias */
1566 unsigned normal_shift = 23 - mant_bits;
1567 unsigned bias_shift = 127 - ((1 << (exp_bits - 1)) - 1);
1568 LLVMValueRef shifted, normal;
1569
1570 shifted = LLVMBuildShl(ctx->builder, src, LLVMConstInt(ctx->i32, normal_shift, false), "");
1571 normal = LLVMBuildAdd(ctx->builder, shifted, LLVMConstInt(ctx->i32, bias_shift << 23, false), "");
1572
1573 /* Converting nan/inf numbers is the same, but with a different exponent update */
1574 LLVMValueRef naninf;
1575 naninf = LLVMBuildOr(ctx->builder, normal, LLVMConstInt(ctx->i32, 0xff << 23, false), "");
1576
1577 /* Converting denormals is the complex case: determine the leading zeros of the
1578 * mantissa to obtain the correct shift for the mantissa and exponent correction.
1579 */
1580 LLVMValueRef denormal;
1581 LLVMValueRef params[2] = {
1582 mantissa,
1583 ctx->i1true, /* result can be undef when arg is 0 */
1584 };
1585 LLVMValueRef ctlz = ac_build_intrinsic(ctx, "llvm.ctlz.i32", ctx->i32,
1586 params, 2, AC_FUNC_ATTR_READNONE);
1587
1588 /* Shift such that the leading 1 ends up as the LSB of the exponent field. */
1589 tmp = LLVMBuildSub(ctx->builder, ctlz, LLVMConstInt(ctx->i32, 8, false), "");
1590 denormal = LLVMBuildShl(ctx->builder, mantissa, tmp, "");
1591
1592 unsigned denormal_exp = bias_shift + (32 - mant_bits) - 1;
1593 tmp = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, denormal_exp, false), ctlz, "");
1594 tmp = LLVMBuildShl(ctx->builder, tmp, LLVMConstInt(ctx->i32, 23, false), "");
1595 denormal = LLVMBuildAdd(ctx->builder, denormal, tmp, "");
1596
1597 /* Select the final result. */
1598 LLVMValueRef result;
1599
1600 tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src,
1601 LLVMConstInt(ctx->i32, ((1 << exp_bits) - 1) << mant_bits, false), "");
1602 result = LLVMBuildSelect(ctx->builder, tmp, naninf, normal, "");
1603
1604 tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src,
1605 LLVMConstInt(ctx->i32, 1 << mant_bits, false), "");
1606 result = LLVMBuildSelect(ctx->builder, tmp, result, denormal, "");
1607
1608 tmp = LLVMBuildICmp(ctx->builder, LLVMIntNE, src, ctx->i32_0, "");
1609 result = LLVMBuildSelect(ctx->builder, tmp, result, ctx->i32_0, "");
1610
1611 return ac_to_float(ctx, result);
1612 }
1613
1614 /**
1615 * Generate a fully general open coded buffer format fetch with all required
1616 * fixups suitable for vertex fetch, using non-format buffer loads.
1617 *
1618 * Some combinations of argument values have special interpretations:
1619 * - size = 8 bytes, format = fixed indicates PIPE_FORMAT_R11G11B10_FLOAT
1620 * - size = 8 bytes, format != {float,fixed} indicates a 2_10_10_10 data format
1621 *
1622 * \param log_size log(size of channel in bytes)
1623 * \param num_channels number of channels (1 to 4)
1624 * \param format AC_FETCH_FORMAT_xxx value
1625 * \param reverse whether XYZ channels are reversed
1626 * \param known_aligned whether the source is known to be aligned to hardware's
1627 * effective element size for loading the given format
1628 * (note: this means dword alignment for 8_8_8_8, 16_16, etc.)
1629 * \param rsrc buffer resource descriptor
1630 * \return the resulting vector of floats or integers bitcast to <4 x i32>
1631 */
1632 LLVMValueRef
1633 ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
1634 unsigned log_size,
1635 unsigned num_channels,
1636 unsigned format,
1637 bool reverse,
1638 bool known_aligned,
1639 LLVMValueRef rsrc,
1640 LLVMValueRef vindex,
1641 LLVMValueRef voffset,
1642 LLVMValueRef soffset,
1643 unsigned cache_policy,
1644 bool can_speculate)
1645 {
1646 LLVMValueRef tmp;
1647 unsigned load_log_size = log_size;
1648 unsigned load_num_channels = num_channels;
1649 if (log_size == 3) {
1650 load_log_size = 2;
1651 if (format == AC_FETCH_FORMAT_FLOAT) {
1652 load_num_channels = 2 * num_channels;
1653 } else {
1654 load_num_channels = 1; /* 10_11_11 or 2_10_10_10 */
1655 }
1656 }
1657
1658 int log_recombine = 0;
1659 if ((ctx->chip_class == GFX6 || ctx->chip_class >= GFX10) && !known_aligned) {
1660 /* Avoid alignment restrictions by loading one byte at a time. */
1661 load_num_channels <<= load_log_size;
1662 log_recombine = load_log_size;
1663 load_log_size = 0;
1664 } else if (load_num_channels == 2 || load_num_channels == 4) {
1665 log_recombine = -util_logbase2(load_num_channels);
1666 load_num_channels = 1;
1667 load_log_size += -log_recombine;
1668 }
1669
1670 assert(load_log_size >= 2 || LLVM_VERSION_MAJOR >= 9);
1671
1672 LLVMValueRef loads[32]; /* up to 32 bytes */
1673 for (unsigned i = 0; i < load_num_channels; ++i) {
1674 tmp = LLVMBuildAdd(ctx->builder, soffset,
1675 LLVMConstInt(ctx->i32, i << load_log_size, false), "");
1676 LLVMTypeRef channel_type = load_log_size == 0 ? ctx->i8 :
1677 load_log_size == 1 ? ctx->i16 : ctx->i32;
1678 unsigned num_channels = 1 << (MAX2(load_log_size, 2) - 2);
1679 loads[i] = ac_build_buffer_load_common(
1680 ctx, rsrc, vindex, voffset, tmp,
1681 num_channels, channel_type, cache_policy,
1682 can_speculate, false, true);
1683 if (load_log_size >= 2)
1684 loads[i] = ac_to_integer(ctx, loads[i]);
1685 }
1686
1687 if (log_recombine > 0) {
1688 /* Recombine bytes if necessary (GFX6 only) */
1689 LLVMTypeRef dst_type = log_recombine == 2 ? ctx->i32 : ctx->i16;
1690
1691 for (unsigned src = 0, dst = 0; src < load_num_channels; ++dst) {
1692 LLVMValueRef accum = NULL;
1693 for (unsigned i = 0; i < (1 << log_recombine); ++i, ++src) {
1694 tmp = LLVMBuildZExt(ctx->builder, loads[src], dst_type, "");
1695 if (i == 0) {
1696 accum = tmp;
1697 } else {
1698 tmp = LLVMBuildShl(ctx->builder, tmp,
1699 LLVMConstInt(dst_type, 8 * i, false), "");
1700 accum = LLVMBuildOr(ctx->builder, accum, tmp, "");
1701 }
1702 }
1703 loads[dst] = accum;
1704 }
1705 } else if (log_recombine < 0) {
1706 /* Split vectors of dwords */
1707 if (load_log_size > 2) {
1708 assert(load_num_channels == 1);
1709 LLVMValueRef loaded = loads[0];
1710 unsigned log_split = load_log_size - 2;
1711 log_recombine += log_split;
1712 load_num_channels = 1 << log_split;
1713 load_log_size = 2;
1714 for (unsigned i = 0; i < load_num_channels; ++i) {
1715 tmp = LLVMConstInt(ctx->i32, i, false);
1716 loads[i] = LLVMBuildExtractElement(ctx->builder, loaded, tmp, "");
1717 }
1718 }
1719
1720 /* Further split dwords and shorts if required */
1721 if (log_recombine < 0) {
1722 for (unsigned src = load_num_channels,
1723 dst = load_num_channels << -log_recombine;
1724 src > 0; --src) {
1725 unsigned dst_bits = 1 << (3 + load_log_size + log_recombine);
1726 LLVMTypeRef dst_type = LLVMIntTypeInContext(ctx->context, dst_bits);
1727 LLVMValueRef loaded = loads[src - 1];
1728 LLVMTypeRef loaded_type = LLVMTypeOf(loaded);
1729 for (unsigned i = 1 << -log_recombine; i > 0; --i, --dst) {
1730 tmp = LLVMConstInt(loaded_type, dst_bits * (i - 1), false);
1731 tmp = LLVMBuildLShr(ctx->builder, loaded, tmp, "");
1732 loads[dst - 1] = LLVMBuildTrunc(ctx->builder, tmp, dst_type, "");
1733 }
1734 }
1735 }
1736 }
1737
1738 if (log_size == 3) {
1739 if (format == AC_FETCH_FORMAT_FLOAT) {
1740 for (unsigned i = 0; i < num_channels; ++i) {
1741 tmp = ac_build_gather_values(ctx, &loads[2 * i], 2);
1742 loads[i] = LLVMBuildBitCast(ctx->builder, tmp, ctx->f64, "");
1743 }
1744 } else if (format == AC_FETCH_FORMAT_FIXED) {
1745 /* 10_11_11_FLOAT */
1746 LLVMValueRef data = loads[0];
1747 LLVMValueRef i32_2047 = LLVMConstInt(ctx->i32, 2047, false);
1748 LLVMValueRef r = LLVMBuildAnd(ctx->builder, data, i32_2047, "");
1749 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 11, false), "");
1750 LLVMValueRef g = LLVMBuildAnd(ctx->builder, tmp, i32_2047, "");
1751 LLVMValueRef b = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 22, false), "");
1752
1753 loads[0] = ac_to_integer(ctx, ac_ufN_to_float(ctx, r, 5, 6));
1754 loads[1] = ac_to_integer(ctx, ac_ufN_to_float(ctx, g, 5, 6));
1755 loads[2] = ac_to_integer(ctx, ac_ufN_to_float(ctx, b, 5, 5));
1756
1757 num_channels = 3;
1758 log_size = 2;
1759 format = AC_FETCH_FORMAT_FLOAT;
1760 } else {
1761 /* 2_10_10_10 data formats */
1762 LLVMValueRef data = loads[0];
1763 LLVMTypeRef i10 = LLVMIntTypeInContext(ctx->context, 10);
1764 LLVMTypeRef i2 = LLVMIntTypeInContext(ctx->context, 2);
1765 loads[0] = LLVMBuildTrunc(ctx->builder, data, i10, "");
1766 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 10, false), "");
1767 loads[1] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
1768 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 20, false), "");
1769 loads[2] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
1770 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 30, false), "");
1771 loads[3] = LLVMBuildTrunc(ctx->builder, tmp, i2, "");
1772
1773 num_channels = 4;
1774 }
1775 }
1776
1777 if (format == AC_FETCH_FORMAT_FLOAT) {
1778 if (log_size != 2) {
1779 for (unsigned chan = 0; chan < num_channels; ++chan) {
1780 tmp = ac_to_float(ctx, loads[chan]);
1781 if (log_size == 3)
1782 tmp = LLVMBuildFPTrunc(ctx->builder, tmp, ctx->f32, "");
1783 else if (log_size == 1)
1784 tmp = LLVMBuildFPExt(ctx->builder, tmp, ctx->f32, "");
1785 loads[chan] = ac_to_integer(ctx, tmp);
1786 }
1787 }
1788 } else if (format == AC_FETCH_FORMAT_UINT) {
1789 if (log_size != 2) {
1790 for (unsigned chan = 0; chan < num_channels; ++chan)
1791 loads[chan] = LLVMBuildZExt(ctx->builder, loads[chan], ctx->i32, "");
1792 }
1793 } else if (format == AC_FETCH_FORMAT_SINT) {
1794 if (log_size != 2) {
1795 for (unsigned chan = 0; chan < num_channels; ++chan)
1796 loads[chan] = LLVMBuildSExt(ctx->builder, loads[chan], ctx->i32, "");
1797 }
1798 } else {
1799 bool unsign = format == AC_FETCH_FORMAT_UNORM ||
1800 format == AC_FETCH_FORMAT_USCALED ||
1801 format == AC_FETCH_FORMAT_UINT;
1802
1803 for (unsigned chan = 0; chan < num_channels; ++chan) {
1804 if (unsign) {
1805 tmp = LLVMBuildUIToFP(ctx->builder, loads[chan], ctx->f32, "");
1806 } else {
1807 tmp = LLVMBuildSIToFP(ctx->builder, loads[chan], ctx->f32, "");
1808 }
1809
1810 LLVMValueRef scale = NULL;
1811 if (format == AC_FETCH_FORMAT_FIXED) {
1812 assert(log_size == 2);
1813 scale = LLVMConstReal(ctx->f32, 1.0 / 0x10000);
1814 } else if (format == AC_FETCH_FORMAT_UNORM) {
1815 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
1816 scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << bits) - 1));
1817 } else if (format == AC_FETCH_FORMAT_SNORM) {
1818 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
1819 scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << (bits - 1)) - 1));
1820 }
1821 if (scale)
1822 tmp = LLVMBuildFMul(ctx->builder, tmp, scale, "");
1823
1824 if (format == AC_FETCH_FORMAT_SNORM) {
1825 /* Clamp to [-1, 1] */
1826 LLVMValueRef neg_one = LLVMConstReal(ctx->f32, -1.0);
1827 LLVMValueRef clamp =
1828 LLVMBuildFCmp(ctx->builder, LLVMRealULT, tmp, neg_one, "");
1829 tmp = LLVMBuildSelect(ctx->builder, clamp, neg_one, tmp, "");
1830 }
1831
1832 loads[chan] = ac_to_integer(ctx, tmp);
1833 }
1834 }
1835
1836 while (num_channels < 4) {
1837 if (format == AC_FETCH_FORMAT_UINT || format == AC_FETCH_FORMAT_SINT) {
1838 loads[num_channels] = num_channels == 3 ? ctx->i32_1 : ctx->i32_0;
1839 } else {
1840 loads[num_channels] = ac_to_integer(ctx, num_channels == 3 ? ctx->f32_1 : ctx->f32_0);
1841 }
1842 num_channels++;
1843 }
1844
1845 if (reverse) {
1846 tmp = loads[0];
1847 loads[0] = loads[2];
1848 loads[2] = tmp;
1849 }
1850
1851 return ac_build_gather_values(ctx, loads, 4);
1852 }
1853
1854 static void
1855 ac_build_tbuffer_store(struct ac_llvm_context *ctx,
1856 LLVMValueRef rsrc,
1857 LLVMValueRef vdata,
1858 LLVMValueRef vindex,
1859 LLVMValueRef voffset,
1860 LLVMValueRef soffset,
1861 LLVMValueRef immoffset,
1862 unsigned num_channels,
1863 unsigned dfmt,
1864 unsigned nfmt,
1865 unsigned cache_policy,
1866 bool structurized)
1867 {
1868 voffset = LLVMBuildAdd(ctx->builder, voffset ? voffset : ctx->i32_0,
1869 immoffset, "");
1870
1871 LLVMValueRef args[7];
1872 int idx = 0;
1873 args[idx++] = vdata;
1874 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1875 if (structurized)
1876 args[idx++] = vindex ? vindex : ctx->i32_0;
1877 args[idx++] = voffset ? voffset : ctx->i32_0;
1878 args[idx++] = soffset ? soffset : ctx->i32_0;
1879 args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx->chip_class, dfmt, nfmt), 0);
1880 args[idx++] = LLVMConstInt(ctx->i32, cache_policy, 0);
1881 unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
1882 const char *indexing_kind = structurized ? "struct" : "raw";
1883 char name[256], type_name[8];
1884
1885 LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
1886 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1887
1888 snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.store.%s",
1889 indexing_kind, type_name);
1890
1891 ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
1892 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
1893 }
1894
1895 void
1896 ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
1897 LLVMValueRef rsrc,
1898 LLVMValueRef vdata,
1899 LLVMValueRef vindex,
1900 LLVMValueRef voffset,
1901 LLVMValueRef soffset,
1902 LLVMValueRef immoffset,
1903 unsigned num_channels,
1904 unsigned dfmt,
1905 unsigned nfmt,
1906 unsigned cache_policy)
1907 {
1908 ac_build_tbuffer_store(ctx, rsrc, vdata, vindex, voffset, soffset,
1909 immoffset, num_channels, dfmt, nfmt, cache_policy,
1910 true);
1911 }
1912
1913 void
1914 ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
1915 LLVMValueRef rsrc,
1916 LLVMValueRef vdata,
1917 LLVMValueRef voffset,
1918 LLVMValueRef soffset,
1919 LLVMValueRef immoffset,
1920 unsigned num_channels,
1921 unsigned dfmt,
1922 unsigned nfmt,
1923 unsigned cache_policy)
1924 {
1925 ac_build_tbuffer_store(ctx, rsrc, vdata, NULL, voffset, soffset,
1926 immoffset, num_channels, dfmt, nfmt, cache_policy,
1927 false);
1928 }
1929
1930 void
1931 ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
1932 LLVMValueRef rsrc,
1933 LLVMValueRef vdata,
1934 LLVMValueRef voffset,
1935 LLVMValueRef soffset,
1936 unsigned cache_policy)
1937 {
1938 vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i16, "");
1939
1940 if (LLVM_VERSION_MAJOR >= 9) {
1941 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1942 ac_build_buffer_store_common(ctx, rsrc, vdata, NULL,
1943 voffset, soffset, cache_policy,
1944 false, false);
1945 } else {
1946 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
1947 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1948
1949 vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
1950
1951 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1952 ctx->i32_0, 1, dfmt, nfmt, cache_policy);
1953 }
1954 }
1955
1956 void
1957 ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
1958 LLVMValueRef rsrc,
1959 LLVMValueRef vdata,
1960 LLVMValueRef voffset,
1961 LLVMValueRef soffset,
1962 unsigned cache_policy)
1963 {
1964 vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i8, "");
1965
1966 if (LLVM_VERSION_MAJOR >= 9) {
1967 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1968 ac_build_buffer_store_common(ctx, rsrc, vdata, NULL,
1969 voffset, soffset, cache_policy,
1970 false, false);
1971 } else {
1972 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
1973 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1974
1975 vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
1976
1977 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1978 ctx->i32_0, 1, dfmt, nfmt, cache_policy);
1979 }
1980 }
1981 /**
1982 * Set range metadata on an instruction. This can only be used on load and
1983 * call instructions. If you know an instruction can only produce the values
1984 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1985 * \p lo is the minimum value inclusive.
1986 * \p hi is the maximum value exclusive.
1987 */
1988 static void set_range_metadata(struct ac_llvm_context *ctx,
1989 LLVMValueRef value, unsigned lo, unsigned hi)
1990 {
1991 LLVMValueRef range_md, md_args[2];
1992 LLVMTypeRef type = LLVMTypeOf(value);
1993 LLVMContextRef context = LLVMGetTypeContext(type);
1994
1995 md_args[0] = LLVMConstInt(type, lo, false);
1996 md_args[1] = LLVMConstInt(type, hi, false);
1997 range_md = LLVMMDNodeInContext(context, md_args, 2);
1998 LLVMSetMetadata(value, ctx->range_md_kind, range_md);
1999 }
2000
2001 LLVMValueRef
2002 ac_get_thread_id(struct ac_llvm_context *ctx)
2003 {
2004 LLVMValueRef tid;
2005
2006 LLVMValueRef tid_args[2];
2007 tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
2008 tid_args[1] = ctx->i32_0;
2009 tid_args[1] = ac_build_intrinsic(ctx,
2010 "llvm.amdgcn.mbcnt.lo", ctx->i32,
2011 tid_args, 2, AC_FUNC_ATTR_READNONE);
2012
2013 if (ctx->wave_size == 32) {
2014 tid = tid_args[1];
2015 } else {
2016 tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
2017 ctx->i32, tid_args,
2018 2, AC_FUNC_ATTR_READNONE);
2019 }
2020 set_range_metadata(ctx, tid, 0, ctx->wave_size);
2021 return tid;
2022 }
2023
2024 /*
2025 * AMD GCN implements derivatives using the local data store (LDS)
2026 * All writes to the LDS happen in all executing threads at
2027 * the same time. TID is the Thread ID for the current
2028 * thread and is a value between 0 and 63, representing
2029 * the thread's position in the wavefront.
2030 *
2031 * For the pixel shader threads are grouped into quads of four pixels.
2032 * The TIDs of the pixels of a quad are:
2033 *
2034 * +------+------+
2035 * |4n + 0|4n + 1|
2036 * +------+------+
2037 * |4n + 2|4n + 3|
2038 * +------+------+
2039 *
2040 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
2041 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
2042 * the current pixel's column, and masking with 0xfffffffe yields the TID
2043 * of the left pixel of the current pixel's row.
2044 *
2045 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
2046 * adding 2 yields the TID of the pixel below the top pixel.
2047 */
2048 LLVMValueRef
2049 ac_build_ddxy(struct ac_llvm_context *ctx,
2050 uint32_t mask,
2051 int idx,
2052 LLVMValueRef val)
2053 {
2054 unsigned tl_lanes[4], trbl_lanes[4];
2055 char name[32], type[8];
2056 LLVMValueRef tl, trbl;
2057 LLVMTypeRef result_type;
2058 LLVMValueRef result;
2059
2060 result_type = ac_to_float_type(ctx, LLVMTypeOf(val));
2061
2062 if (result_type == ctx->f16)
2063 val = LLVMBuildZExt(ctx->builder, val, ctx->i32, "");
2064 else if (result_type == ctx->v2f16)
2065 val = LLVMBuildBitCast(ctx->builder, val, ctx->i32, "");
2066
2067 for (unsigned i = 0; i < 4; ++i) {
2068 tl_lanes[i] = i & mask;
2069 trbl_lanes[i] = (i & mask) + idx;
2070 }
2071
2072 tl = ac_build_quad_swizzle(ctx, val,
2073 tl_lanes[0], tl_lanes[1],
2074 tl_lanes[2], tl_lanes[3]);
2075 trbl = ac_build_quad_swizzle(ctx, val,
2076 trbl_lanes[0], trbl_lanes[1],
2077 trbl_lanes[2], trbl_lanes[3]);
2078
2079 if (result_type == ctx->f16) {
2080 tl = LLVMBuildTrunc(ctx->builder, tl, ctx->i16, "");
2081 trbl = LLVMBuildTrunc(ctx->builder, trbl, ctx->i16, "");
2082 }
2083
2084 tl = LLVMBuildBitCast(ctx->builder, tl, result_type, "");
2085 trbl = LLVMBuildBitCast(ctx->builder, trbl, result_type, "");
2086 result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
2087
2088 ac_build_type_name_for_intr(result_type, type, sizeof(type));
2089 snprintf(name, sizeof(name), "llvm.amdgcn.wqm.%s", type);
2090
2091 return ac_build_intrinsic(ctx, name, result_type, &result, 1, 0);
2092 }
2093
2094 void
2095 ac_build_sendmsg(struct ac_llvm_context *ctx,
2096 uint32_t msg,
2097 LLVMValueRef wave_id)
2098 {
2099 LLVMValueRef args[2];
2100 args[0] = LLVMConstInt(ctx->i32, msg, false);
2101 args[1] = wave_id;
2102 ac_build_intrinsic(ctx, "llvm.amdgcn.s.sendmsg", ctx->voidt, args, 2, 0);
2103 }
2104
2105 LLVMValueRef
2106 ac_build_imsb(struct ac_llvm_context *ctx,
2107 LLVMValueRef arg,
2108 LLVMTypeRef dst_type)
2109 {
2110 LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.amdgcn.sffbh.i32",
2111 dst_type, &arg, 1,
2112 AC_FUNC_ATTR_READNONE);
2113
2114 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
2115 * the index from LSB. Invert it by doing "31 - msb". */
2116 msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
2117 msb, "");
2118
2119 LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
2120 LLVMValueRef cond = LLVMBuildOr(ctx->builder,
2121 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
2122 arg, ctx->i32_0, ""),
2123 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
2124 arg, all_ones, ""), "");
2125
2126 return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
2127 }
2128
2129 LLVMValueRef
2130 ac_build_umsb(struct ac_llvm_context *ctx,
2131 LLVMValueRef arg,
2132 LLVMTypeRef dst_type)
2133 {
2134 const char *intrin_name;
2135 LLVMTypeRef type;
2136 LLVMValueRef highest_bit;
2137 LLVMValueRef zero;
2138 unsigned bitsize;
2139
2140 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(arg));
2141 switch (bitsize) {
2142 case 64:
2143 intrin_name = "llvm.ctlz.i64";
2144 type = ctx->i64;
2145 highest_bit = LLVMConstInt(ctx->i64, 63, false);
2146 zero = ctx->i64_0;
2147 break;
2148 case 32:
2149 intrin_name = "llvm.ctlz.i32";
2150 type = ctx->i32;
2151 highest_bit = LLVMConstInt(ctx->i32, 31, false);
2152 zero = ctx->i32_0;
2153 break;
2154 case 16:
2155 intrin_name = "llvm.ctlz.i16";
2156 type = ctx->i16;
2157 highest_bit = LLVMConstInt(ctx->i16, 15, false);
2158 zero = ctx->i16_0;
2159 break;
2160 case 8:
2161 intrin_name = "llvm.ctlz.i8";
2162 type = ctx->i8;
2163 highest_bit = LLVMConstInt(ctx->i8, 7, false);
2164 zero = ctx->i8_0;
2165 break;
2166 default:
2167 unreachable(!"invalid bitsize");
2168 break;
2169 }
2170
2171 LLVMValueRef params[2] = {
2172 arg,
2173 ctx->i1true,
2174 };
2175
2176 LLVMValueRef msb = ac_build_intrinsic(ctx, intrin_name, type,
2177 params, 2,
2178 AC_FUNC_ATTR_READNONE);
2179
2180 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
2181 * the index from LSB. Invert it by doing "31 - msb". */
2182 msb = LLVMBuildSub(ctx->builder, highest_bit, msb, "");
2183
2184 if (bitsize == 64) {
2185 msb = LLVMBuildTrunc(ctx->builder, msb, ctx->i32, "");
2186 } else if (bitsize < 32) {
2187 msb = LLVMBuildSExt(ctx->builder, msb, ctx->i32, "");
2188 }
2189
2190 /* check for zero */
2191 return LLVMBuildSelect(ctx->builder,
2192 LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg, zero, ""),
2193 LLVMConstInt(ctx->i32, -1, true), msb, "");
2194 }
2195
2196 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
2197 LLVMValueRef b)
2198 {
2199 char name[64], type[64];
2200
2201 ac_build_type_name_for_intr(LLVMTypeOf(a), type, sizeof(type));
2202 snprintf(name, sizeof(name), "llvm.minnum.%s", type);
2203 LLVMValueRef args[2] = {a, b};
2204 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
2205 AC_FUNC_ATTR_READNONE);
2206 }
2207
2208 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
2209 LLVMValueRef b)
2210 {
2211 char name[64], type[64];
2212
2213 ac_build_type_name_for_intr(LLVMTypeOf(a), type, sizeof(type));
2214 snprintf(name, sizeof(name), "llvm.maxnum.%s", type);
2215 LLVMValueRef args[2] = {a, b};
2216 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
2217 AC_FUNC_ATTR_READNONE);
2218 }
2219
2220 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
2221 LLVMValueRef b)
2222 {
2223 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSLE, a, b, "");
2224 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2225 }
2226
2227 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
2228 LLVMValueRef b)
2229 {
2230 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, a, b, "");
2231 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2232 }
2233
2234 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a,
2235 LLVMValueRef b)
2236 {
2237 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntULE, a, b, "");
2238 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2239 }
2240
2241 LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a,
2242 LLVMValueRef b)
2243 {
2244 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, a, b, "");
2245 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2246 }
2247
2248 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value)
2249 {
2250 LLVMTypeRef t = LLVMTypeOf(value);
2251 return ac_build_fmin(ctx, ac_build_fmax(ctx, value, LLVMConstReal(t, 0.0)),
2252 LLVMConstReal(t, 1.0));
2253 }
2254
2255 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a)
2256 {
2257 LLVMValueRef args[9];
2258
2259 args[0] = LLVMConstInt(ctx->i32, a->target, 0);
2260 args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
2261
2262 if (a->compr) {
2263 args[2] = LLVMBuildBitCast(ctx->builder, a->out[0],
2264 ctx->v2i16, "");
2265 args[3] = LLVMBuildBitCast(ctx->builder, a->out[1],
2266 ctx->v2i16, "");
2267 args[4] = LLVMConstInt(ctx->i1, a->done, 0);
2268 args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
2269
2270 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16",
2271 ctx->voidt, args, 6, 0);
2272 } else {
2273 args[2] = a->out[0];
2274 args[3] = a->out[1];
2275 args[4] = a->out[2];
2276 args[5] = a->out[3];
2277 args[6] = LLVMConstInt(ctx->i1, a->done, 0);
2278 args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
2279
2280 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.f32",
2281 ctx->voidt, args, 8, 0);
2282 }
2283 }
2284
2285 void ac_build_export_null(struct ac_llvm_context *ctx)
2286 {
2287 struct ac_export_args args;
2288
2289 args.enabled_channels = 0x0; /* enabled channels */
2290 args.valid_mask = 1; /* whether the EXEC mask is valid */
2291 args.done = 1; /* DONE bit */
2292 args.target = V_008DFC_SQ_EXP_NULL;
2293 args.compr = 0; /* COMPR flag (0 = 32-bit export) */
2294 args.out[0] = LLVMGetUndef(ctx->f32); /* R */
2295 args.out[1] = LLVMGetUndef(ctx->f32); /* G */
2296 args.out[2] = LLVMGetUndef(ctx->f32); /* B */
2297 args.out[3] = LLVMGetUndef(ctx->f32); /* A */
2298
2299 ac_build_export(ctx, &args);
2300 }
2301
2302 static unsigned ac_num_coords(enum ac_image_dim dim)
2303 {
2304 switch (dim) {
2305 case ac_image_1d:
2306 return 1;
2307 case ac_image_2d:
2308 case ac_image_1darray:
2309 return 2;
2310 case ac_image_3d:
2311 case ac_image_cube:
2312 case ac_image_2darray:
2313 case ac_image_2dmsaa:
2314 return 3;
2315 case ac_image_2darraymsaa:
2316 return 4;
2317 default:
2318 unreachable("ac_num_coords: bad dim");
2319 }
2320 }
2321
2322 static unsigned ac_num_derivs(enum ac_image_dim dim)
2323 {
2324 switch (dim) {
2325 case ac_image_1d:
2326 case ac_image_1darray:
2327 return 2;
2328 case ac_image_2d:
2329 case ac_image_2darray:
2330 case ac_image_cube:
2331 return 4;
2332 case ac_image_3d:
2333 return 6;
2334 case ac_image_2dmsaa:
2335 case ac_image_2darraymsaa:
2336 default:
2337 unreachable("derivatives not supported");
2338 }
2339 }
2340
2341 static const char *get_atomic_name(enum ac_atomic_op op)
2342 {
2343 switch (op) {
2344 case ac_atomic_swap: return "swap";
2345 case ac_atomic_add: return "add";
2346 case ac_atomic_sub: return "sub";
2347 case ac_atomic_smin: return "smin";
2348 case ac_atomic_umin: return "umin";
2349 case ac_atomic_smax: return "smax";
2350 case ac_atomic_umax: return "umax";
2351 case ac_atomic_and: return "and";
2352 case ac_atomic_or: return "or";
2353 case ac_atomic_xor: return "xor";
2354 case ac_atomic_inc_wrap: return "inc";
2355 case ac_atomic_dec_wrap: return "dec";
2356 }
2357 unreachable("bad atomic op");
2358 }
2359
2360 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
2361 struct ac_image_args *a)
2362 {
2363 const char *overload[3] = { "", "", "" };
2364 unsigned num_overloads = 0;
2365 LLVMValueRef args[18];
2366 unsigned num_args = 0;
2367 enum ac_image_dim dim = a->dim;
2368
2369 assert(!a->lod || a->lod == ctx->i32_0 || a->lod == ctx->f32_0 ||
2370 !a->level_zero);
2371 assert((a->opcode != ac_image_get_resinfo && a->opcode != ac_image_load_mip &&
2372 a->opcode != ac_image_store_mip) ||
2373 a->lod);
2374 assert(a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
2375 (!a->compare && !a->offset));
2376 assert((a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
2377 a->opcode == ac_image_get_lod) ||
2378 !a->bias);
2379 assert((a->bias ? 1 : 0) +
2380 (a->lod ? 1 : 0) +
2381 (a->level_zero ? 1 : 0) +
2382 (a->derivs[0] ? 1 : 0) <= 1);
2383 assert((a->min_lod ? 1 : 0) +
2384 (a->lod ? 1 : 0) +
2385 (a->level_zero ? 1 : 0) <= 1);
2386 assert(!a->d16 || (ctx->chip_class >= GFX8 &&
2387 a->opcode != ac_image_atomic &&
2388 a->opcode != ac_image_atomic_cmpswap &&
2389 a->opcode != ac_image_get_lod &&
2390 a->opcode != ac_image_get_resinfo));
2391
2392 if (a->opcode == ac_image_get_lod) {
2393 switch (dim) {
2394 case ac_image_1darray:
2395 dim = ac_image_1d;
2396 break;
2397 case ac_image_2darray:
2398 case ac_image_cube:
2399 dim = ac_image_2d;
2400 break;
2401 default:
2402 break;
2403 }
2404 }
2405
2406 bool sample = a->opcode == ac_image_sample ||
2407 a->opcode == ac_image_gather4 ||
2408 a->opcode == ac_image_get_lod;
2409 bool atomic = a->opcode == ac_image_atomic ||
2410 a->opcode == ac_image_atomic_cmpswap;
2411 bool load = a->opcode == ac_image_sample ||
2412 a->opcode == ac_image_gather4 ||
2413 a->opcode == ac_image_load ||
2414 a->opcode == ac_image_load_mip;
2415 LLVMTypeRef coord_type = sample ? ctx->f32 : ctx->i32;
2416
2417 if (atomic || a->opcode == ac_image_store || a->opcode == ac_image_store_mip) {
2418 args[num_args++] = a->data[0];
2419 if (a->opcode == ac_image_atomic_cmpswap)
2420 args[num_args++] = a->data[1];
2421 }
2422
2423 if (!atomic)
2424 args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, false);
2425
2426 if (a->offset)
2427 args[num_args++] = ac_to_integer(ctx, a->offset);
2428 if (a->bias) {
2429 args[num_args++] = ac_to_float(ctx, a->bias);
2430 overload[num_overloads++] = ".f32";
2431 }
2432 if (a->compare)
2433 args[num_args++] = ac_to_float(ctx, a->compare);
2434 if (a->derivs[0]) {
2435 unsigned count = ac_num_derivs(dim);
2436 for (unsigned i = 0; i < count; ++i)
2437 args[num_args++] = ac_to_float(ctx, a->derivs[i]);
2438 overload[num_overloads++] = ".f32";
2439 }
2440 unsigned num_coords =
2441 a->opcode != ac_image_get_resinfo ? ac_num_coords(dim) : 0;
2442 for (unsigned i = 0; i < num_coords; ++i)
2443 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->coords[i], coord_type, "");
2444 if (a->lod)
2445 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->lod, coord_type, "");
2446 if (a->min_lod)
2447 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->min_lod, coord_type, "");
2448
2449 overload[num_overloads++] = sample ? ".f32" : ".i32";
2450
2451 args[num_args++] = a->resource;
2452 if (sample) {
2453 args[num_args++] = a->sampler;
2454 args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, false);
2455 }
2456
2457 args[num_args++] = ctx->i32_0; /* texfailctrl */
2458 args[num_args++] = LLVMConstInt(ctx->i32,
2459 load ? get_load_cache_policy(ctx, a->cache_policy) :
2460 a->cache_policy, false);
2461
2462 const char *name;
2463 const char *atomic_subop = "";
2464 switch (a->opcode) {
2465 case ac_image_sample: name = "sample"; break;
2466 case ac_image_gather4: name = "gather4"; break;
2467 case ac_image_load: name = "load"; break;
2468 case ac_image_load_mip: name = "load.mip"; break;
2469 case ac_image_store: name = "store"; break;
2470 case ac_image_store_mip: name = "store.mip"; break;
2471 case ac_image_atomic:
2472 name = "atomic.";
2473 atomic_subop = get_atomic_name(a->atomic);
2474 break;
2475 case ac_image_atomic_cmpswap:
2476 name = "atomic.";
2477 atomic_subop = "cmpswap";
2478 break;
2479 case ac_image_get_lod: name = "getlod"; break;
2480 case ac_image_get_resinfo: name = "getresinfo"; break;
2481 default: unreachable("invalid image opcode");
2482 }
2483
2484 const char *dimname;
2485 switch (dim) {
2486 case ac_image_1d: dimname = "1d"; break;
2487 case ac_image_2d: dimname = "2d"; break;
2488 case ac_image_3d: dimname = "3d"; break;
2489 case ac_image_cube: dimname = "cube"; break;
2490 case ac_image_1darray: dimname = "1darray"; break;
2491 case ac_image_2darray: dimname = "2darray"; break;
2492 case ac_image_2dmsaa: dimname = "2dmsaa"; break;
2493 case ac_image_2darraymsaa: dimname = "2darraymsaa"; break;
2494 default: unreachable("invalid dim");
2495 }
2496
2497 bool lod_suffix =
2498 a->lod && (a->opcode == ac_image_sample || a->opcode == ac_image_gather4);
2499 char intr_name[96];
2500 snprintf(intr_name, sizeof(intr_name),
2501 "llvm.amdgcn.image.%s%s" /* base name */
2502 "%s%s%s%s" /* sample/gather modifiers */
2503 ".%s.%s%s%s%s", /* dimension and type overloads */
2504 name, atomic_subop,
2505 a->compare ? ".c" : "",
2506 a->bias ? ".b" :
2507 lod_suffix ? ".l" :
2508 a->derivs[0] ? ".d" :
2509 a->level_zero ? ".lz" : "",
2510 a->min_lod ? ".cl" : "",
2511 a->offset ? ".o" : "",
2512 dimname,
2513 atomic ? "i32" : (a->d16 ? "v4f16" : "v4f32"),
2514 overload[0], overload[1], overload[2]);
2515
2516 LLVMTypeRef retty;
2517 if (atomic)
2518 retty = ctx->i32;
2519 else if (a->opcode == ac_image_store || a->opcode == ac_image_store_mip)
2520 retty = ctx->voidt;
2521 else
2522 retty = a->d16 ? ctx->v4f16 : ctx->v4f32;
2523
2524 LLVMValueRef result =
2525 ac_build_intrinsic(ctx, intr_name, retty, args, num_args,
2526 a->attributes);
2527 if (!sample && !atomic && retty != ctx->voidt)
2528 result = ac_to_integer(ctx, result);
2529
2530 return result;
2531 }
2532
2533 LLVMValueRef ac_build_image_get_sample_count(struct ac_llvm_context *ctx,
2534 LLVMValueRef rsrc)
2535 {
2536 LLVMValueRef samples;
2537
2538 /* Read the samples from the descriptor directly.
2539 * Hardware doesn't have any instruction for this.
2540 */
2541 samples = LLVMBuildExtractElement(ctx->builder, rsrc,
2542 LLVMConstInt(ctx->i32, 3, 0), "");
2543 samples = LLVMBuildLShr(ctx->builder, samples,
2544 LLVMConstInt(ctx->i32, 16, 0), "");
2545 samples = LLVMBuildAnd(ctx->builder, samples,
2546 LLVMConstInt(ctx->i32, 0xf, 0), "");
2547 samples = LLVMBuildShl(ctx->builder, ctx->i32_1,
2548 samples, "");
2549 return samples;
2550 }
2551
2552 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
2553 LLVMValueRef args[2])
2554 {
2555 return ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz", ctx->v2f16,
2556 args, 2, AC_FUNC_ATTR_READNONE);
2557 }
2558
2559 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
2560 LLVMValueRef args[2])
2561 {
2562 LLVMValueRef res =
2563 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16",
2564 ctx->v2i16, args, 2,
2565 AC_FUNC_ATTR_READNONE);
2566 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2567 }
2568
2569 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
2570 LLVMValueRef args[2])
2571 {
2572 LLVMValueRef res =
2573 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16",
2574 ctx->v2i16, args, 2,
2575 AC_FUNC_ATTR_READNONE);
2576 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2577 }
2578
2579 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2580 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
2581 LLVMValueRef args[2], unsigned bits, bool hi)
2582 {
2583 assert(bits == 8 || bits == 10 || bits == 16);
2584
2585 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2586 bits == 8 ? 127 : bits == 10 ? 511 : 32767, 0);
2587 LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
2588 bits == 8 ? -128 : bits == 10 ? -512 : -32768, 0);
2589 LLVMValueRef max_alpha =
2590 bits != 10 ? max_rgb : ctx->i32_1;
2591 LLVMValueRef min_alpha =
2592 bits != 10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
2593
2594 /* Clamp. */
2595 if (bits != 16) {
2596 for (int i = 0; i < 2; i++) {
2597 bool alpha = hi && i == 1;
2598 args[i] = ac_build_imin(ctx, args[i],
2599 alpha ? max_alpha : max_rgb);
2600 args[i] = ac_build_imax(ctx, args[i],
2601 alpha ? min_alpha : min_rgb);
2602 }
2603 }
2604
2605 LLVMValueRef res =
2606 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16",
2607 ctx->v2i16, args, 2,
2608 AC_FUNC_ATTR_READNONE);
2609 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2610 }
2611
2612 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2613 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
2614 LLVMValueRef args[2], unsigned bits, bool hi)
2615 {
2616 assert(bits == 8 || bits == 10 || bits == 16);
2617
2618 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2619 bits == 8 ? 255 : bits == 10 ? 1023 : 65535, 0);
2620 LLVMValueRef max_alpha =
2621 bits != 10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
2622
2623 /* Clamp. */
2624 if (bits != 16) {
2625 for (int i = 0; i < 2; i++) {
2626 bool alpha = hi && i == 1;
2627 args[i] = ac_build_umin(ctx, args[i],
2628 alpha ? max_alpha : max_rgb);
2629 }
2630 }
2631
2632 LLVMValueRef res =
2633 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16",
2634 ctx->v2i16, args, 2,
2635 AC_FUNC_ATTR_READNONE);
2636 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2637 }
2638
2639 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1)
2640 {
2641 return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1,
2642 &i1, 1, AC_FUNC_ATTR_READNONE);
2643 }
2644
2645 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1)
2646 {
2647 ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt,
2648 &i1, 1, 0);
2649 }
2650
2651 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
2652 LLVMValueRef offset, LLVMValueRef width,
2653 bool is_signed)
2654 {
2655 LLVMValueRef args[] = {
2656 input,
2657 offset,
2658 width,
2659 };
2660
2661 return ac_build_intrinsic(ctx, is_signed ? "llvm.amdgcn.sbfe.i32" :
2662 "llvm.amdgcn.ubfe.i32",
2663 ctx->i32, args, 3, AC_FUNC_ATTR_READNONE);
2664
2665 }
2666
2667 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2668 LLVMValueRef s1, LLVMValueRef s2)
2669 {
2670 return LLVMBuildAdd(ctx->builder,
2671 LLVMBuildMul(ctx->builder, s0, s1, ""), s2, "");
2672 }
2673
2674 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2675 LLVMValueRef s1, LLVMValueRef s2)
2676 {
2677 /* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
2678 if (ctx->chip_class >= GFX10) {
2679 return ac_build_intrinsic(ctx, "llvm.fma.f32", ctx->f32,
2680 (LLVMValueRef []) {s0, s1, s2}, 3,
2681 AC_FUNC_ATTR_READNONE);
2682 }
2683
2684 return LLVMBuildFAdd(ctx->builder,
2685 LLVMBuildFMul(ctx->builder, s0, s1, ""), s2, "");
2686 }
2687
2688 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags)
2689 {
2690 if (!wait_flags)
2691 return;
2692
2693 unsigned lgkmcnt = 63;
2694 unsigned vmcnt = ctx->chip_class >= GFX9 ? 63 : 15;
2695 unsigned vscnt = 63;
2696
2697 if (wait_flags & AC_WAIT_LGKM)
2698 lgkmcnt = 0;
2699 if (wait_flags & AC_WAIT_VLOAD)
2700 vmcnt = 0;
2701
2702 if (wait_flags & AC_WAIT_VSTORE) {
2703 if (ctx->chip_class >= GFX10)
2704 vscnt = 0;
2705 else
2706 vmcnt = 0;
2707 }
2708
2709 /* There is no intrinsic for vscnt(0), so use a fence. */
2710 if ((wait_flags & AC_WAIT_LGKM &&
2711 wait_flags & AC_WAIT_VLOAD &&
2712 wait_flags & AC_WAIT_VSTORE) ||
2713 vscnt == 0) {
2714 LLVMBuildFence(ctx->builder, LLVMAtomicOrderingRelease, false, "");
2715 return;
2716 }
2717
2718 unsigned simm16 = (lgkmcnt << 8) |
2719 (7 << 4) | /* expcnt */
2720 (vmcnt & 0xf) |
2721 ((vmcnt >> 4) << 14);
2722
2723 LLVMValueRef args[1] = {
2724 LLVMConstInt(ctx->i32, simm16, false),
2725 };
2726 ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt",
2727 ctx->voidt, args, 1, 0);
2728 }
2729
2730 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
2731 unsigned bitsize)
2732 {
2733 LLVMTypeRef type;
2734 char *intr;
2735
2736 if (bitsize == 16) {
2737 intr = "llvm.amdgcn.fract.f16";
2738 type = ctx->f16;
2739 } else if (bitsize == 32) {
2740 intr = "llvm.amdgcn.fract.f32";
2741 type = ctx->f32;
2742 } else {
2743 intr = "llvm.amdgcn.fract.f64";
2744 type = ctx->f64;
2745 }
2746
2747 LLVMValueRef params[] = {
2748 src0,
2749 };
2750 return ac_build_intrinsic(ctx, intr, type, params, 1,
2751 AC_FUNC_ATTR_READNONE);
2752 }
2753
2754 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2755 unsigned bitsize)
2756 {
2757 LLVMTypeRef type = LLVMIntTypeInContext(ctx->context, bitsize);
2758 LLVMValueRef zero = LLVMConstInt(type, 0, false);
2759 LLVMValueRef one = LLVMConstInt(type, 1, false);
2760
2761 LLVMValueRef cmp, val;
2762 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, src0, zero, "");
2763 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2764 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGE, val, zero, "");
2765 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstInt(type, -1, true), "");
2766 return val;
2767 }
2768
2769 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2770 unsigned bitsize)
2771 {
2772 LLVMValueRef cmp, val, zero, one;
2773 LLVMTypeRef type;
2774
2775 if (bitsize == 16) {
2776 type = ctx->f16;
2777 zero = ctx->f16_0;
2778 one = ctx->f16_1;
2779 } else if (bitsize == 32) {
2780 type = ctx->f32;
2781 zero = ctx->f32_0;
2782 one = ctx->f32_1;
2783 } else {
2784 type = ctx->f64;
2785 zero = ctx->f64_0;
2786 one = ctx->f64_1;
2787 }
2788
2789 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src0, zero, "");
2790 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2791 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGE, val, zero, "");
2792 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstReal(type, -1.0), "");
2793 return val;
2794 }
2795
2796 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0)
2797 {
2798 LLVMValueRef result;
2799 unsigned bitsize;
2800
2801 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2802
2803 switch (bitsize) {
2804 case 128:
2805 result = ac_build_intrinsic(ctx, "llvm.ctpop.i128", ctx->i128,
2806 (LLVMValueRef []) { src0 }, 1,
2807 AC_FUNC_ATTR_READNONE);
2808 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
2809 break;
2810 case 64:
2811 result = ac_build_intrinsic(ctx, "llvm.ctpop.i64", ctx->i64,
2812 (LLVMValueRef []) { src0 }, 1,
2813 AC_FUNC_ATTR_READNONE);
2814
2815 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
2816 break;
2817 case 32:
2818 result = ac_build_intrinsic(ctx, "llvm.ctpop.i32", ctx->i32,
2819 (LLVMValueRef []) { src0 }, 1,
2820 AC_FUNC_ATTR_READNONE);
2821 break;
2822 case 16:
2823 result = ac_build_intrinsic(ctx, "llvm.ctpop.i16", ctx->i16,
2824 (LLVMValueRef []) { src0 }, 1,
2825 AC_FUNC_ATTR_READNONE);
2826
2827 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2828 break;
2829 case 8:
2830 result = ac_build_intrinsic(ctx, "llvm.ctpop.i8", ctx->i8,
2831 (LLVMValueRef []) { src0 }, 1,
2832 AC_FUNC_ATTR_READNONE);
2833
2834 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2835 break;
2836 default:
2837 unreachable(!"invalid bitsize");
2838 break;
2839 }
2840
2841 return result;
2842 }
2843
2844 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
2845 LLVMValueRef src0)
2846 {
2847 LLVMValueRef result;
2848 unsigned bitsize;
2849
2850 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2851
2852 switch (bitsize) {
2853 case 64:
2854 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i64", ctx->i64,
2855 (LLVMValueRef []) { src0 }, 1,
2856 AC_FUNC_ATTR_READNONE);
2857
2858 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
2859 break;
2860 case 32:
2861 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i32", ctx->i32,
2862 (LLVMValueRef []) { src0 }, 1,
2863 AC_FUNC_ATTR_READNONE);
2864 break;
2865 case 16:
2866 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i16", ctx->i16,
2867 (LLVMValueRef []) { src0 }, 1,
2868 AC_FUNC_ATTR_READNONE);
2869
2870 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2871 break;
2872 case 8:
2873 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i8", ctx->i8,
2874 (LLVMValueRef []) { src0 }, 1,
2875 AC_FUNC_ATTR_READNONE);
2876
2877 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2878 break;
2879 default:
2880 unreachable(!"invalid bitsize");
2881 break;
2882 }
2883
2884 return result;
2885 }
2886
2887 #define AC_EXP_TARGET 0
2888 #define AC_EXP_ENABLED_CHANNELS 1
2889 #define AC_EXP_OUT0 2
2890
2891 enum ac_ir_type {
2892 AC_IR_UNDEF,
2893 AC_IR_CONST,
2894 AC_IR_VALUE,
2895 };
2896
2897 struct ac_vs_exp_chan
2898 {
2899 LLVMValueRef value;
2900 float const_float;
2901 enum ac_ir_type type;
2902 };
2903
2904 struct ac_vs_exp_inst {
2905 unsigned offset;
2906 LLVMValueRef inst;
2907 struct ac_vs_exp_chan chan[4];
2908 };
2909
2910 struct ac_vs_exports {
2911 unsigned num;
2912 struct ac_vs_exp_inst exp[VARYING_SLOT_MAX];
2913 };
2914
2915 /* Return true if the PARAM export has been eliminated. */
2916 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset,
2917 uint32_t num_outputs,
2918 struct ac_vs_exp_inst *exp)
2919 {
2920 unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
2921 bool is_zero[4] = {}, is_one[4] = {};
2922
2923 for (i = 0; i < 4; i++) {
2924 /* It's a constant expression. Undef outputs are eliminated too. */
2925 if (exp->chan[i].type == AC_IR_UNDEF) {
2926 is_zero[i] = true;
2927 is_one[i] = true;
2928 } else if (exp->chan[i].type == AC_IR_CONST) {
2929 if (exp->chan[i].const_float == 0)
2930 is_zero[i] = true;
2931 else if (exp->chan[i].const_float == 1)
2932 is_one[i] = true;
2933 else
2934 return false; /* other constant */
2935 } else
2936 return false;
2937 }
2938
2939 /* Only certain combinations of 0 and 1 can be eliminated. */
2940 if (is_zero[0] && is_zero[1] && is_zero[2])
2941 default_val = is_zero[3] ? 0 : 1;
2942 else if (is_one[0] && is_one[1] && is_one[2])
2943 default_val = is_zero[3] ? 2 : 3;
2944 else
2945 return false;
2946
2947 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
2948 LLVMInstructionEraseFromParent(exp->inst);
2949
2950 /* Change OFFSET to DEFAULT_VAL. */
2951 for (i = 0; i < num_outputs; i++) {
2952 if (vs_output_param_offset[i] == exp->offset) {
2953 vs_output_param_offset[i] =
2954 AC_EXP_PARAM_DEFAULT_VAL_0000 + default_val;
2955 break;
2956 }
2957 }
2958 return true;
2959 }
2960
2961 static bool ac_eliminate_duplicated_output(struct ac_llvm_context *ctx,
2962 uint8_t *vs_output_param_offset,
2963 uint32_t num_outputs,
2964 struct ac_vs_exports *processed,
2965 struct ac_vs_exp_inst *exp)
2966 {
2967 unsigned p, copy_back_channels = 0;
2968
2969 /* See if the output is already in the list of processed outputs.
2970 * The LLVMValueRef comparison relies on SSA.
2971 */
2972 for (p = 0; p < processed->num; p++) {
2973 bool different = false;
2974
2975 for (unsigned j = 0; j < 4; j++) {
2976 struct ac_vs_exp_chan *c1 = &processed->exp[p].chan[j];
2977 struct ac_vs_exp_chan *c2 = &exp->chan[j];
2978
2979 /* Treat undef as a match. */
2980 if (c2->type == AC_IR_UNDEF)
2981 continue;
2982
2983 /* If c1 is undef but c2 isn't, we can copy c2 to c1
2984 * and consider the instruction duplicated.
2985 */
2986 if (c1->type == AC_IR_UNDEF) {
2987 copy_back_channels |= 1 << j;
2988 continue;
2989 }
2990
2991 /* Test whether the channels are not equal. */
2992 if (c1->type != c2->type ||
2993 (c1->type == AC_IR_CONST &&
2994 c1->const_float != c2->const_float) ||
2995 (c1->type == AC_IR_VALUE &&
2996 c1->value != c2->value)) {
2997 different = true;
2998 break;
2999 }
3000 }
3001 if (!different)
3002 break;
3003
3004 copy_back_channels = 0;
3005 }
3006 if (p == processed->num)
3007 return false;
3008
3009 /* If a match was found, but the matching export has undef where the new
3010 * one has a normal value, copy the normal value to the undef channel.
3011 */
3012 struct ac_vs_exp_inst *match = &processed->exp[p];
3013
3014 /* Get current enabled channels mask. */
3015 LLVMValueRef arg = LLVMGetOperand(match->inst, AC_EXP_ENABLED_CHANNELS);
3016 unsigned enabled_channels = LLVMConstIntGetZExtValue(arg);
3017
3018 while (copy_back_channels) {
3019 unsigned chan = u_bit_scan(&copy_back_channels);
3020
3021 assert(match->chan[chan].type == AC_IR_UNDEF);
3022 LLVMSetOperand(match->inst, AC_EXP_OUT0 + chan,
3023 exp->chan[chan].value);
3024 match->chan[chan] = exp->chan[chan];
3025
3026 /* Update number of enabled channels because the original mask
3027 * is not always 0xf.
3028 */
3029 enabled_channels |= (1 << chan);
3030 LLVMSetOperand(match->inst, AC_EXP_ENABLED_CHANNELS,
3031 LLVMConstInt(ctx->i32, enabled_channels, 0));
3032 }
3033
3034 /* The PARAM export is duplicated. Kill it. */
3035 LLVMInstructionEraseFromParent(exp->inst);
3036
3037 /* Change OFFSET to the matching export. */
3038 for (unsigned i = 0; i < num_outputs; i++) {
3039 if (vs_output_param_offset[i] == exp->offset) {
3040 vs_output_param_offset[i] = match->offset;
3041 break;
3042 }
3043 }
3044 return true;
3045 }
3046
3047 void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
3048 LLVMValueRef main_fn,
3049 uint8_t *vs_output_param_offset,
3050 uint32_t num_outputs,
3051 uint32_t skip_output_mask,
3052 uint8_t *num_param_exports)
3053 {
3054 LLVMBasicBlockRef bb;
3055 bool removed_any = false;
3056 struct ac_vs_exports exports;
3057
3058 exports.num = 0;
3059
3060 /* Process all LLVM instructions. */
3061 bb = LLVMGetFirstBasicBlock(main_fn);
3062 while (bb) {
3063 LLVMValueRef inst = LLVMGetFirstInstruction(bb);
3064
3065 while (inst) {
3066 LLVMValueRef cur = inst;
3067 inst = LLVMGetNextInstruction(inst);
3068 struct ac_vs_exp_inst exp;
3069
3070 if (LLVMGetInstructionOpcode(cur) != LLVMCall)
3071 continue;
3072
3073 LLVMValueRef callee = ac_llvm_get_called_value(cur);
3074
3075 if (!ac_llvm_is_function(callee))
3076 continue;
3077
3078 const char *name = LLVMGetValueName(callee);
3079 unsigned num_args = LLVMCountParams(callee);
3080
3081 /* Check if this is an export instruction. */
3082 if ((num_args != 9 && num_args != 8) ||
3083 (strcmp(name, "llvm.SI.export") &&
3084 strcmp(name, "llvm.amdgcn.exp.f32")))
3085 continue;
3086
3087 LLVMValueRef arg = LLVMGetOperand(cur, AC_EXP_TARGET);
3088 unsigned target = LLVMConstIntGetZExtValue(arg);
3089
3090 if (target < V_008DFC_SQ_EXP_PARAM)
3091 continue;
3092
3093 target -= V_008DFC_SQ_EXP_PARAM;
3094
3095 /* Parse the instruction. */
3096 memset(&exp, 0, sizeof(exp));
3097 exp.offset = target;
3098 exp.inst = cur;
3099
3100 for (unsigned i = 0; i < 4; i++) {
3101 LLVMValueRef v = LLVMGetOperand(cur, AC_EXP_OUT0 + i);
3102
3103 exp.chan[i].value = v;
3104
3105 if (LLVMIsUndef(v)) {
3106 exp.chan[i].type = AC_IR_UNDEF;
3107 } else if (LLVMIsAConstantFP(v)) {
3108 LLVMBool loses_info;
3109 exp.chan[i].type = AC_IR_CONST;
3110 exp.chan[i].const_float =
3111 LLVMConstRealGetDouble(v, &loses_info);
3112 } else {
3113 exp.chan[i].type = AC_IR_VALUE;
3114 }
3115 }
3116
3117 /* Eliminate constant and duplicated PARAM exports. */
3118 if (!((1u << target) & skip_output_mask) &&
3119 (ac_eliminate_const_output(vs_output_param_offset,
3120 num_outputs, &exp) ||
3121 ac_eliminate_duplicated_output(ctx,
3122 vs_output_param_offset,
3123 num_outputs, &exports,
3124 &exp))) {
3125 removed_any = true;
3126 } else {
3127 exports.exp[exports.num++] = exp;
3128 }
3129 }
3130 bb = LLVMGetNextBasicBlock(bb);
3131 }
3132
3133 /* Remove holes in export memory due to removed PARAM exports.
3134 * This is done by renumbering all PARAM exports.
3135 */
3136 if (removed_any) {
3137 uint8_t old_offset[VARYING_SLOT_MAX];
3138 unsigned out, i;
3139
3140 /* Make a copy of the offsets. We need the old version while
3141 * we are modifying some of them. */
3142 memcpy(old_offset, vs_output_param_offset,
3143 sizeof(old_offset));
3144
3145 for (i = 0; i < exports.num; i++) {
3146 unsigned offset = exports.exp[i].offset;
3147
3148 /* Update vs_output_param_offset. Multiple outputs can
3149 * have the same offset.
3150 */
3151 for (out = 0; out < num_outputs; out++) {
3152 if (old_offset[out] == offset)
3153 vs_output_param_offset[out] = i;
3154 }
3155
3156 /* Change the PARAM offset in the instruction. */
3157 LLVMSetOperand(exports.exp[i].inst, AC_EXP_TARGET,
3158 LLVMConstInt(ctx->i32,
3159 V_008DFC_SQ_EXP_PARAM + i, 0));
3160 }
3161 *num_param_exports = exports.num;
3162 }
3163 }
3164
3165 void ac_init_exec_full_mask(struct ac_llvm_context *ctx)
3166 {
3167 LLVMValueRef full_mask = LLVMConstInt(ctx->i64, ~0ull, 0);
3168 ac_build_intrinsic(ctx,
3169 "llvm.amdgcn.init.exec", ctx->voidt,
3170 &full_mask, 1, AC_FUNC_ATTR_CONVERGENT);
3171 }
3172
3173 void ac_declare_lds_as_pointer(struct ac_llvm_context *ctx)
3174 {
3175 unsigned lds_size = ctx->chip_class >= GFX7 ? 65536 : 32768;
3176 ctx->lds = LLVMBuildIntToPtr(ctx->builder, ctx->i32_0,
3177 LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), AC_ADDR_SPACE_LDS),
3178 "lds");
3179 }
3180
3181 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
3182 LLVMValueRef dw_addr)
3183 {
3184 return LLVMBuildLoad(ctx->builder, ac_build_gep0(ctx, ctx->lds, dw_addr), "");
3185 }
3186
3187 void ac_lds_store(struct ac_llvm_context *ctx,
3188 LLVMValueRef dw_addr,
3189 LLVMValueRef value)
3190 {
3191 value = ac_to_integer(ctx, value);
3192 ac_build_indexed_store(ctx, ctx->lds,
3193 dw_addr, value);
3194 }
3195
3196 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
3197 LLVMTypeRef dst_type,
3198 LLVMValueRef src0)
3199 {
3200 unsigned src0_bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
3201 const char *intrin_name;
3202 LLVMTypeRef type;
3203 LLVMValueRef zero;
3204
3205 switch (src0_bitsize) {
3206 case 64:
3207 intrin_name = "llvm.cttz.i64";
3208 type = ctx->i64;
3209 zero = ctx->i64_0;
3210 break;
3211 case 32:
3212 intrin_name = "llvm.cttz.i32";
3213 type = ctx->i32;
3214 zero = ctx->i32_0;
3215 break;
3216 case 16:
3217 intrin_name = "llvm.cttz.i16";
3218 type = ctx->i16;
3219 zero = ctx->i16_0;
3220 break;
3221 case 8:
3222 intrin_name = "llvm.cttz.i8";
3223 type = ctx->i8;
3224 zero = ctx->i8_0;
3225 break;
3226 default:
3227 unreachable(!"invalid bitsize");
3228 }
3229
3230 LLVMValueRef params[2] = {
3231 src0,
3232
3233 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
3234 * add special code to check for x=0. The reason is that
3235 * the LLVM behavior for x=0 is different from what we
3236 * need here. However, LLVM also assumes that ffs(x) is
3237 * in [0, 31], but GLSL expects that ffs(0) = -1, so
3238 * a conditional assignment to handle 0 is still required.
3239 *
3240 * The hardware already implements the correct behavior.
3241 */
3242 ctx->i1true,
3243 };
3244
3245 LLVMValueRef lsb = ac_build_intrinsic(ctx, intrin_name, type,
3246 params, 2,
3247 AC_FUNC_ATTR_READNONE);
3248
3249 if (src0_bitsize == 64) {
3250 lsb = LLVMBuildTrunc(ctx->builder, lsb, ctx->i32, "");
3251 } else if (src0_bitsize < 32) {
3252 lsb = LLVMBuildSExt(ctx->builder, lsb, ctx->i32, "");
3253 }
3254
3255 /* TODO: We need an intrinsic to skip this conditional. */
3256 /* Check for zero: */
3257 return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder,
3258 LLVMIntEQ, src0,
3259 zero, ""),
3260 LLVMConstInt(ctx->i32, -1, 0), lsb, "");
3261 }
3262
3263 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type)
3264 {
3265 return LLVMPointerType(elem_type, AC_ADDR_SPACE_CONST);
3266 }
3267
3268 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type)
3269 {
3270 return LLVMPointerType(elem_type, AC_ADDR_SPACE_CONST_32BIT);
3271 }
3272
3273 static struct ac_llvm_flow *
3274 get_current_flow(struct ac_llvm_context *ctx)
3275 {
3276 if (ctx->flow->depth > 0)
3277 return &ctx->flow->stack[ctx->flow->depth - 1];
3278 return NULL;
3279 }
3280
3281 static struct ac_llvm_flow *
3282 get_innermost_loop(struct ac_llvm_context *ctx)
3283 {
3284 for (unsigned i = ctx->flow->depth; i > 0; --i) {
3285 if (ctx->flow->stack[i - 1].loop_entry_block)
3286 return &ctx->flow->stack[i - 1];
3287 }
3288 return NULL;
3289 }
3290
3291 static struct ac_llvm_flow *
3292 push_flow(struct ac_llvm_context *ctx)
3293 {
3294 struct ac_llvm_flow *flow;
3295
3296 if (ctx->flow->depth >= ctx->flow->depth_max) {
3297 unsigned new_max = MAX2(ctx->flow->depth << 1,
3298 AC_LLVM_INITIAL_CF_DEPTH);
3299
3300 ctx->flow->stack = realloc(ctx->flow->stack, new_max * sizeof(*ctx->flow->stack));
3301 ctx->flow->depth_max = new_max;
3302 }
3303
3304 flow = &ctx->flow->stack[ctx->flow->depth];
3305 ctx->flow->depth++;
3306
3307 flow->next_block = NULL;
3308 flow->loop_entry_block = NULL;
3309 return flow;
3310 }
3311
3312 static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base,
3313 int label_id)
3314 {
3315 char buf[32];
3316 snprintf(buf, sizeof(buf), "%s%d", base, label_id);
3317 LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
3318 }
3319
3320 /* Append a basic block at the level of the parent flow.
3321 */
3322 static LLVMBasicBlockRef append_basic_block(struct ac_llvm_context *ctx,
3323 const char *name)
3324 {
3325 assert(ctx->flow->depth >= 1);
3326
3327 if (ctx->flow->depth >= 2) {
3328 struct ac_llvm_flow *flow = &ctx->flow->stack[ctx->flow->depth - 2];
3329
3330 return LLVMInsertBasicBlockInContext(ctx->context,
3331 flow->next_block, name);
3332 }
3333
3334 LLVMValueRef main_fn =
3335 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->builder));
3336 return LLVMAppendBasicBlockInContext(ctx->context, main_fn, name);
3337 }
3338
3339 /* Emit a branch to the given default target for the current block if
3340 * applicable -- that is, if the current block does not already contain a
3341 * branch from a break or continue.
3342 */
3343 static void emit_default_branch(LLVMBuilderRef builder,
3344 LLVMBasicBlockRef target)
3345 {
3346 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
3347 LLVMBuildBr(builder, target);
3348 }
3349
3350 void ac_build_bgnloop(struct ac_llvm_context *ctx, int label_id)
3351 {
3352 struct ac_llvm_flow *flow = push_flow(ctx);
3353 flow->loop_entry_block = append_basic_block(ctx, "LOOP");
3354 flow->next_block = append_basic_block(ctx, "ENDLOOP");
3355 set_basicblock_name(flow->loop_entry_block, "loop", label_id);
3356 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
3357 LLVMPositionBuilderAtEnd(ctx->builder, flow->loop_entry_block);
3358 }
3359
3360 void ac_build_break(struct ac_llvm_context *ctx)
3361 {
3362 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
3363 LLVMBuildBr(ctx->builder, flow->next_block);
3364 }
3365
3366 void ac_build_continue(struct ac_llvm_context *ctx)
3367 {
3368 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
3369 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
3370 }
3371
3372 void ac_build_else(struct ac_llvm_context *ctx, int label_id)
3373 {
3374 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
3375 LLVMBasicBlockRef endif_block;
3376
3377 assert(!current_branch->loop_entry_block);
3378
3379 endif_block = append_basic_block(ctx, "ENDIF");
3380 emit_default_branch(ctx->builder, endif_block);
3381
3382 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
3383 set_basicblock_name(current_branch->next_block, "else", label_id);
3384
3385 current_branch->next_block = endif_block;
3386 }
3387
3388 void ac_build_endif(struct ac_llvm_context *ctx, int label_id)
3389 {
3390 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
3391
3392 assert(!current_branch->loop_entry_block);
3393
3394 emit_default_branch(ctx->builder, current_branch->next_block);
3395 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
3396 set_basicblock_name(current_branch->next_block, "endif", label_id);
3397
3398 ctx->flow->depth--;
3399 }
3400
3401 void ac_build_endloop(struct ac_llvm_context *ctx, int label_id)
3402 {
3403 struct ac_llvm_flow *current_loop = get_current_flow(ctx);
3404
3405 assert(current_loop->loop_entry_block);
3406
3407 emit_default_branch(ctx->builder, current_loop->loop_entry_block);
3408
3409 LLVMPositionBuilderAtEnd(ctx->builder, current_loop->next_block);
3410 set_basicblock_name(current_loop->next_block, "endloop", label_id);
3411 ctx->flow->depth--;
3412 }
3413
3414 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id)
3415 {
3416 struct ac_llvm_flow *flow = push_flow(ctx);
3417 LLVMBasicBlockRef if_block;
3418
3419 if_block = append_basic_block(ctx, "IF");
3420 flow->next_block = append_basic_block(ctx, "ELSE");
3421 set_basicblock_name(if_block, "if", label_id);
3422 LLVMBuildCondBr(ctx->builder, cond, if_block, flow->next_block);
3423 LLVMPositionBuilderAtEnd(ctx->builder, if_block);
3424 }
3425
3426 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
3427 int label_id)
3428 {
3429 LLVMValueRef cond = LLVMBuildFCmp(ctx->builder, LLVMRealUNE,
3430 value, ctx->f32_0, "");
3431 ac_build_ifcc(ctx, cond, label_id);
3432 }
3433
3434 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
3435 int label_id)
3436 {
3437 LLVMValueRef cond = LLVMBuildICmp(ctx->builder, LLVMIntNE,
3438 ac_to_integer(ctx, value),
3439 ctx->i32_0, "");
3440 ac_build_ifcc(ctx, cond, label_id);
3441 }
3442
3443 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
3444 const char *name)
3445 {
3446 LLVMBuilderRef builder = ac->builder;
3447 LLVMBasicBlockRef current_block = LLVMGetInsertBlock(builder);
3448 LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
3449 LLVMBasicBlockRef first_block = LLVMGetEntryBasicBlock(function);
3450 LLVMValueRef first_instr = LLVMGetFirstInstruction(first_block);
3451 LLVMBuilderRef first_builder = LLVMCreateBuilderInContext(ac->context);
3452 LLVMValueRef res;
3453
3454 if (first_instr) {
3455 LLVMPositionBuilderBefore(first_builder, first_instr);
3456 } else {
3457 LLVMPositionBuilderAtEnd(first_builder, first_block);
3458 }
3459
3460 res = LLVMBuildAlloca(first_builder, type, name);
3461 LLVMDisposeBuilder(first_builder);
3462 return res;
3463 }
3464
3465 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac,
3466 LLVMTypeRef type, const char *name)
3467 {
3468 LLVMValueRef ptr = ac_build_alloca_undef(ac, type, name);
3469 LLVMBuildStore(ac->builder, LLVMConstNull(type), ptr);
3470 return ptr;
3471 }
3472
3473 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
3474 LLVMTypeRef type)
3475 {
3476 int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
3477 return LLVMBuildBitCast(ctx->builder, ptr,
3478 LLVMPointerType(type, addr_space), "");
3479 }
3480
3481 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
3482 unsigned count)
3483 {
3484 unsigned num_components = ac_get_llvm_num_components(value);
3485 if (count == num_components)
3486 return value;
3487
3488 LLVMValueRef masks[MAX2(count, 2)];
3489 masks[0] = ctx->i32_0;
3490 masks[1] = ctx->i32_1;
3491 for (unsigned i = 2; i < count; i++)
3492 masks[i] = LLVMConstInt(ctx->i32, i, false);
3493
3494 if (count == 1)
3495 return LLVMBuildExtractElement(ctx->builder, value, masks[0],
3496 "");
3497
3498 LLVMValueRef swizzle = LLVMConstVector(masks, count);
3499 return LLVMBuildShuffleVector(ctx->builder, value, value, swizzle, "");
3500 }
3501
3502 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
3503 unsigned rshift, unsigned bitwidth)
3504 {
3505 LLVMValueRef value = param;
3506 if (rshift)
3507 value = LLVMBuildLShr(ctx->builder, value,
3508 LLVMConstInt(ctx->i32, rshift, false), "");
3509
3510 if (rshift + bitwidth < 32) {
3511 unsigned mask = (1 << bitwidth) - 1;
3512 value = LLVMBuildAnd(ctx->builder, value,
3513 LLVMConstInt(ctx->i32, mask, false), "");
3514 }
3515 return value;
3516 }
3517
3518 /* Adjust the sample index according to FMASK.
3519 *
3520 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3521 * which is the identity mapping. Each nibble says which physical sample
3522 * should be fetched to get that sample.
3523 *
3524 * For example, 0x11111100 means there are only 2 samples stored and
3525 * the second sample covers 3/4 of the pixel. When reading samples 0
3526 * and 1, return physical sample 0 (determined by the first two 0s
3527 * in FMASK), otherwise return physical sample 1.
3528 *
3529 * The sample index should be adjusted as follows:
3530 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
3531 */
3532 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
3533 LLVMValueRef *addr, bool is_array_tex)
3534 {
3535 struct ac_image_args fmask_load = {};
3536 fmask_load.opcode = ac_image_load;
3537 fmask_load.resource = fmask;
3538 fmask_load.dmask = 0xf;
3539 fmask_load.dim = is_array_tex ? ac_image_2darray : ac_image_2d;
3540 fmask_load.attributes = AC_FUNC_ATTR_READNONE;
3541
3542 fmask_load.coords[0] = addr[0];
3543 fmask_load.coords[1] = addr[1];
3544 if (is_array_tex)
3545 fmask_load.coords[2] = addr[2];
3546
3547 LLVMValueRef fmask_value = ac_build_image_opcode(ac, &fmask_load);
3548 fmask_value = LLVMBuildExtractElement(ac->builder, fmask_value,
3549 ac->i32_0, "");
3550
3551 /* Apply the formula. */
3552 unsigned sample_chan = is_array_tex ? 3 : 2;
3553 LLVMValueRef final_sample;
3554 final_sample = LLVMBuildMul(ac->builder, addr[sample_chan],
3555 LLVMConstInt(ac->i32, 4, 0), "");
3556 final_sample = LLVMBuildLShr(ac->builder, fmask_value, final_sample, "");
3557 /* Mask the sample index by 0x7, because 0x8 means an unknown value
3558 * with EQAA, so those will map to 0. */
3559 final_sample = LLVMBuildAnd(ac->builder, final_sample,
3560 LLVMConstInt(ac->i32, 0x7, 0), "");
3561
3562 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3563 * resource descriptor is 0 (invalid).
3564 */
3565 LLVMValueRef tmp;
3566 tmp = LLVMBuildBitCast(ac->builder, fmask, ac->v8i32, "");
3567 tmp = LLVMBuildExtractElement(ac->builder, tmp, ac->i32_1, "");
3568 tmp = LLVMBuildICmp(ac->builder, LLVMIntNE, tmp, ac->i32_0, "");
3569
3570 /* Replace the MSAA sample index. */
3571 addr[sample_chan] = LLVMBuildSelect(ac->builder, tmp, final_sample,
3572 addr[sample_chan], "");
3573 }
3574
3575 static LLVMValueRef
3576 _ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src,
3577 LLVMValueRef lane, bool with_opt_barrier)
3578 {
3579 LLVMTypeRef type = LLVMTypeOf(src);
3580 LLVMValueRef result;
3581
3582 if (with_opt_barrier)
3583 ac_build_optimization_barrier(ctx, &src);
3584
3585 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
3586 if (lane)
3587 lane = LLVMBuildZExt(ctx->builder, lane, ctx->i32, "");
3588
3589 result = ac_build_intrinsic(ctx,
3590 lane == NULL ? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
3591 ctx->i32, (LLVMValueRef []) { src, lane },
3592 lane == NULL ? 1 : 2,
3593 AC_FUNC_ATTR_READNONE |
3594 AC_FUNC_ATTR_CONVERGENT);
3595
3596 return LLVMBuildTrunc(ctx->builder, result, type, "");
3597 }
3598
3599 static LLVMValueRef
3600 ac_build_readlane_common(struct ac_llvm_context *ctx,
3601 LLVMValueRef src, LLVMValueRef lane,
3602 bool with_opt_barrier)
3603 {
3604 LLVMTypeRef src_type = LLVMTypeOf(src);
3605 src = ac_to_integer(ctx, src);
3606 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3607 LLVMValueRef ret;
3608
3609 if (bits > 32) {
3610 assert(bits % 32 == 0);
3611 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3612 LLVMValueRef src_vector =
3613 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3614 ret = LLVMGetUndef(vec_type);
3615 for (unsigned i = 0; i < bits / 32; i++) {
3616 LLVMValueRef ret_comp;
3617
3618 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3619 LLVMConstInt(ctx->i32, i, 0), "");
3620
3621 ret_comp = _ac_build_readlane(ctx, src, lane,
3622 with_opt_barrier);
3623
3624 ret = LLVMBuildInsertElement(ctx->builder, ret, ret_comp,
3625 LLVMConstInt(ctx->i32, i, 0), "");
3626 }
3627 } else {
3628 ret = _ac_build_readlane(ctx, src, lane, with_opt_barrier);
3629 }
3630
3631 if (LLVMGetTypeKind(src_type) == LLVMPointerTypeKind)
3632 return LLVMBuildIntToPtr(ctx->builder, ret, src_type, "");
3633 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3634 }
3635
3636 /**
3637 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
3638 *
3639 * The optimization barrier is not needed if the value is the same in all lanes
3640 * or if this is called in the outermost block.
3641 *
3642 * @param ctx
3643 * @param src
3644 * @param lane - id of the lane or NULL for the first active lane
3645 * @return value of the lane
3646 */
3647 LLVMValueRef ac_build_readlane_no_opt_barrier(struct ac_llvm_context *ctx,
3648 LLVMValueRef src, LLVMValueRef lane)
3649 {
3650 return ac_build_readlane_common(ctx, src, lane, false);
3651 }
3652
3653
3654 LLVMValueRef
3655 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
3656 {
3657 return ac_build_readlane_common(ctx, src, lane, true);
3658 }
3659
3660 LLVMValueRef
3661 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane)
3662 {
3663 return ac_build_intrinsic(ctx, "llvm.amdgcn.writelane", ctx->i32,
3664 (LLVMValueRef []) {value, lane, src}, 3,
3665 AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3666 }
3667
3668 LLVMValueRef
3669 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask)
3670 {
3671 if (ctx->wave_size == 32) {
3672 return ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
3673 (LLVMValueRef []) { mask, ctx->i32_0 },
3674 2, AC_FUNC_ATTR_READNONE);
3675 }
3676 LLVMValueRef mask_vec = LLVMBuildBitCast(ctx->builder, mask, ctx->v2i32, "");
3677 LLVMValueRef mask_lo = LLVMBuildExtractElement(ctx->builder, mask_vec,
3678 ctx->i32_0, "");
3679 LLVMValueRef mask_hi = LLVMBuildExtractElement(ctx->builder, mask_vec,
3680 ctx->i32_1, "");
3681 LLVMValueRef val =
3682 ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
3683 (LLVMValueRef []) { mask_lo, ctx->i32_0 },
3684 2, AC_FUNC_ATTR_READNONE);
3685 val = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi", ctx->i32,
3686 (LLVMValueRef []) { mask_hi, val },
3687 2, AC_FUNC_ATTR_READNONE);
3688 return val;
3689 }
3690
3691 enum dpp_ctrl {
3692 _dpp_quad_perm = 0x000,
3693 _dpp_row_sl = 0x100,
3694 _dpp_row_sr = 0x110,
3695 _dpp_row_rr = 0x120,
3696 dpp_wf_sl1 = 0x130,
3697 dpp_wf_rl1 = 0x134,
3698 dpp_wf_sr1 = 0x138,
3699 dpp_wf_rr1 = 0x13C,
3700 dpp_row_mirror = 0x140,
3701 dpp_row_half_mirror = 0x141,
3702 dpp_row_bcast15 = 0x142,
3703 dpp_row_bcast31 = 0x143
3704 };
3705
3706 static inline enum dpp_ctrl
3707 dpp_quad_perm(unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
3708 {
3709 assert(lane0 < 4 && lane1 < 4 && lane2 < 4 && lane3 < 4);
3710 return _dpp_quad_perm | lane0 | (lane1 << 2) | (lane2 << 4) | (lane3 << 6);
3711 }
3712
3713 static inline enum dpp_ctrl
3714 dpp_row_sl(unsigned amount)
3715 {
3716 assert(amount > 0 && amount < 16);
3717 return _dpp_row_sl | amount;
3718 }
3719
3720 static inline enum dpp_ctrl
3721 dpp_row_sr(unsigned amount)
3722 {
3723 assert(amount > 0 && amount < 16);
3724 return _dpp_row_sr | amount;
3725 }
3726
3727 static LLVMValueRef
3728 _ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3729 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3730 bool bound_ctrl)
3731 {
3732 LLVMTypeRef type = LLVMTypeOf(src);
3733 LLVMValueRef res;
3734
3735 old = LLVMBuildZExt(ctx->builder, old, ctx->i32, "");
3736 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
3737
3738 res = ac_build_intrinsic(ctx, "llvm.amdgcn.update.dpp.i32", ctx->i32,
3739 (LLVMValueRef[]) {
3740 old, src,
3741 LLVMConstInt(ctx->i32, dpp_ctrl, 0),
3742 LLVMConstInt(ctx->i32, row_mask, 0),
3743 LLVMConstInt(ctx->i32, bank_mask, 0),
3744 LLVMConstInt(ctx->i1, bound_ctrl, 0) },
3745 6, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3746
3747 return LLVMBuildTrunc(ctx->builder, res, type, "");
3748 }
3749
3750 static LLVMValueRef
3751 ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3752 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3753 bool bound_ctrl)
3754 {
3755 LLVMTypeRef src_type = LLVMTypeOf(src);
3756 src = ac_to_integer(ctx, src);
3757 old = ac_to_integer(ctx, old);
3758 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3759 LLVMValueRef ret;
3760 if (bits > 32) {
3761 assert(bits % 32 == 0);
3762 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3763 LLVMValueRef src_vector =
3764 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3765 LLVMValueRef old_vector =
3766 LLVMBuildBitCast(ctx->builder, old, vec_type, "");
3767 ret = LLVMGetUndef(vec_type);
3768 for (unsigned i = 0; i < bits / 32; i++) {
3769 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3770 LLVMConstInt(ctx->i32, i,
3771 0), "");
3772 old = LLVMBuildExtractElement(ctx->builder, old_vector,
3773 LLVMConstInt(ctx->i32, i,
3774 0), "");
3775 LLVMValueRef ret_comp = _ac_build_dpp(ctx, old, src,
3776 dpp_ctrl,
3777 row_mask,
3778 bank_mask,
3779 bound_ctrl);
3780 ret = LLVMBuildInsertElement(ctx->builder, ret,
3781 ret_comp,
3782 LLVMConstInt(ctx->i32, i,
3783 0), "");
3784 }
3785 } else {
3786 ret = _ac_build_dpp(ctx, old, src, dpp_ctrl, row_mask,
3787 bank_mask, bound_ctrl);
3788 }
3789 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3790 }
3791
3792 static LLVMValueRef
3793 _ac_build_permlane16(struct ac_llvm_context *ctx, LLVMValueRef src, uint64_t sel,
3794 bool exchange_rows, bool bound_ctrl)
3795 {
3796 LLVMTypeRef type = LLVMTypeOf(src);
3797 LLVMValueRef result;
3798
3799 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
3800
3801 LLVMValueRef args[6] = {
3802 src,
3803 src,
3804 LLVMConstInt(ctx->i32, sel, false),
3805 LLVMConstInt(ctx->i32, sel >> 32, false),
3806 ctx->i1true, /* fi */
3807 bound_ctrl ? ctx->i1true : ctx->i1false,
3808 };
3809
3810 result = ac_build_intrinsic(ctx, exchange_rows ? "llvm.amdgcn.permlanex16"
3811 : "llvm.amdgcn.permlane16",
3812 ctx->i32, args, 6,
3813 AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3814
3815 return LLVMBuildTrunc(ctx->builder, result, type, "");
3816 }
3817
3818 static LLVMValueRef
3819 ac_build_permlane16(struct ac_llvm_context *ctx, LLVMValueRef src, uint64_t sel,
3820 bool exchange_rows, bool bound_ctrl)
3821 {
3822 LLVMTypeRef src_type = LLVMTypeOf(src);
3823 src = ac_to_integer(ctx, src);
3824 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3825 LLVMValueRef ret;
3826 if (bits > 32) {
3827 assert(bits % 32 == 0);
3828 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3829 LLVMValueRef src_vector =
3830 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3831 ret = LLVMGetUndef(vec_type);
3832 for (unsigned i = 0; i < bits / 32; i++) {
3833 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3834 LLVMConstInt(ctx->i32, i,
3835 0), "");
3836 LLVMValueRef ret_comp =
3837 _ac_build_permlane16(ctx, src, sel,
3838 exchange_rows,
3839 bound_ctrl);
3840 ret = LLVMBuildInsertElement(ctx->builder, ret,
3841 ret_comp,
3842 LLVMConstInt(ctx->i32, i,
3843 0), "");
3844 }
3845 } else {
3846 ret = _ac_build_permlane16(ctx, src, sel, exchange_rows,
3847 bound_ctrl);
3848 }
3849 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3850 }
3851
3852 static inline unsigned
3853 ds_pattern_bitmode(unsigned and_mask, unsigned or_mask, unsigned xor_mask)
3854 {
3855 assert(and_mask < 32 && or_mask < 32 && xor_mask < 32);
3856 return and_mask | (or_mask << 5) | (xor_mask << 10);
3857 }
3858
3859 static LLVMValueRef
3860 _ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
3861 {
3862 LLVMTypeRef src_type = LLVMTypeOf(src);
3863 LLVMValueRef ret;
3864
3865 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
3866
3867 ret = ac_build_intrinsic(ctx, "llvm.amdgcn.ds.swizzle", ctx->i32,
3868 (LLVMValueRef []) {
3869 src, LLVMConstInt(ctx->i32, mask, 0) },
3870 2, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3871
3872 return LLVMBuildTrunc(ctx->builder, ret, src_type, "");
3873 }
3874
3875 LLVMValueRef
3876 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
3877 {
3878 LLVMTypeRef src_type = LLVMTypeOf(src);
3879 src = ac_to_integer(ctx, src);
3880 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3881 LLVMValueRef ret;
3882 if (bits > 32) {
3883 assert(bits % 32 == 0);
3884 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3885 LLVMValueRef src_vector =
3886 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3887 ret = LLVMGetUndef(vec_type);
3888 for (unsigned i = 0; i < bits / 32; i++) {
3889 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3890 LLVMConstInt(ctx->i32, i,
3891 0), "");
3892 LLVMValueRef ret_comp = _ac_build_ds_swizzle(ctx, src,
3893 mask);
3894 ret = LLVMBuildInsertElement(ctx->builder, ret,
3895 ret_comp,
3896 LLVMConstInt(ctx->i32, i,
3897 0), "");
3898 }
3899 } else {
3900 ret = _ac_build_ds_swizzle(ctx, src, mask);
3901 }
3902 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3903 }
3904
3905 static LLVMValueRef
3906 ac_build_wwm(struct ac_llvm_context *ctx, LLVMValueRef src)
3907 {
3908 LLVMTypeRef src_type = LLVMTypeOf(src);
3909 unsigned bitsize = ac_get_elem_bits(ctx, src_type);
3910 char name[32], type[8];
3911 LLVMValueRef ret;
3912
3913 src = ac_to_integer(ctx, src);
3914
3915 if (bitsize < 32)
3916 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
3917
3918 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
3919 snprintf(name, sizeof(name), "llvm.amdgcn.wwm.%s", type);
3920 ret = ac_build_intrinsic(ctx, name, LLVMTypeOf(src),
3921 (LLVMValueRef []) { src }, 1,
3922 AC_FUNC_ATTR_READNONE);
3923
3924 if (bitsize < 32)
3925 ret = LLVMBuildTrunc(ctx->builder, ret,
3926 ac_to_integer_type(ctx, src_type), "");
3927
3928 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3929 }
3930
3931 static LLVMValueRef
3932 ac_build_set_inactive(struct ac_llvm_context *ctx, LLVMValueRef src,
3933 LLVMValueRef inactive)
3934 {
3935 char name[33], type[8];
3936 LLVMTypeRef src_type = LLVMTypeOf(src);
3937 unsigned bitsize = ac_get_elem_bits(ctx, src_type);
3938 src = ac_to_integer(ctx, src);
3939 inactive = ac_to_integer(ctx, inactive);
3940
3941 if (bitsize < 32) {
3942 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
3943 inactive = LLVMBuildZExt(ctx->builder, inactive, ctx->i32, "");
3944 }
3945
3946 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
3947 snprintf(name, sizeof(name), "llvm.amdgcn.set.inactive.%s", type);
3948 LLVMValueRef ret =
3949 ac_build_intrinsic(ctx, name,
3950 LLVMTypeOf(src), (LLVMValueRef []) {
3951 src, inactive }, 2,
3952 AC_FUNC_ATTR_READNONE |
3953 AC_FUNC_ATTR_CONVERGENT);
3954 if (bitsize < 32)
3955 ret = LLVMBuildTrunc(ctx->builder, ret, src_type, "");
3956
3957 return ret;
3958 }
3959
3960 static LLVMValueRef
3961 get_reduction_identity(struct ac_llvm_context *ctx, nir_op op, unsigned type_size)
3962 {
3963 if (type_size == 1) {
3964 switch (op) {
3965 case nir_op_iadd: return ctx->i8_0;
3966 case nir_op_imul: return ctx->i8_1;
3967 case nir_op_imin: return LLVMConstInt(ctx->i8, INT8_MAX, 0);
3968 case nir_op_umin: return LLVMConstInt(ctx->i8, UINT8_MAX, 0);
3969 case nir_op_imax: return LLVMConstInt(ctx->i8, INT8_MIN, 0);
3970 case nir_op_umax: return ctx->i8_0;
3971 case nir_op_iand: return LLVMConstInt(ctx->i8, -1, 0);
3972 case nir_op_ior: return ctx->i8_0;
3973 case nir_op_ixor: return ctx->i8_0;
3974 default:
3975 unreachable("bad reduction intrinsic");
3976 }
3977 } else if (type_size == 2) {
3978 switch (op) {
3979 case nir_op_iadd: return ctx->i16_0;
3980 case nir_op_fadd: return ctx->f16_0;
3981 case nir_op_imul: return ctx->i16_1;
3982 case nir_op_fmul: return ctx->f16_1;
3983 case nir_op_imin: return LLVMConstInt(ctx->i16, INT16_MAX, 0);
3984 case nir_op_umin: return LLVMConstInt(ctx->i16, UINT16_MAX, 0);
3985 case nir_op_fmin: return LLVMConstReal(ctx->f16, INFINITY);
3986 case nir_op_imax: return LLVMConstInt(ctx->i16, INT16_MIN, 0);
3987 case nir_op_umax: return ctx->i16_0;
3988 case nir_op_fmax: return LLVMConstReal(ctx->f16, -INFINITY);
3989 case nir_op_iand: return LLVMConstInt(ctx->i16, -1, 0);
3990 case nir_op_ior: return ctx->i16_0;
3991 case nir_op_ixor: return ctx->i16_0;
3992 default:
3993 unreachable("bad reduction intrinsic");
3994 }
3995 } else if (type_size == 4) {
3996 switch (op) {
3997 case nir_op_iadd: return ctx->i32_0;
3998 case nir_op_fadd: return ctx->f32_0;
3999 case nir_op_imul: return ctx->i32_1;
4000 case nir_op_fmul: return ctx->f32_1;
4001 case nir_op_imin: return LLVMConstInt(ctx->i32, INT32_MAX, 0);
4002 case nir_op_umin: return LLVMConstInt(ctx->i32, UINT32_MAX, 0);
4003 case nir_op_fmin: return LLVMConstReal(ctx->f32, INFINITY);
4004 case nir_op_imax: return LLVMConstInt(ctx->i32, INT32_MIN, 0);
4005 case nir_op_umax: return ctx->i32_0;
4006 case nir_op_fmax: return LLVMConstReal(ctx->f32, -INFINITY);
4007 case nir_op_iand: return LLVMConstInt(ctx->i32, -1, 0);
4008 case nir_op_ior: return ctx->i32_0;
4009 case nir_op_ixor: return ctx->i32_0;
4010 default:
4011 unreachable("bad reduction intrinsic");
4012 }
4013 } else { /* type_size == 64bit */
4014 switch (op) {
4015 case nir_op_iadd: return ctx->i64_0;
4016 case nir_op_fadd: return ctx->f64_0;
4017 case nir_op_imul: return ctx->i64_1;
4018 case nir_op_fmul: return ctx->f64_1;
4019 case nir_op_imin: return LLVMConstInt(ctx->i64, INT64_MAX, 0);
4020 case nir_op_umin: return LLVMConstInt(ctx->i64, UINT64_MAX, 0);
4021 case nir_op_fmin: return LLVMConstReal(ctx->f64, INFINITY);
4022 case nir_op_imax: return LLVMConstInt(ctx->i64, INT64_MIN, 0);
4023 case nir_op_umax: return ctx->i64_0;
4024 case nir_op_fmax: return LLVMConstReal(ctx->f64, -INFINITY);
4025 case nir_op_iand: return LLVMConstInt(ctx->i64, -1, 0);
4026 case nir_op_ior: return ctx->i64_0;
4027 case nir_op_ixor: return ctx->i64_0;
4028 default:
4029 unreachable("bad reduction intrinsic");
4030 }
4031 }
4032 }
4033
4034 static LLVMValueRef
4035 ac_build_alu_op(struct ac_llvm_context *ctx, LLVMValueRef lhs, LLVMValueRef rhs, nir_op op)
4036 {
4037 bool _64bit = ac_get_type_size(LLVMTypeOf(lhs)) == 8;
4038 bool _32bit = ac_get_type_size(LLVMTypeOf(lhs)) == 4;
4039 switch (op) {
4040 case nir_op_iadd: return LLVMBuildAdd(ctx->builder, lhs, rhs, "");
4041 case nir_op_fadd: return LLVMBuildFAdd(ctx->builder, lhs, rhs, "");
4042 case nir_op_imul: return LLVMBuildMul(ctx->builder, lhs, rhs, "");
4043 case nir_op_fmul: return LLVMBuildFMul(ctx->builder, lhs, rhs, "");
4044 case nir_op_imin: return LLVMBuildSelect(ctx->builder,
4045 LLVMBuildICmp(ctx->builder, LLVMIntSLT, lhs, rhs, ""),
4046 lhs, rhs, "");
4047 case nir_op_umin: return LLVMBuildSelect(ctx->builder,
4048 LLVMBuildICmp(ctx->builder, LLVMIntULT, lhs, rhs, ""),
4049 lhs, rhs, "");
4050 case nir_op_fmin: return ac_build_intrinsic(ctx,
4051 _64bit ? "llvm.minnum.f64" : _32bit ? "llvm.minnum.f32" : "llvm.minnum.f16",
4052 _64bit ? ctx->f64 : _32bit ? ctx->f32 : ctx->f16,
4053 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
4054 case nir_op_imax: return LLVMBuildSelect(ctx->builder,
4055 LLVMBuildICmp(ctx->builder, LLVMIntSGT, lhs, rhs, ""),
4056 lhs, rhs, "");
4057 case nir_op_umax: return LLVMBuildSelect(ctx->builder,
4058 LLVMBuildICmp(ctx->builder, LLVMIntUGT, lhs, rhs, ""),
4059 lhs, rhs, "");
4060 case nir_op_fmax: return ac_build_intrinsic(ctx,
4061 _64bit ? "llvm.maxnum.f64" : _32bit ? "llvm.maxnum.f32" : "llvm.maxnum.f16",
4062 _64bit ? ctx->f64 : _32bit ? ctx->f32 : ctx->f16,
4063 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
4064 case nir_op_iand: return LLVMBuildAnd(ctx->builder, lhs, rhs, "");
4065 case nir_op_ior: return LLVMBuildOr(ctx->builder, lhs, rhs, "");
4066 case nir_op_ixor: return LLVMBuildXor(ctx->builder, lhs, rhs, "");
4067 default:
4068 unreachable("bad reduction intrinsic");
4069 }
4070 }
4071
4072 /**
4073 * \param src The value to shift.
4074 * \param identity The value to use the first lane.
4075 * \param maxprefix specifies that the result only needs to be correct for a
4076 * prefix of this many threads
4077 * \return src, shifted 1 lane up, and identity shifted into lane 0.
4078 */
4079 static LLVMValueRef
4080 ac_wavefront_shift_right_1(struct ac_llvm_context *ctx, LLVMValueRef src,
4081 LLVMValueRef identity, unsigned maxprefix)
4082 {
4083 if (ctx->chip_class >= GFX10) {
4084 /* wavefront shift_right by 1 on GFX10 (emulate dpp_wf_sr1) */
4085 LLVMValueRef active, tmp1, tmp2;
4086 LLVMValueRef tid = ac_get_thread_id(ctx);
4087
4088 tmp1 = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
4089
4090 tmp2 = ac_build_permlane16(ctx, src, (uint64_t)~0, true, false);
4091
4092 if (maxprefix > 32) {
4093 active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid,
4094 LLVMConstInt(ctx->i32, 32, false), "");
4095
4096 tmp2 = LLVMBuildSelect(ctx->builder, active,
4097 ac_build_readlane(ctx, src,
4098 LLVMConstInt(ctx->i32, 31, false)),
4099 tmp2, "");
4100
4101 active = LLVMBuildOr(ctx->builder, active,
4102 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
4103 LLVMBuildAnd(ctx->builder, tid,
4104 LLVMConstInt(ctx->i32, 0x1f, false), ""),
4105 LLVMConstInt(ctx->i32, 0x10, false), ""), "");
4106 return LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
4107 } else if (maxprefix > 16) {
4108 active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid,
4109 LLVMConstInt(ctx->i32, 16, false), "");
4110
4111 return LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
4112 }
4113 } else if (ctx->chip_class >= GFX8) {
4114 return ac_build_dpp(ctx, identity, src, dpp_wf_sr1, 0xf, 0xf, false);
4115 }
4116
4117 /* wavefront shift_right by 1 on SI/CI */
4118 LLVMValueRef active, tmp1, tmp2;
4119 LLVMValueRef tid = ac_get_thread_id(ctx);
4120 tmp1 = ac_build_ds_swizzle(ctx, src, (1 << 15) | dpp_quad_perm(0, 0, 1, 2));
4121 tmp2 = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x18, 0x03, 0x00));
4122 active = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
4123 LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 0x7, 0), ""),
4124 LLVMConstInt(ctx->i32, 0x4, 0), "");
4125 tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
4126 tmp2 = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x10, 0x07, 0x00));
4127 active = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
4128 LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 0xf, 0), ""),
4129 LLVMConstInt(ctx->i32, 0x8, 0), "");
4130 tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
4131 tmp2 = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x00, 0x0f, 0x00));
4132 active = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
4133 LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 0x1f, 0), ""),
4134 LLVMConstInt(ctx->i32, 0x10, 0), "");
4135 tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
4136 tmp2 = ac_build_readlane(ctx, src, LLVMConstInt(ctx->i32, 31, 0));
4137 active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid, LLVMConstInt(ctx->i32, 32, 0), "");
4138 tmp1 = LLVMBuildSelect(ctx->builder, active, tmp2, tmp1, "");
4139 active = LLVMBuildICmp(ctx->builder, LLVMIntEQ, tid, LLVMConstInt(ctx->i32, 0, 0), "");
4140 return LLVMBuildSelect(ctx->builder, active, identity, tmp1, "");
4141 }
4142
4143 /**
4144 * \param maxprefix specifies that the result only needs to be correct for a
4145 * prefix of this many threads
4146 */
4147 static LLVMValueRef
4148 ac_build_scan(struct ac_llvm_context *ctx, nir_op op, LLVMValueRef src, LLVMValueRef identity,
4149 unsigned maxprefix, bool inclusive)
4150 {
4151 LLVMValueRef result, tmp;
4152
4153 if (!inclusive)
4154 src = ac_wavefront_shift_right_1(ctx, src, identity, maxprefix);
4155
4156 result = src;
4157
4158 if (ctx->chip_class <= GFX7) {
4159 assert(maxprefix == 64);
4160 LLVMValueRef tid = ac_get_thread_id(ctx);
4161 LLVMValueRef active;
4162 tmp = ac_build_ds_swizzle(ctx, src, ds_pattern_bitmode(0x1e, 0x00, 0x00));
4163 active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
4164 LLVMBuildAnd(ctx->builder, tid, ctx->i32_1, ""),
4165 ctx->i32_0, "");
4166 tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
4167 result = ac_build_alu_op(ctx, result, tmp, op);
4168 tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1c, 0x01, 0x00));
4169 active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
4170 LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 2, 0), ""),
4171 ctx->i32_0, "");
4172 tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
4173 result = ac_build_alu_op(ctx, result, tmp, op);
4174 tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x18, 0x03, 0x00));
4175 active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
4176 LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 4, 0), ""),
4177 ctx->i32_0, "");
4178 tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
4179 result = ac_build_alu_op(ctx, result, tmp, op);
4180 tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x10, 0x07, 0x00));
4181 active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
4182 LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 8, 0), ""),
4183 ctx->i32_0, "");
4184 tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
4185 result = ac_build_alu_op(ctx, result, tmp, op);
4186 tmp = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x00, 0x0f, 0x00));
4187 active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
4188 LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 16, 0), ""),
4189 ctx->i32_0, "");
4190 tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
4191 result = ac_build_alu_op(ctx, result, tmp, op);
4192 tmp = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, 0));
4193 active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
4194 LLVMBuildAnd(ctx->builder, tid, LLVMConstInt(ctx->i32, 32, 0), ""),
4195 ctx->i32_0, "");
4196 tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
4197 result = ac_build_alu_op(ctx, result, tmp, op);
4198 return result;
4199 }
4200
4201 if (maxprefix <= 1)
4202 return result;
4203 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
4204 result = ac_build_alu_op(ctx, result, tmp, op);
4205 if (maxprefix <= 2)
4206 return result;
4207 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(2), 0xf, 0xf, false);
4208 result = ac_build_alu_op(ctx, result, tmp, op);
4209 if (maxprefix <= 3)
4210 return result;
4211 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(3), 0xf, 0xf, false);
4212 result = ac_build_alu_op(ctx, result, tmp, op);
4213 if (maxprefix <= 4)
4214 return result;
4215 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(4), 0xf, 0xe, false);
4216 result = ac_build_alu_op(ctx, result, tmp, op);
4217 if (maxprefix <= 8)
4218 return result;
4219 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(8), 0xf, 0xc, false);
4220 result = ac_build_alu_op(ctx, result, tmp, op);
4221 if (maxprefix <= 16)
4222 return result;
4223
4224 if (ctx->chip_class >= GFX10) {
4225 LLVMValueRef tid = ac_get_thread_id(ctx);
4226 LLVMValueRef active;
4227
4228 tmp = ac_build_permlane16(ctx, result, ~(uint64_t)0, true, false);
4229
4230 active = LLVMBuildICmp(ctx->builder, LLVMIntNE,
4231 LLVMBuildAnd(ctx->builder, tid,
4232 LLVMConstInt(ctx->i32, 16, false), ""),
4233 ctx->i32_0, "");
4234
4235 tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
4236
4237 result = ac_build_alu_op(ctx, result, tmp, op);
4238
4239 if (maxprefix <= 32)
4240 return result;
4241
4242 tmp = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, false));
4243
4244 active = LLVMBuildICmp(ctx->builder, LLVMIntUGE, tid,
4245 LLVMConstInt(ctx->i32, 32, false), "");
4246
4247 tmp = LLVMBuildSelect(ctx->builder, active, tmp, identity, "");
4248
4249 result = ac_build_alu_op(ctx, result, tmp, op);
4250 return result;
4251 }
4252
4253 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
4254 result = ac_build_alu_op(ctx, result, tmp, op);
4255 if (maxprefix <= 32)
4256 return result;
4257 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
4258 result = ac_build_alu_op(ctx, result, tmp, op);
4259 return result;
4260 }
4261
4262 LLVMValueRef
4263 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
4264 {
4265 LLVMValueRef result;
4266
4267 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
4268 LLVMBuilderRef builder = ctx->builder;
4269 src = LLVMBuildZExt(builder, src, ctx->i32, "");
4270 result = ac_build_ballot(ctx, src);
4271 result = ac_build_mbcnt(ctx, result);
4272 result = LLVMBuildAdd(builder, result, src, "");
4273 return result;
4274 }
4275
4276 ac_build_optimization_barrier(ctx, &src);
4277
4278 LLVMValueRef identity =
4279 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
4280 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
4281 LLVMTypeOf(identity), "");
4282 result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, true);
4283
4284 return ac_build_wwm(ctx, result);
4285 }
4286
4287 LLVMValueRef
4288 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
4289 {
4290 LLVMValueRef result;
4291
4292 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
4293 LLVMBuilderRef builder = ctx->builder;
4294 src = LLVMBuildZExt(builder, src, ctx->i32, "");
4295 result = ac_build_ballot(ctx, src);
4296 result = ac_build_mbcnt(ctx, result);
4297 return result;
4298 }
4299
4300 ac_build_optimization_barrier(ctx, &src);
4301
4302 LLVMValueRef identity =
4303 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
4304 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
4305 LLVMTypeOf(identity), "");
4306 result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, false);
4307
4308 return ac_build_wwm(ctx, result);
4309 }
4310
4311 LLVMValueRef
4312 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size)
4313 {
4314 if (cluster_size == 1) return src;
4315 ac_build_optimization_barrier(ctx, &src);
4316 LLVMValueRef result, swap;
4317 LLVMValueRef identity = get_reduction_identity(ctx, op,
4318 ac_get_type_size(LLVMTypeOf(src)));
4319 result = LLVMBuildBitCast(ctx->builder,
4320 ac_build_set_inactive(ctx, src, identity),
4321 LLVMTypeOf(identity), "");
4322 swap = ac_build_quad_swizzle(ctx, result, 1, 0, 3, 2);
4323 result = ac_build_alu_op(ctx, result, swap, op);
4324 if (cluster_size == 2) return ac_build_wwm(ctx, result);
4325
4326 swap = ac_build_quad_swizzle(ctx, result, 2, 3, 0, 1);
4327 result = ac_build_alu_op(ctx, result, swap, op);
4328 if (cluster_size == 4) return ac_build_wwm(ctx, result);
4329
4330 if (ctx->chip_class >= GFX8)
4331 swap = ac_build_dpp(ctx, identity, result, dpp_row_half_mirror, 0xf, 0xf, false);
4332 else
4333 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x04));
4334 result = ac_build_alu_op(ctx, result, swap, op);
4335 if (cluster_size == 8) return ac_build_wwm(ctx, result);
4336
4337 if (ctx->chip_class >= GFX8)
4338 swap = ac_build_dpp(ctx, identity, result, dpp_row_mirror, 0xf, 0xf, false);
4339 else
4340 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x08));
4341 result = ac_build_alu_op(ctx, result, swap, op);
4342 if (cluster_size == 16) return ac_build_wwm(ctx, result);
4343
4344 if (ctx->chip_class >= GFX10)
4345 swap = ac_build_permlane16(ctx, result, 0, true, false);
4346 else if (ctx->chip_class >= GFX8 && cluster_size != 32)
4347 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
4348 else
4349 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x10));
4350 result = ac_build_alu_op(ctx, result, swap, op);
4351 if (cluster_size == 32) return ac_build_wwm(ctx, result);
4352
4353 if (ctx->chip_class >= GFX8) {
4354 if (ctx->wave_size == 64) {
4355 if (ctx->chip_class >= GFX10)
4356 swap = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, false));
4357 else
4358 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
4359 result = ac_build_alu_op(ctx, result, swap, op);
4360 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 63, 0));
4361 }
4362
4363 return ac_build_wwm(ctx, result);
4364 } else {
4365 swap = ac_build_readlane(ctx, result, ctx->i32_0);
4366 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 32, 0));
4367 result = ac_build_alu_op(ctx, result, swap, op);
4368 return ac_build_wwm(ctx, result);
4369 }
4370 }
4371
4372 /**
4373 * "Top half" of a scan that reduces per-wave values across an entire
4374 * workgroup.
4375 *
4376 * The source value must be present in the highest lane of the wave, and the
4377 * highest lane must be live.
4378 */
4379 void
4380 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4381 {
4382 if (ws->maxwaves <= 1)
4383 return;
4384
4385 const LLVMValueRef last_lane = LLVMConstInt(ctx->i32, ctx->wave_size - 1, false);
4386 LLVMBuilderRef builder = ctx->builder;
4387 LLVMValueRef tid = ac_get_thread_id(ctx);
4388 LLVMValueRef tmp;
4389
4390 tmp = LLVMBuildICmp(builder, LLVMIntEQ, tid, last_lane, "");
4391 ac_build_ifcc(ctx, tmp, 1000);
4392 LLVMBuildStore(builder, ws->src, LLVMBuildGEP(builder, ws->scratch, &ws->waveidx, 1, ""));
4393 ac_build_endif(ctx, 1000);
4394 }
4395
4396 /**
4397 * "Bottom half" of a scan that reduces per-wave values across an entire
4398 * workgroup.
4399 *
4400 * The caller must place a barrier between the top and bottom halves.
4401 */
4402 void
4403 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4404 {
4405 const LLVMTypeRef type = LLVMTypeOf(ws->src);
4406 const LLVMValueRef identity =
4407 get_reduction_identity(ctx, ws->op, ac_get_type_size(type));
4408
4409 if (ws->maxwaves <= 1) {
4410 ws->result_reduce = ws->src;
4411 ws->result_inclusive = ws->src;
4412 ws->result_exclusive = identity;
4413 return;
4414 }
4415 assert(ws->maxwaves <= 32);
4416
4417 LLVMBuilderRef builder = ctx->builder;
4418 LLVMValueRef tid = ac_get_thread_id(ctx);
4419 LLVMBasicBlockRef bbs[2];
4420 LLVMValueRef phivalues_scan[2];
4421 LLVMValueRef tmp, tmp2;
4422
4423 bbs[0] = LLVMGetInsertBlock(builder);
4424 phivalues_scan[0] = LLVMGetUndef(type);
4425
4426 if (ws->enable_reduce)
4427 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->numwaves, "");
4428 else if (ws->enable_inclusive)
4429 tmp = LLVMBuildICmp(builder, LLVMIntULE, tid, ws->waveidx, "");
4430 else
4431 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->waveidx, "");
4432 ac_build_ifcc(ctx, tmp, 1001);
4433 {
4434 tmp = LLVMBuildLoad(builder, LLVMBuildGEP(builder, ws->scratch, &tid, 1, ""), "");
4435
4436 ac_build_optimization_barrier(ctx, &tmp);
4437
4438 bbs[1] = LLVMGetInsertBlock(builder);
4439 phivalues_scan[1] = ac_build_scan(ctx, ws->op, tmp, identity, ws->maxwaves, true);
4440 }
4441 ac_build_endif(ctx, 1001);
4442
4443 const LLVMValueRef scan = ac_build_phi(ctx, type, 2, phivalues_scan, bbs);
4444
4445 if (ws->enable_reduce) {
4446 tmp = LLVMBuildSub(builder, ws->numwaves, ctx->i32_1, "");
4447 ws->result_reduce = ac_build_readlane(ctx, scan, tmp);
4448 }
4449 if (ws->enable_inclusive)
4450 ws->result_inclusive = ac_build_readlane(ctx, scan, ws->waveidx);
4451 if (ws->enable_exclusive) {
4452 tmp = LLVMBuildSub(builder, ws->waveidx, ctx->i32_1, "");
4453 tmp = ac_build_readlane(ctx, scan, tmp);
4454 tmp2 = LLVMBuildICmp(builder, LLVMIntEQ, ws->waveidx, ctx->i32_0, "");
4455 ws->result_exclusive = LLVMBuildSelect(builder, tmp2, identity, tmp, "");
4456 }
4457 }
4458
4459 /**
4460 * Inclusive scan of a per-wave value across an entire workgroup.
4461 *
4462 * This implies an s_barrier instruction.
4463 *
4464 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
4465 * of the workgroup are live. (This requirement cannot easily be relaxed in a
4466 * useful manner because of the barrier in the algorithm.)
4467 */
4468 void
4469 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4470 {
4471 ac_build_wg_wavescan_top(ctx, ws);
4472 ac_build_s_barrier(ctx);
4473 ac_build_wg_wavescan_bottom(ctx, ws);
4474 }
4475
4476 /**
4477 * "Top half" of a scan that reduces per-thread values across an entire
4478 * workgroup.
4479 *
4480 * All lanes must be active when this code runs.
4481 */
4482 void
4483 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4484 {
4485 if (ws->enable_exclusive) {
4486 ws->extra = ac_build_exclusive_scan(ctx, ws->src, ws->op);
4487 if (LLVMTypeOf(ws->src) == ctx->i1 && ws->op == nir_op_iadd)
4488 ws->src = LLVMBuildZExt(ctx->builder, ws->src, ctx->i32, "");
4489 ws->src = ac_build_alu_op(ctx, ws->extra, ws->src, ws->op);
4490 } else {
4491 ws->src = ac_build_inclusive_scan(ctx, ws->src, ws->op);
4492 }
4493
4494 bool enable_inclusive = ws->enable_inclusive;
4495 bool enable_exclusive = ws->enable_exclusive;
4496 ws->enable_inclusive = false;
4497 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
4498 ac_build_wg_wavescan_top(ctx, ws);
4499 ws->enable_inclusive = enable_inclusive;
4500 ws->enable_exclusive = enable_exclusive;
4501 }
4502
4503 /**
4504 * "Bottom half" of a scan that reduces per-thread values across an entire
4505 * workgroup.
4506 *
4507 * The caller must place a barrier between the top and bottom halves.
4508 */
4509 void
4510 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4511 {
4512 bool enable_inclusive = ws->enable_inclusive;
4513 bool enable_exclusive = ws->enable_exclusive;
4514 ws->enable_inclusive = false;
4515 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
4516 ac_build_wg_wavescan_bottom(ctx, ws);
4517 ws->enable_inclusive = enable_inclusive;
4518 ws->enable_exclusive = enable_exclusive;
4519
4520 /* ws->result_reduce is already the correct value */
4521 if (ws->enable_inclusive)
4522 ws->result_inclusive = ac_build_alu_op(ctx, ws->result_inclusive, ws->src, ws->op);
4523 if (ws->enable_exclusive)
4524 ws->result_exclusive = ac_build_alu_op(ctx, ws->result_exclusive, ws->extra, ws->op);
4525 }
4526
4527 /**
4528 * A scan that reduces per-thread values across an entire workgroup.
4529 *
4530 * The caller must ensure that all lanes are active when this code runs
4531 * (WWM is insufficient!), because there is an implied barrier.
4532 */
4533 void
4534 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4535 {
4536 ac_build_wg_scan_top(ctx, ws);
4537 ac_build_s_barrier(ctx);
4538 ac_build_wg_scan_bottom(ctx, ws);
4539 }
4540
4541 LLVMValueRef
4542 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
4543 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
4544 {
4545 unsigned mask = dpp_quad_perm(lane0, lane1, lane2, lane3);
4546 if (ctx->chip_class >= GFX8) {
4547 return ac_build_dpp(ctx, src, src, mask, 0xf, 0xf, false);
4548 } else {
4549 return ac_build_ds_swizzle(ctx, src, (1 << 15) | mask);
4550 }
4551 }
4552
4553 LLVMValueRef
4554 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index)
4555 {
4556 LLVMTypeRef type = LLVMTypeOf(src);
4557 LLVMValueRef result;
4558
4559 index = LLVMBuildMul(ctx->builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
4560 src = LLVMBuildZExt(ctx->builder, src, ctx->i32, "");
4561
4562 result = ac_build_intrinsic(ctx, "llvm.amdgcn.ds.bpermute", ctx->i32,
4563 (LLVMValueRef []) {index, src}, 2,
4564 AC_FUNC_ATTR_READNONE |
4565 AC_FUNC_ATTR_CONVERGENT);
4566 return LLVMBuildTrunc(ctx->builder, result, type, "");
4567 }
4568
4569 LLVMValueRef
4570 ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
4571 unsigned bitsize)
4572 {
4573 LLVMTypeRef type;
4574 char *intr;
4575
4576 if (bitsize == 16) {
4577 intr = "llvm.amdgcn.frexp.exp.i16.f16";
4578 type = ctx->i16;
4579 } else if (bitsize == 32) {
4580 intr = "llvm.amdgcn.frexp.exp.i32.f32";
4581 type = ctx->i32;
4582 } else {
4583 intr = "llvm.amdgcn.frexp.exp.i32.f64";
4584 type = ctx->i32;
4585 }
4586
4587 LLVMValueRef params[] = {
4588 src0,
4589 };
4590 return ac_build_intrinsic(ctx, intr, type, params, 1,
4591 AC_FUNC_ATTR_READNONE);
4592 }
4593 LLVMValueRef
4594 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
4595 unsigned bitsize)
4596 {
4597 LLVMTypeRef type;
4598 char *intr;
4599
4600 if (bitsize == 16) {
4601 intr = "llvm.amdgcn.frexp.mant.f16";
4602 type = ctx->f16;
4603 } else if (bitsize == 32) {
4604 intr = "llvm.amdgcn.frexp.mant.f32";
4605 type = ctx->f32;
4606 } else {
4607 intr = "llvm.amdgcn.frexp.mant.f64";
4608 type = ctx->f64;
4609 }
4610
4611 LLVMValueRef params[] = {
4612 src0,
4613 };
4614 return ac_build_intrinsic(ctx, intr, type, params, 1,
4615 AC_FUNC_ATTR_READNONE);
4616 }
4617
4618 LLVMValueRef
4619 ac_build_canonicalize(struct ac_llvm_context *ctx, LLVMValueRef src0,
4620 unsigned bitsize)
4621 {
4622 LLVMTypeRef type;
4623 char *intr;
4624
4625 if (bitsize == 16) {
4626 intr = "llvm.canonicalize.f16";
4627 type = ctx->f16;
4628 } else if (bitsize == 32) {
4629 intr = "llvm.canonicalize.f32";
4630 type = ctx->f32;
4631 } else {
4632 intr = "llvm.canonicalize.f64";
4633 type = ctx->f64;
4634 }
4635
4636 LLVMValueRef params[] = {
4637 src0,
4638 };
4639 return ac_build_intrinsic(ctx, intr, type, params, 1,
4640 AC_FUNC_ATTR_READNONE);
4641 }
4642
4643 /*
4644 * this takes an I,J coordinate pair,
4645 * and works out the X and Y derivatives.
4646 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
4647 */
4648 LLVMValueRef
4649 ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij)
4650 {
4651 LLVMValueRef result[4], a;
4652 unsigned i;
4653
4654 for (i = 0; i < 2; i++) {
4655 a = LLVMBuildExtractElement(ctx->builder, interp_ij,
4656 LLVMConstInt(ctx->i32, i, false), "");
4657 result[i] = ac_build_ddxy(ctx, AC_TID_MASK_TOP_LEFT, 1, a);
4658 result[2+i] = ac_build_ddxy(ctx, AC_TID_MASK_TOP_LEFT, 2, a);
4659 }
4660 return ac_build_gather_values(ctx, result, 4);
4661 }
4662
4663 LLVMValueRef
4664 ac_build_load_helper_invocation(struct ac_llvm_context *ctx)
4665 {
4666 LLVMValueRef result = ac_build_intrinsic(ctx, "llvm.amdgcn.ps.live",
4667 ctx->i1, NULL, 0,
4668 AC_FUNC_ATTR_READNONE);
4669 result = LLVMBuildNot(ctx->builder, result, "");
4670 return LLVMBuildSExt(ctx->builder, result, ctx->i32, "");
4671 }
4672
4673 LLVMValueRef
4674 ac_build_is_helper_invocation(struct ac_llvm_context *ctx)
4675 {
4676 if (!ctx->postponed_kill)
4677 return ac_build_load_helper_invocation(ctx);
4678
4679 /* !(exact && postponed) */
4680 LLVMValueRef exact = ac_build_intrinsic(ctx, "llvm.amdgcn.ps.live",
4681 ctx->i1, NULL, 0,
4682 AC_FUNC_ATTR_READNONE);
4683
4684 LLVMValueRef postponed = LLVMBuildLoad(ctx->builder, ctx->postponed_kill, "");
4685 LLVMValueRef result = LLVMBuildAnd(ctx->builder, exact, postponed, "");
4686
4687 return LLVMBuildSelect(ctx->builder, result, ctx->i32_0,
4688 LLVMConstInt(ctx->i32, 0xFFFFFFFF, false), "");
4689 }
4690
4691 LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
4692 LLVMValueRef *args, unsigned num_args)
4693 {
4694 LLVMValueRef ret = LLVMBuildCall(ctx->builder, func, args, num_args, "");
4695 LLVMSetInstructionCallConv(ret, LLVMGetFunctionCallConv(func));
4696 return ret;
4697 }
4698
4699 void
4700 ac_export_mrt_z(struct ac_llvm_context *ctx, LLVMValueRef depth,
4701 LLVMValueRef stencil, LLVMValueRef samplemask,
4702 struct ac_export_args *args)
4703 {
4704 unsigned mask = 0;
4705 unsigned format = ac_get_spi_shader_z_format(depth != NULL,
4706 stencil != NULL,
4707 samplemask != NULL);
4708
4709 assert(depth || stencil || samplemask);
4710
4711 memset(args, 0, sizeof(*args));
4712
4713 args->valid_mask = 1; /* whether the EXEC mask is valid */
4714 args->done = 1; /* DONE bit */
4715
4716 /* Specify the target we are exporting */
4717 args->target = V_008DFC_SQ_EXP_MRTZ;
4718
4719 args->compr = 0; /* COMP flag */
4720 args->out[0] = LLVMGetUndef(ctx->f32); /* R, depth */
4721 args->out[1] = LLVMGetUndef(ctx->f32); /* G, stencil test val[0:7], stencil op val[8:15] */
4722 args->out[2] = LLVMGetUndef(ctx->f32); /* B, sample mask */
4723 args->out[3] = LLVMGetUndef(ctx->f32); /* A, alpha to mask */
4724
4725 if (format == V_028710_SPI_SHADER_UINT16_ABGR) {
4726 assert(!depth);
4727 args->compr = 1; /* COMPR flag */
4728
4729 if (stencil) {
4730 /* Stencil should be in X[23:16]. */
4731 stencil = ac_to_integer(ctx, stencil);
4732 stencil = LLVMBuildShl(ctx->builder, stencil,
4733 LLVMConstInt(ctx->i32, 16, 0), "");
4734 args->out[0] = ac_to_float(ctx, stencil);
4735 mask |= 0x3;
4736 }
4737 if (samplemask) {
4738 /* SampleMask should be in Y[15:0]. */
4739 args->out[1] = samplemask;
4740 mask |= 0xc;
4741 }
4742 } else {
4743 if (depth) {
4744 args->out[0] = depth;
4745 mask |= 0x1;
4746 }
4747 if (stencil) {
4748 args->out[1] = stencil;
4749 mask |= 0x2;
4750 }
4751 if (samplemask) {
4752 args->out[2] = samplemask;
4753 mask |= 0x4;
4754 }
4755 }
4756
4757 /* GFX6 (except OLAND and HAINAN) has a bug that it only looks
4758 * at the X writemask component. */
4759 if (ctx->chip_class == GFX6 &&
4760 ctx->family != CHIP_OLAND &&
4761 ctx->family != CHIP_HAINAN)
4762 mask |= 0x1;
4763
4764 /* Specify which components to enable */
4765 args->enabled_channels = mask;
4766 }
4767
4768 /* Send GS Alloc Req message from the first wave of the group to SPI.
4769 * Message payload is:
4770 * - bits 0..10: vertices in group
4771 * - bits 12..22: primitives in group
4772 */
4773 void ac_build_sendmsg_gs_alloc_req(struct ac_llvm_context *ctx, LLVMValueRef wave_id,
4774 LLVMValueRef vtx_cnt, LLVMValueRef prim_cnt)
4775 {
4776 LLVMBuilderRef builder = ctx->builder;
4777 LLVMValueRef tmp;
4778 bool export_dummy_prim = false;
4779
4780 /* HW workaround for a GPU hang with 100% culling.
4781 * We always have to export at least 1 primitive.
4782 * Export a degenerate triangle using vertex 0 for all 3 vertices.
4783 */
4784 if (prim_cnt == ctx->i32_0 && ctx->chip_class == GFX10) {
4785 assert(vtx_cnt == ctx->i32_0);
4786 prim_cnt = ctx->i32_1;
4787 vtx_cnt = ctx->i32_1;
4788 export_dummy_prim = true;
4789 }
4790
4791 ac_build_ifcc(ctx, LLVMBuildICmp(builder, LLVMIntEQ, wave_id, ctx->i32_0, ""), 5020);
4792
4793 tmp = LLVMBuildShl(builder, prim_cnt, LLVMConstInt(ctx->i32, 12, false),"");
4794 tmp = LLVMBuildOr(builder, tmp, vtx_cnt, "");
4795 ac_build_sendmsg(ctx, AC_SENDMSG_GS_ALLOC_REQ, tmp);
4796
4797 if (export_dummy_prim) {
4798 struct ac_ngg_prim prim = {};
4799 /* The vertex indices are 0,0,0. */
4800 prim.passthrough = ctx->i32_0;
4801
4802 struct ac_export_args pos = {};
4803 pos.out[0] = pos.out[1] = pos.out[2] = pos.out[3] = ctx->f32_0;
4804 pos.target = V_008DFC_SQ_EXP_POS;
4805 pos.enabled_channels = 0xf;
4806 pos.done = true;
4807
4808 ac_build_ifcc(ctx, LLVMBuildICmp(builder, LLVMIntEQ, ac_get_thread_id(ctx),
4809 ctx->i32_0, ""), 5021);
4810 ac_build_export_prim(ctx, &prim);
4811 ac_build_export(ctx, &pos);
4812 ac_build_endif(ctx, 5021);
4813 }
4814
4815 ac_build_endif(ctx, 5020);
4816 }
4817
4818 LLVMValueRef ac_pack_prim_export(struct ac_llvm_context *ctx,
4819 const struct ac_ngg_prim *prim)
4820 {
4821 /* The prim export format is:
4822 * - bits 0..8: index 0
4823 * - bit 9: edge flag 0
4824 * - bits 10..18: index 1
4825 * - bit 19: edge flag 1
4826 * - bits 20..28: index 2
4827 * - bit 29: edge flag 2
4828 * - bit 31: null primitive (skip)
4829 */
4830 LLVMBuilderRef builder = ctx->builder;
4831 LLVMValueRef tmp = LLVMBuildZExt(builder, prim->isnull, ctx->i32, "");
4832 LLVMValueRef result = LLVMBuildShl(builder, tmp, LLVMConstInt(ctx->i32, 31, false), "");
4833
4834 for (unsigned i = 0; i < prim->num_vertices; ++i) {
4835 tmp = LLVMBuildShl(builder, prim->index[i],
4836 LLVMConstInt(ctx->i32, 10 * i, false), "");
4837 result = LLVMBuildOr(builder, result, tmp, "");
4838 tmp = LLVMBuildZExt(builder, prim->edgeflag[i], ctx->i32, "");
4839 tmp = LLVMBuildShl(builder, tmp,
4840 LLVMConstInt(ctx->i32, 10 * i + 9, false), "");
4841 result = LLVMBuildOr(builder, result, tmp, "");
4842 }
4843 return result;
4844 }
4845
4846 void ac_build_export_prim(struct ac_llvm_context *ctx,
4847 const struct ac_ngg_prim *prim)
4848 {
4849 struct ac_export_args args;
4850
4851 if (prim->passthrough) {
4852 args.out[0] = prim->passthrough;
4853 } else {
4854 args.out[0] = ac_pack_prim_export(ctx, prim);
4855 }
4856
4857 args.out[0] = LLVMBuildBitCast(ctx->builder, args.out[0], ctx->f32, "");
4858 args.out[1] = LLVMGetUndef(ctx->f32);
4859 args.out[2] = LLVMGetUndef(ctx->f32);
4860 args.out[3] = LLVMGetUndef(ctx->f32);
4861
4862 args.target = V_008DFC_SQ_EXP_PRIM;
4863 args.enabled_channels = 1;
4864 args.done = true;
4865 args.valid_mask = false;
4866 args.compr = false;
4867
4868 ac_build_export(ctx, &args);
4869 }
4870
4871 static LLVMTypeRef
4872 arg_llvm_type(enum ac_arg_type type, unsigned size, struct ac_llvm_context *ctx)
4873 {
4874 if (type == AC_ARG_FLOAT) {
4875 return size == 1 ? ctx->f32 : LLVMVectorType(ctx->f32, size);
4876 } else if (type == AC_ARG_INT) {
4877 return size == 1 ? ctx->i32 : LLVMVectorType(ctx->i32, size);
4878 } else {
4879 LLVMTypeRef ptr_type;
4880 switch (type) {
4881 case AC_ARG_CONST_PTR:
4882 ptr_type = ctx->i8;
4883 break;
4884 case AC_ARG_CONST_FLOAT_PTR:
4885 ptr_type = ctx->f32;
4886 break;
4887 case AC_ARG_CONST_PTR_PTR:
4888 ptr_type = ac_array_in_const32_addr_space(ctx->i8);
4889 break;
4890 case AC_ARG_CONST_DESC_PTR:
4891 ptr_type = ctx->v4i32;
4892 break;
4893 case AC_ARG_CONST_IMAGE_PTR:
4894 ptr_type = ctx->v8i32;
4895 break;
4896 default:
4897 unreachable("unknown arg type");
4898 }
4899 if (size == 1) {
4900 return ac_array_in_const32_addr_space(ptr_type);
4901 } else {
4902 assert(size == 2);
4903 return ac_array_in_const_addr_space(ptr_type);
4904 }
4905 }
4906 }
4907
4908 LLVMValueRef
4909 ac_build_main(const struct ac_shader_args *args,
4910 struct ac_llvm_context *ctx,
4911 enum ac_llvm_calling_convention convention,
4912 const char *name, LLVMTypeRef ret_type,
4913 LLVMModuleRef module)
4914 {
4915 LLVMTypeRef arg_types[AC_MAX_ARGS];
4916
4917 for (unsigned i = 0; i < args->arg_count; i++) {
4918 arg_types[i] = arg_llvm_type(args->args[i].type,
4919 args->args[i].size, ctx);
4920 }
4921
4922 LLVMTypeRef main_function_type =
4923 LLVMFunctionType(ret_type, arg_types, args->arg_count, 0);
4924
4925 LLVMValueRef main_function =
4926 LLVMAddFunction(module, name, main_function_type);
4927 LLVMBasicBlockRef main_function_body =
4928 LLVMAppendBasicBlockInContext(ctx->context, main_function, "main_body");
4929 LLVMPositionBuilderAtEnd(ctx->builder, main_function_body);
4930
4931 LLVMSetFunctionCallConv(main_function, convention);
4932 for (unsigned i = 0; i < args->arg_count; ++i) {
4933 LLVMValueRef P = LLVMGetParam(main_function, i);
4934
4935 if (args->args[i].file != AC_ARG_SGPR)
4936 continue;
4937
4938 ac_add_function_attr(ctx->context, main_function, i + 1, AC_FUNC_ATTR_INREG);
4939
4940 if (LLVMGetTypeKind(LLVMTypeOf(P)) == LLVMPointerTypeKind) {
4941 ac_add_function_attr(ctx->context, main_function, i + 1, AC_FUNC_ATTR_NOALIAS);
4942 ac_add_attr_dereferenceable(P, UINT64_MAX);
4943 ac_add_attr_alignment(P, 32);
4944 }
4945 }
4946
4947 ctx->main_function = main_function;
4948
4949 if (LLVM_VERSION_MAJOR >= 11) {
4950 /* Enable denormals for FP16 and FP64: */
4951 LLVMAddTargetDependentFunctionAttr(main_function, "denormal-fp-math",
4952 "ieee,ieee");
4953 /* Disable denormals for FP32: */
4954 LLVMAddTargetDependentFunctionAttr(main_function, "denormal-fp-math-f32",
4955 "preserve-sign,preserve-sign");
4956 }
4957 return main_function;
4958 }
4959
4960 void ac_build_s_endpgm(struct ac_llvm_context *ctx)
4961 {
4962 LLVMTypeRef calltype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
4963 LLVMValueRef code = LLVMConstInlineAsm(calltype, "s_endpgm", "", true, false);
4964 LLVMBuildCall(ctx->builder, code, NULL, 0, "");
4965 }
4966
4967 LLVMValueRef ac_prefix_bitcount(struct ac_llvm_context *ctx,
4968 LLVMValueRef mask, LLVMValueRef index)
4969 {
4970 LLVMBuilderRef builder = ctx->builder;
4971 LLVMTypeRef type = LLVMTypeOf(mask);
4972
4973 LLVMValueRef bit = LLVMBuildShl(builder, LLVMConstInt(type, 1, 0),
4974 LLVMBuildZExt(builder, index, type, ""), "");
4975 LLVMValueRef prefix_bits = LLVMBuildSub(builder, bit, LLVMConstInt(type, 1, 0), "");
4976 LLVMValueRef prefix_mask = LLVMBuildAnd(builder, mask, prefix_bits, "");
4977 return ac_build_bit_count(ctx, prefix_mask);
4978 }
4979
4980 /* Compute the prefix sum of the "mask" bit array with 128 elements (bits). */
4981 LLVMValueRef ac_prefix_bitcount_2x64(struct ac_llvm_context *ctx,
4982 LLVMValueRef mask[2], LLVMValueRef index)
4983 {
4984 LLVMBuilderRef builder = ctx->builder;
4985 #if 0
4986 /* Reference version using i128. */
4987 LLVMValueRef input_mask =
4988 LLVMBuildBitCast(builder, ac_build_gather_values(ctx, mask, 2), ctx->i128, "");
4989
4990 return ac_prefix_bitcount(ctx, input_mask, index);
4991 #else
4992 /* Optimized version using 2 64-bit masks. */
4993 LLVMValueRef is_hi, is_0, c64, c128, all_bits;
4994 LLVMValueRef prefix_mask[2], shift[2], mask_bcnt0, prefix_bcnt[2];
4995
4996 /* Compute the 128-bit prefix mask. */
4997 c64 = LLVMConstInt(ctx->i32, 64, 0);
4998 c128 = LLVMConstInt(ctx->i32, 128, 0);
4999 all_bits = LLVMConstInt(ctx->i64, UINT64_MAX, 0);
5000 /* The first index that can have non-zero high bits in the prefix mask is 65. */
5001 is_hi = LLVMBuildICmp(builder, LLVMIntUGT, index, c64, "");
5002 is_0 = LLVMBuildICmp(builder, LLVMIntEQ, index, ctx->i32_0, "");
5003 mask_bcnt0 = ac_build_bit_count(ctx, mask[0]);
5004
5005 for (unsigned i = 0; i < 2; i++) {
5006 shift[i] = LLVMBuildSub(builder, i ? c128 : c64, index, "");
5007 /* For i==0, index==0, the right shift by 64 doesn't give the desired result,
5008 * so we handle it by the is_0 select.
5009 * For i==1, index==64, same story, so we handle it by the last is_hi select.
5010 * For i==0, index==64, we shift by 0, which is what we want.
5011 */
5012 prefix_mask[i] = LLVMBuildLShr(builder, all_bits,
5013 LLVMBuildZExt(builder, shift[i], ctx->i64, ""), "");
5014 prefix_mask[i] = LLVMBuildAnd(builder, mask[i], prefix_mask[i], "");
5015 prefix_bcnt[i] = ac_build_bit_count(ctx, prefix_mask[i]);
5016 }
5017
5018 prefix_bcnt[0] = LLVMBuildSelect(builder, is_0, ctx->i32_0, prefix_bcnt[0], "");
5019 prefix_bcnt[0] = LLVMBuildSelect(builder, is_hi, mask_bcnt0, prefix_bcnt[0], "");
5020 prefix_bcnt[1] = LLVMBuildSelect(builder, is_hi, prefix_bcnt[1], ctx->i32_0, "");
5021
5022 return LLVMBuildAdd(builder, prefix_bcnt[0], prefix_bcnt[1], "");
5023 #endif
5024 }
5025
5026 /**
5027 * Convert triangle strip indices to triangle indices. This is used to decompose
5028 * triangle strips into triangles.
5029 */
5030 void ac_build_triangle_strip_indices_to_triangle(struct ac_llvm_context *ctx,
5031 LLVMValueRef is_odd,
5032 LLVMValueRef flatshade_first,
5033 LLVMValueRef index[3])
5034 {
5035 LLVMBuilderRef builder = ctx->builder;
5036 LLVMValueRef out[3];
5037
5038 /* We need to change the vertex order for odd triangles to get correct
5039 * front/back facing by swapping 2 vertex indices, but we also have to
5040 * keep the provoking vertex in the same place.
5041 *
5042 * If the first vertex is provoking, swap index 1 and 2.
5043 * If the last vertex is provoking, swap index 0 and 1.
5044 */
5045 out[0] = LLVMBuildSelect(builder, flatshade_first,
5046 index[0],
5047 LLVMBuildSelect(builder, is_odd,
5048 index[1], index[0], ""), "");
5049 out[1] = LLVMBuildSelect(builder, flatshade_first,
5050 LLVMBuildSelect(builder, is_odd,
5051 index[2], index[1], ""),
5052 LLVMBuildSelect(builder, is_odd,
5053 index[0], index[1], ""), "");
5054 out[2] = LLVMBuildSelect(builder, flatshade_first,
5055 LLVMBuildSelect(builder, is_odd,
5056 index[1], index[2], ""),
5057 index[2], "");
5058 memcpy(index, out, sizeof(out));
5059 }