ac/nir: support 16-bit data in buffer_load_format opcodes
[mesa.git] / src / amd / llvm / ac_llvm_build.c
1 /*
2 * Copyright 2014 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the
6 * "Software"), to deal in the Software without restriction, including
7 * without limitation the rights to use, copy, modify, merge, publish,
8 * distribute, sub license, and/or sell copies of the Software, and to
9 * permit persons to whom the Software is furnished to do so, subject to
10 * the following conditions:
11 *
12 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
13 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
14 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
15 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
16 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
17 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
18 * USE OR OTHER DEALINGS IN THE SOFTWARE.
19 *
20 * The above copyright notice and this permission notice (including the
21 * next paragraph) shall be included in all copies or substantial portions
22 * of the Software.
23 *
24 */
25 /* based on pieces from si_pipe.c and radeon_llvm_emit.c */
26 #include "ac_llvm_build.h"
27
28 #include <llvm-c/Core.h>
29 #include <llvm/Config/llvm-config.h>
30
31 #include "c11/threads.h"
32
33 #include <assert.h>
34 #include <stdio.h>
35
36 #include "ac_llvm_util.h"
37 #include "ac_shader_util.h"
38 #include "ac_exp_param.h"
39 #include "util/bitscan.h"
40 #include "util/macros.h"
41 #include "util/u_atomic.h"
42 #include "util/u_math.h"
43 #include "sid.h"
44
45 #include "shader_enums.h"
46
47 #define AC_LLVM_INITIAL_CF_DEPTH 4
48
49 /* Data for if/else/endif and bgnloop/endloop control flow structures.
50 */
51 struct ac_llvm_flow {
52 /* Loop exit or next part of if/else/endif. */
53 LLVMBasicBlockRef next_block;
54 LLVMBasicBlockRef loop_entry_block;
55 };
56
57 /* Initialize module-independent parts of the context.
58 *
59 * The caller is responsible for initializing ctx::module and ctx::builder.
60 */
61 void
62 ac_llvm_context_init(struct ac_llvm_context *ctx,
63 struct ac_llvm_compiler *compiler,
64 enum chip_class chip_class, enum radeon_family family,
65 enum ac_float_mode float_mode, unsigned wave_size,
66 unsigned ballot_mask_bits)
67 {
68 ctx->context = LLVMContextCreate();
69
70 ctx->chip_class = chip_class;
71 ctx->family = family;
72 ctx->wave_size = wave_size;
73 ctx->ballot_mask_bits = ballot_mask_bits;
74 ctx->float_mode = float_mode;
75 ctx->module = ac_create_module(wave_size == 32 ? compiler->tm_wave32
76 : compiler->tm,
77 ctx->context);
78 ctx->builder = ac_create_builder(ctx->context, float_mode);
79
80 ctx->voidt = LLVMVoidTypeInContext(ctx->context);
81 ctx->i1 = LLVMInt1TypeInContext(ctx->context);
82 ctx->i8 = LLVMInt8TypeInContext(ctx->context);
83 ctx->i16 = LLVMIntTypeInContext(ctx->context, 16);
84 ctx->i32 = LLVMIntTypeInContext(ctx->context, 32);
85 ctx->i64 = LLVMIntTypeInContext(ctx->context, 64);
86 ctx->i128 = LLVMIntTypeInContext(ctx->context, 128);
87 ctx->intptr = ctx->i32;
88 ctx->f16 = LLVMHalfTypeInContext(ctx->context);
89 ctx->f32 = LLVMFloatTypeInContext(ctx->context);
90 ctx->f64 = LLVMDoubleTypeInContext(ctx->context);
91 ctx->v2i16 = LLVMVectorType(ctx->i16, 2);
92 ctx->v4i16 = LLVMVectorType(ctx->i16, 4);
93 ctx->v2f16 = LLVMVectorType(ctx->f16, 2);
94 ctx->v4f16 = LLVMVectorType(ctx->f16, 4);
95 ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
96 ctx->v3i32 = LLVMVectorType(ctx->i32, 3);
97 ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
98 ctx->v2f32 = LLVMVectorType(ctx->f32, 2);
99 ctx->v3f32 = LLVMVectorType(ctx->f32, 3);
100 ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
101 ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
102 ctx->iN_wavemask = LLVMIntTypeInContext(ctx->context, ctx->wave_size);
103 ctx->iN_ballotmask = LLVMIntTypeInContext(ctx->context, ballot_mask_bits);
104
105 ctx->i8_0 = LLVMConstInt(ctx->i8, 0, false);
106 ctx->i8_1 = LLVMConstInt(ctx->i8, 1, false);
107 ctx->i16_0 = LLVMConstInt(ctx->i16, 0, false);
108 ctx->i16_1 = LLVMConstInt(ctx->i16, 1, false);
109 ctx->i32_0 = LLVMConstInt(ctx->i32, 0, false);
110 ctx->i32_1 = LLVMConstInt(ctx->i32, 1, false);
111 ctx->i64_0 = LLVMConstInt(ctx->i64, 0, false);
112 ctx->i64_1 = LLVMConstInt(ctx->i64, 1, false);
113 ctx->i128_0 = LLVMConstInt(ctx->i128, 0, false);
114 ctx->i128_1 = LLVMConstInt(ctx->i128, 1, false);
115 ctx->f16_0 = LLVMConstReal(ctx->f16, 0.0);
116 ctx->f16_1 = LLVMConstReal(ctx->f16, 1.0);
117 ctx->f32_0 = LLVMConstReal(ctx->f32, 0.0);
118 ctx->f32_1 = LLVMConstReal(ctx->f32, 1.0);
119 ctx->f64_0 = LLVMConstReal(ctx->f64, 0.0);
120 ctx->f64_1 = LLVMConstReal(ctx->f64, 1.0);
121
122 ctx->i1false = LLVMConstInt(ctx->i1, 0, false);
123 ctx->i1true = LLVMConstInt(ctx->i1, 1, false);
124
125 ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context,
126 "range", 5);
127
128 ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context,
129 "invariant.load", 14);
130
131 ctx->uniform_md_kind = LLVMGetMDKindIDInContext(ctx->context,
132 "amdgpu.uniform", 14);
133
134 ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
135 ctx->flow = calloc(1, sizeof(*ctx->flow));
136 }
137
138 void
139 ac_llvm_context_dispose(struct ac_llvm_context *ctx)
140 {
141 free(ctx->flow->stack);
142 free(ctx->flow);
143 ctx->flow = NULL;
144 }
145
146 int
147 ac_get_llvm_num_components(LLVMValueRef value)
148 {
149 LLVMTypeRef type = LLVMTypeOf(value);
150 unsigned num_components = LLVMGetTypeKind(type) == LLVMVectorTypeKind
151 ? LLVMGetVectorSize(type)
152 : 1;
153 return num_components;
154 }
155
156 LLVMValueRef
157 ac_llvm_extract_elem(struct ac_llvm_context *ac,
158 LLVMValueRef value,
159 int index)
160 {
161 if (LLVMGetTypeKind(LLVMTypeOf(value)) != LLVMVectorTypeKind) {
162 assert(index == 0);
163 return value;
164 }
165
166 return LLVMBuildExtractElement(ac->builder, value,
167 LLVMConstInt(ac->i32, index, false), "");
168 }
169
170 int
171 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type)
172 {
173 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
174 type = LLVMGetElementType(type);
175
176 if (LLVMGetTypeKind(type) == LLVMIntegerTypeKind)
177 return LLVMGetIntTypeWidth(type);
178
179 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
180 if (LLVMGetPointerAddressSpace(type) == AC_ADDR_SPACE_LDS)
181 return 32;
182 }
183
184 if (type == ctx->f16)
185 return 16;
186 if (type == ctx->f32)
187 return 32;
188 if (type == ctx->f64)
189 return 64;
190
191 unreachable("Unhandled type kind in get_elem_bits");
192 }
193
194 unsigned
195 ac_get_type_size(LLVMTypeRef type)
196 {
197 LLVMTypeKind kind = LLVMGetTypeKind(type);
198
199 switch (kind) {
200 case LLVMIntegerTypeKind:
201 return LLVMGetIntTypeWidth(type) / 8;
202 case LLVMHalfTypeKind:
203 return 2;
204 case LLVMFloatTypeKind:
205 return 4;
206 case LLVMDoubleTypeKind:
207 return 8;
208 case LLVMPointerTypeKind:
209 if (LLVMGetPointerAddressSpace(type) == AC_ADDR_SPACE_CONST_32BIT)
210 return 4;
211 return 8;
212 case LLVMVectorTypeKind:
213 return LLVMGetVectorSize(type) *
214 ac_get_type_size(LLVMGetElementType(type));
215 case LLVMArrayTypeKind:
216 return LLVMGetArrayLength(type) *
217 ac_get_type_size(LLVMGetElementType(type));
218 default:
219 assert(0);
220 return 0;
221 }
222 }
223
224 static LLVMTypeRef to_integer_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
225 {
226 if (t == ctx->i8)
227 return ctx->i8;
228 else if (t == ctx->f16 || t == ctx->i16)
229 return ctx->i16;
230 else if (t == ctx->f32 || t == ctx->i32)
231 return ctx->i32;
232 else if (t == ctx->f64 || t == ctx->i64)
233 return ctx->i64;
234 else
235 unreachable("Unhandled integer size");
236 }
237
238 LLVMTypeRef
239 ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
240 {
241 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
242 LLVMTypeRef elem_type = LLVMGetElementType(t);
243 return LLVMVectorType(to_integer_type_scalar(ctx, elem_type),
244 LLVMGetVectorSize(t));
245 }
246 if (LLVMGetTypeKind(t) == LLVMPointerTypeKind) {
247 switch (LLVMGetPointerAddressSpace(t)) {
248 case AC_ADDR_SPACE_GLOBAL:
249 return ctx->i64;
250 case AC_ADDR_SPACE_CONST_32BIT:
251 case AC_ADDR_SPACE_LDS:
252 return ctx->i32;
253 default:
254 unreachable("unhandled address space");
255 }
256 }
257 return to_integer_type_scalar(ctx, t);
258 }
259
260 LLVMValueRef
261 ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v)
262 {
263 LLVMTypeRef type = LLVMTypeOf(v);
264 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
265 return LLVMBuildPtrToInt(ctx->builder, v, ac_to_integer_type(ctx, type), "");
266 }
267 return LLVMBuildBitCast(ctx->builder, v, ac_to_integer_type(ctx, type), "");
268 }
269
270 LLVMValueRef
271 ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v)
272 {
273 LLVMTypeRef type = LLVMTypeOf(v);
274 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind)
275 return v;
276 return ac_to_integer(ctx, v);
277 }
278
279 static LLVMTypeRef to_float_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
280 {
281 if (t == ctx->i8)
282 return ctx->i8;
283 else if (t == ctx->i16 || t == ctx->f16)
284 return ctx->f16;
285 else if (t == ctx->i32 || t == ctx->f32)
286 return ctx->f32;
287 else if (t == ctx->i64 || t == ctx->f64)
288 return ctx->f64;
289 else
290 unreachable("Unhandled float size");
291 }
292
293 LLVMTypeRef
294 ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
295 {
296 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
297 LLVMTypeRef elem_type = LLVMGetElementType(t);
298 return LLVMVectorType(to_float_type_scalar(ctx, elem_type),
299 LLVMGetVectorSize(t));
300 }
301 return to_float_type_scalar(ctx, t);
302 }
303
304 LLVMValueRef
305 ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v)
306 {
307 LLVMTypeRef type = LLVMTypeOf(v);
308 return LLVMBuildBitCast(ctx->builder, v, ac_to_float_type(ctx, type), "");
309 }
310
311
312 LLVMValueRef
313 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
314 LLVMTypeRef return_type, LLVMValueRef *params,
315 unsigned param_count, unsigned attrib_mask)
316 {
317 LLVMValueRef function, call;
318 bool set_callsite_attrs = !(attrib_mask & AC_FUNC_ATTR_LEGACY);
319
320 function = LLVMGetNamedFunction(ctx->module, name);
321 if (!function) {
322 LLVMTypeRef param_types[32], function_type;
323 unsigned i;
324
325 assert(param_count <= 32);
326
327 for (i = 0; i < param_count; ++i) {
328 assert(params[i]);
329 param_types[i] = LLVMTypeOf(params[i]);
330 }
331 function_type =
332 LLVMFunctionType(return_type, param_types, param_count, 0);
333 function = LLVMAddFunction(ctx->module, name, function_type);
334
335 LLVMSetFunctionCallConv(function, LLVMCCallConv);
336 LLVMSetLinkage(function, LLVMExternalLinkage);
337
338 if (!set_callsite_attrs)
339 ac_add_func_attributes(ctx->context, function, attrib_mask);
340 }
341
342 call = LLVMBuildCall(ctx->builder, function, params, param_count, "");
343 if (set_callsite_attrs)
344 ac_add_func_attributes(ctx->context, call, attrib_mask);
345 return call;
346 }
347
348 /**
349 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
350 * intrinsic names).
351 */
352 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize)
353 {
354 LLVMTypeRef elem_type = type;
355
356 assert(bufsize >= 8);
357
358 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
359 int ret = snprintf(buf, bufsize, "v%u",
360 LLVMGetVectorSize(type));
361 if (ret < 0) {
362 char *type_name = LLVMPrintTypeToString(type);
363 fprintf(stderr, "Error building type name for: %s\n",
364 type_name);
365 LLVMDisposeMessage(type_name);
366 return;
367 }
368 elem_type = LLVMGetElementType(type);
369 buf += ret;
370 bufsize -= ret;
371 }
372 switch (LLVMGetTypeKind(elem_type)) {
373 default: break;
374 case LLVMIntegerTypeKind:
375 snprintf(buf, bufsize, "i%d", LLVMGetIntTypeWidth(elem_type));
376 break;
377 case LLVMHalfTypeKind:
378 snprintf(buf, bufsize, "f16");
379 break;
380 case LLVMFloatTypeKind:
381 snprintf(buf, bufsize, "f32");
382 break;
383 case LLVMDoubleTypeKind:
384 snprintf(buf, bufsize, "f64");
385 break;
386 }
387 }
388
389 /**
390 * Helper function that builds an LLVM IR PHI node and immediately adds
391 * incoming edges.
392 */
393 LLVMValueRef
394 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
395 unsigned count_incoming, LLVMValueRef *values,
396 LLVMBasicBlockRef *blocks)
397 {
398 LLVMValueRef phi = LLVMBuildPhi(ctx->builder, type, "");
399 LLVMAddIncoming(phi, values, blocks, count_incoming);
400 return phi;
401 }
402
403 void ac_build_s_barrier(struct ac_llvm_context *ctx)
404 {
405 ac_build_intrinsic(ctx, "llvm.amdgcn.s.barrier", ctx->voidt, NULL,
406 0, AC_FUNC_ATTR_CONVERGENT);
407 }
408
409 /* Prevent optimizations (at least of memory accesses) across the current
410 * point in the program by emitting empty inline assembly that is marked as
411 * having side effects.
412 *
413 * Optionally, a value can be passed through the inline assembly to prevent
414 * LLVM from hoisting calls to ReadNone functions.
415 */
416 void
417 ac_build_optimization_barrier(struct ac_llvm_context *ctx,
418 LLVMValueRef *pvgpr)
419 {
420 static int counter = 0;
421
422 LLVMBuilderRef builder = ctx->builder;
423 char code[16];
424
425 snprintf(code, sizeof(code), "; %d", p_atomic_inc_return(&counter));
426
427 if (!pvgpr) {
428 LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
429 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "", true, false);
430 LLVMBuildCall(builder, inlineasm, NULL, 0, "");
431 } else {
432 LLVMTypeRef ftype = LLVMFunctionType(ctx->i32, &ctx->i32, 1, false);
433 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "=v,0", true, false);
434 LLVMTypeRef type = LLVMTypeOf(*pvgpr);
435 unsigned bitsize = ac_get_elem_bits(ctx, type);
436 LLVMValueRef vgpr = *pvgpr;
437 LLVMTypeRef vgpr_type;
438 unsigned vgpr_size;
439 LLVMValueRef vgpr0;
440
441 if (bitsize < 32)
442 vgpr = LLVMBuildZExt(ctx->builder, vgpr, ctx->i32, "");
443
444 vgpr_type = LLVMTypeOf(vgpr);
445 vgpr_size = ac_get_type_size(vgpr_type);
446
447 assert(vgpr_size % 4 == 0);
448
449 vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
450 vgpr0 = LLVMBuildExtractElement(builder, vgpr, ctx->i32_0, "");
451 vgpr0 = LLVMBuildCall(builder, inlineasm, &vgpr0, 1, "");
452 vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
453 vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
454
455 if (bitsize < 32)
456 vgpr = LLVMBuildTrunc(builder, vgpr, type, "");
457
458 *pvgpr = vgpr;
459 }
460 }
461
462 LLVMValueRef
463 ac_build_shader_clock(struct ac_llvm_context *ctx, nir_scope scope)
464 {
465 const char *name = scope == NIR_SCOPE_DEVICE ? "llvm.amdgcn.s.memrealtime" : "llvm.amdgcn.s.memtime";
466 LLVMValueRef tmp = ac_build_intrinsic(ctx, name, ctx->i64, NULL, 0, 0);
467 return LLVMBuildBitCast(ctx->builder, tmp, ctx->v2i32, "");
468 }
469
470 LLVMValueRef
471 ac_build_ballot(struct ac_llvm_context *ctx,
472 LLVMValueRef value)
473 {
474 const char *name;
475
476 if (LLVM_VERSION_MAJOR >= 9) {
477 if (ctx->wave_size == 64)
478 name = "llvm.amdgcn.icmp.i64.i32";
479 else
480 name = "llvm.amdgcn.icmp.i32.i32";
481 } else {
482 name = "llvm.amdgcn.icmp.i32";
483 }
484 LLVMValueRef args[3] = {
485 value,
486 ctx->i32_0,
487 LLVMConstInt(ctx->i32, LLVMIntNE, 0)
488 };
489
490 /* We currently have no other way to prevent LLVM from lifting the icmp
491 * calls to a dominating basic block.
492 */
493 ac_build_optimization_barrier(ctx, &args[0]);
494
495 args[0] = ac_to_integer(ctx, args[0]);
496
497 return ac_build_intrinsic(ctx, name, ctx->iN_wavemask, args, 3,
498 AC_FUNC_ATTR_NOUNWIND |
499 AC_FUNC_ATTR_READNONE |
500 AC_FUNC_ATTR_CONVERGENT);
501 }
502
503 LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
504 LLVMValueRef value)
505 {
506 const char *name;
507
508 if (LLVM_VERSION_MAJOR >= 9) {
509 if (ctx->wave_size == 64)
510 name = "llvm.amdgcn.icmp.i64.i1";
511 else
512 name = "llvm.amdgcn.icmp.i32.i1";
513 } else {
514 name = "llvm.amdgcn.icmp.i1";
515 }
516 LLVMValueRef args[3] = {
517 value,
518 ctx->i1false,
519 LLVMConstInt(ctx->i32, LLVMIntNE, 0),
520 };
521
522 return ac_build_intrinsic(ctx, name, ctx->iN_wavemask, args, 3,
523 AC_FUNC_ATTR_NOUNWIND |
524 AC_FUNC_ATTR_READNONE |
525 AC_FUNC_ATTR_CONVERGENT);
526 }
527
528 LLVMValueRef
529 ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value)
530 {
531 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
532 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
533 return LLVMBuildICmp(ctx->builder, LLVMIntEQ, vote_set, active_set, "");
534 }
535
536 LLVMValueRef
537 ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value)
538 {
539 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
540 return LLVMBuildICmp(ctx->builder, LLVMIntNE, vote_set,
541 LLVMConstInt(ctx->iN_wavemask, 0, 0), "");
542 }
543
544 LLVMValueRef
545 ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value)
546 {
547 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
548 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
549
550 LLVMValueRef all = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
551 vote_set, active_set, "");
552 LLVMValueRef none = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
553 vote_set,
554 LLVMConstInt(ctx->iN_wavemask, 0, 0), "");
555 return LLVMBuildOr(ctx->builder, all, none, "");
556 }
557
558 LLVMValueRef
559 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
560 unsigned value_count, unsigned component)
561 {
562 LLVMValueRef vec = NULL;
563
564 if (value_count == 1) {
565 return values[component];
566 } else if (!value_count)
567 unreachable("value_count is 0");
568
569 for (unsigned i = component; i < value_count + component; i++) {
570 LLVMValueRef value = values[i];
571
572 if (i == component)
573 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
574 LLVMValueRef index = LLVMConstInt(ctx->i32, i - component, false);
575 vec = LLVMBuildInsertElement(ctx->builder, vec, value, index, "");
576 }
577 return vec;
578 }
579
580 LLVMValueRef
581 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
582 LLVMValueRef *values,
583 unsigned value_count,
584 unsigned value_stride,
585 bool load,
586 bool always_vector)
587 {
588 LLVMBuilderRef builder = ctx->builder;
589 LLVMValueRef vec = NULL;
590 unsigned i;
591
592 if (value_count == 1 && !always_vector) {
593 if (load)
594 return LLVMBuildLoad(builder, values[0], "");
595 return values[0];
596 } else if (!value_count)
597 unreachable("value_count is 0");
598
599 for (i = 0; i < value_count; i++) {
600 LLVMValueRef value = values[i * value_stride];
601 if (load)
602 value = LLVMBuildLoad(builder, value, "");
603
604 if (!i)
605 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
606 LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
607 vec = LLVMBuildInsertElement(builder, vec, value, index, "");
608 }
609 return vec;
610 }
611
612 LLVMValueRef
613 ac_build_gather_values(struct ac_llvm_context *ctx,
614 LLVMValueRef *values,
615 unsigned value_count)
616 {
617 return ac_build_gather_values_extended(ctx, values, value_count, 1, false, false);
618 }
619
620 /* Expand a scalar or vector to <dst_channels x type> by filling the remaining
621 * channels with undef. Extract at most src_channels components from the input.
622 */
623 static LLVMValueRef
624 ac_build_expand(struct ac_llvm_context *ctx,
625 LLVMValueRef value,
626 unsigned src_channels,
627 unsigned dst_channels)
628 {
629 LLVMTypeRef elemtype;
630 LLVMValueRef chan[dst_channels];
631
632 if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMVectorTypeKind) {
633 unsigned vec_size = LLVMGetVectorSize(LLVMTypeOf(value));
634
635 if (src_channels == dst_channels && vec_size == dst_channels)
636 return value;
637
638 src_channels = MIN2(src_channels, vec_size);
639
640 for (unsigned i = 0; i < src_channels; i++)
641 chan[i] = ac_llvm_extract_elem(ctx, value, i);
642
643 elemtype = LLVMGetElementType(LLVMTypeOf(value));
644 } else {
645 if (src_channels) {
646 assert(src_channels == 1);
647 chan[0] = value;
648 }
649 elemtype = LLVMTypeOf(value);
650 }
651
652 for (unsigned i = src_channels; i < dst_channels; i++)
653 chan[i] = LLVMGetUndef(elemtype);
654
655 return ac_build_gather_values(ctx, chan, dst_channels);
656 }
657
658 /* Extract components [start, start + channels) from a vector.
659 */
660 LLVMValueRef
661 ac_extract_components(struct ac_llvm_context *ctx,
662 LLVMValueRef value,
663 unsigned start,
664 unsigned channels)
665 {
666 LLVMValueRef chan[channels];
667
668 for (unsigned i = 0; i < channels; i++)
669 chan[i] = ac_llvm_extract_elem(ctx, value, i + start);
670
671 return ac_build_gather_values(ctx, chan, channels);
672 }
673
674 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
675 * with undef. Extract at most num_channels components from the input.
676 */
677 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
678 LLVMValueRef value,
679 unsigned num_channels)
680 {
681 return ac_build_expand(ctx, value, num_channels, 4);
682 }
683
684 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value)
685 {
686 unsigned type_size = ac_get_type_size(LLVMTypeOf(value));
687 const char *name;
688
689 if (type_size == 2)
690 name = "llvm.rint.f16";
691 else if (type_size == 4)
692 name = "llvm.rint.f32";
693 else
694 name = "llvm.rint.f64";
695
696 return ac_build_intrinsic(ctx, name, LLVMTypeOf(value), &value, 1,
697 AC_FUNC_ATTR_READNONE);
698 }
699
700 LLVMValueRef
701 ac_build_fdiv(struct ac_llvm_context *ctx,
702 LLVMValueRef num,
703 LLVMValueRef den)
704 {
705 unsigned type_size = ac_get_type_size(LLVMTypeOf(den));
706 const char *name;
707
708 if (type_size == 2)
709 name = "llvm.amdgcn.rcp.f16";
710 else if (type_size == 4)
711 name = "llvm.amdgcn.rcp.f32";
712 else
713 name = "llvm.amdgcn.rcp.f64";
714
715 LLVMValueRef rcp = ac_build_intrinsic(ctx, name, LLVMTypeOf(den),
716 &den, 1, AC_FUNC_ATTR_READNONE);
717
718 return LLVMBuildFMul(ctx->builder, num, rcp, "");
719 }
720
721 /* See fast_idiv_by_const.h. */
722 /* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
723 LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
724 LLVMValueRef num,
725 LLVMValueRef multiplier,
726 LLVMValueRef pre_shift,
727 LLVMValueRef post_shift,
728 LLVMValueRef increment)
729 {
730 LLVMBuilderRef builder = ctx->builder;
731
732 num = LLVMBuildLShr(builder, num, pre_shift, "");
733 num = LLVMBuildMul(builder,
734 LLVMBuildZExt(builder, num, ctx->i64, ""),
735 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
736 num = LLVMBuildAdd(builder, num,
737 LLVMBuildZExt(builder, increment, ctx->i64, ""), "");
738 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
739 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
740 return LLVMBuildLShr(builder, num, post_shift, "");
741 }
742
743 /* See fast_idiv_by_const.h. */
744 /* If num != UINT_MAX, this more efficient version can be used. */
745 /* Set: increment = util_fast_udiv_info::increment; */
746 LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
747 LLVMValueRef num,
748 LLVMValueRef multiplier,
749 LLVMValueRef pre_shift,
750 LLVMValueRef post_shift,
751 LLVMValueRef increment)
752 {
753 LLVMBuilderRef builder = ctx->builder;
754
755 num = LLVMBuildLShr(builder, num, pre_shift, "");
756 num = LLVMBuildNUWAdd(builder, num, increment, "");
757 num = LLVMBuildMul(builder,
758 LLVMBuildZExt(builder, num, ctx->i64, ""),
759 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
760 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
761 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
762 return LLVMBuildLShr(builder, num, post_shift, "");
763 }
764
765 /* See fast_idiv_by_const.h. */
766 /* Both operands must fit in 31 bits and the divisor must not be 1. */
767 LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
768 LLVMValueRef num,
769 LLVMValueRef multiplier,
770 LLVMValueRef post_shift)
771 {
772 LLVMBuilderRef builder = ctx->builder;
773
774 num = LLVMBuildMul(builder,
775 LLVMBuildZExt(builder, num, ctx->i64, ""),
776 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
777 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
778 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
779 return LLVMBuildLShr(builder, num, post_shift, "");
780 }
781
782 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
783 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
784 * already multiplied by two. id is the cube face number.
785 */
786 struct cube_selection_coords {
787 LLVMValueRef stc[2];
788 LLVMValueRef ma;
789 LLVMValueRef id;
790 };
791
792 static void
793 build_cube_intrinsic(struct ac_llvm_context *ctx,
794 LLVMValueRef in[3],
795 struct cube_selection_coords *out)
796 {
797 LLVMTypeRef f32 = ctx->f32;
798
799 out->stc[1] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubetc",
800 f32, in, 3, AC_FUNC_ATTR_READNONE);
801 out->stc[0] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubesc",
802 f32, in, 3, AC_FUNC_ATTR_READNONE);
803 out->ma = ac_build_intrinsic(ctx, "llvm.amdgcn.cubema",
804 f32, in, 3, AC_FUNC_ATTR_READNONE);
805 out->id = ac_build_intrinsic(ctx, "llvm.amdgcn.cubeid",
806 f32, in, 3, AC_FUNC_ATTR_READNONE);
807 }
808
809 /**
810 * Build a manual selection sequence for cube face sc/tc coordinates and
811 * major axis vector (multiplied by 2 for consistency) for the given
812 * vec3 \p coords, for the face implied by \p selcoords.
813 *
814 * For the major axis, we always adjust the sign to be in the direction of
815 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
816 * the selcoords major axis.
817 */
818 static void build_cube_select(struct ac_llvm_context *ctx,
819 const struct cube_selection_coords *selcoords,
820 const LLVMValueRef *coords,
821 LLVMValueRef *out_st,
822 LLVMValueRef *out_ma)
823 {
824 LLVMBuilderRef builder = ctx->builder;
825 LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
826 LLVMValueRef is_ma_positive;
827 LLVMValueRef sgn_ma;
828 LLVMValueRef is_ma_z, is_not_ma_z;
829 LLVMValueRef is_ma_y;
830 LLVMValueRef is_ma_x;
831 LLVMValueRef sgn;
832 LLVMValueRef tmp;
833
834 is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
835 selcoords->ma, LLVMConstReal(f32, 0.0), "");
836 sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
837 LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
838
839 is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
840 is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
841 is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
842 LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
843 is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
844
845 /* Select sc */
846 tmp = LLVMBuildSelect(builder, is_ma_x, coords[2], coords[0], "");
847 sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
848 LLVMBuildSelect(builder, is_ma_z, sgn_ma,
849 LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
850 out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
851
852 /* Select tc */
853 tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
854 sgn = LLVMBuildSelect(builder, is_ma_y, sgn_ma,
855 LLVMConstReal(f32, -1.0), "");
856 out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
857
858 /* Select ma */
859 tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
860 LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
861 tmp = ac_build_intrinsic(ctx, "llvm.fabs.f32",
862 ctx->f32, &tmp, 1, AC_FUNC_ATTR_READNONE);
863 *out_ma = LLVMBuildFMul(builder, tmp, LLVMConstReal(f32, 2.0), "");
864 }
865
866 void
867 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
868 bool is_deriv, bool is_array, bool is_lod,
869 LLVMValueRef *coords_arg,
870 LLVMValueRef *derivs_arg)
871 {
872
873 LLVMBuilderRef builder = ctx->builder;
874 struct cube_selection_coords selcoords;
875 LLVMValueRef coords[3];
876 LLVMValueRef invma;
877
878 if (is_array && !is_lod) {
879 LLVMValueRef tmp = ac_build_round(ctx, coords_arg[3]);
880
881 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
882 *
883 * "For Array forms, the array layer used will be
884 *
885 * max(0, min(d−1, floor(layer+0.5)))
886 *
887 * where d is the depth of the texture array and layer
888 * comes from the component indicated in the tables below.
889 * Workaroudn for an issue where the layer is taken from a
890 * helper invocation which happens to fall on a different
891 * layer due to extrapolation."
892 *
893 * GFX8 and earlier attempt to implement this in hardware by
894 * clamping the value of coords[2] = (8 * layer) + face.
895 * Unfortunately, this means that the we end up with the wrong
896 * face when clamping occurs.
897 *
898 * Clamp the layer earlier to work around the issue.
899 */
900 if (ctx->chip_class <= GFX8) {
901 LLVMValueRef ge0;
902 ge0 = LLVMBuildFCmp(builder, LLVMRealOGE, tmp, ctx->f32_0, "");
903 tmp = LLVMBuildSelect(builder, ge0, tmp, ctx->f32_0, "");
904 }
905
906 coords_arg[3] = tmp;
907 }
908
909 build_cube_intrinsic(ctx, coords_arg, &selcoords);
910
911 invma = ac_build_intrinsic(ctx, "llvm.fabs.f32",
912 ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
913 invma = ac_build_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
914
915 for (int i = 0; i < 2; ++i)
916 coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
917
918 coords[2] = selcoords.id;
919
920 if (is_deriv && derivs_arg) {
921 LLVMValueRef derivs[4];
922 int axis;
923
924 /* Convert cube derivatives to 2D derivatives. */
925 for (axis = 0; axis < 2; axis++) {
926 LLVMValueRef deriv_st[2];
927 LLVMValueRef deriv_ma;
928
929 /* Transform the derivative alongside the texture
930 * coordinate. Mathematically, the correct formula is
931 * as follows. Assume we're projecting onto the +Z face
932 * and denote by dx/dh the derivative of the (original)
933 * X texture coordinate with respect to horizontal
934 * window coordinates. The projection onto the +Z face
935 * plane is:
936 *
937 * f(x,z) = x/z
938 *
939 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
940 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
941 *
942 * This motivatives the implementation below.
943 *
944 * Whether this actually gives the expected results for
945 * apps that might feed in derivatives obtained via
946 * finite differences is anyone's guess. The OpenGL spec
947 * seems awfully quiet about how textureGrad for cube
948 * maps should be handled.
949 */
950 build_cube_select(ctx, &selcoords, &derivs_arg[axis * 3],
951 deriv_st, &deriv_ma);
952
953 deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
954
955 for (int i = 0; i < 2; ++i)
956 derivs[axis * 2 + i] =
957 LLVMBuildFSub(builder,
958 LLVMBuildFMul(builder, deriv_st[i], invma, ""),
959 LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
960 }
961
962 memcpy(derivs_arg, derivs, sizeof(derivs));
963 }
964
965 /* Shift the texture coordinate. This must be applied after the
966 * derivative calculation.
967 */
968 for (int i = 0; i < 2; ++i)
969 coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
970
971 if (is_array) {
972 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
973 /* coords_arg.w component - array_index for cube arrays */
974 coords[2] = ac_build_fmad(ctx, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), coords[2]);
975 }
976
977 memcpy(coords_arg, coords, sizeof(coords));
978 }
979
980
981 LLVMValueRef
982 ac_build_fs_interp(struct ac_llvm_context *ctx,
983 LLVMValueRef llvm_chan,
984 LLVMValueRef attr_number,
985 LLVMValueRef params,
986 LLVMValueRef i,
987 LLVMValueRef j)
988 {
989 LLVMValueRef args[5];
990 LLVMValueRef p1;
991
992 args[0] = i;
993 args[1] = llvm_chan;
994 args[2] = attr_number;
995 args[3] = params;
996
997 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1",
998 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
999
1000 args[0] = p1;
1001 args[1] = j;
1002 args[2] = llvm_chan;
1003 args[3] = attr_number;
1004 args[4] = params;
1005
1006 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2",
1007 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
1008 }
1009
1010 LLVMValueRef
1011 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
1012 LLVMValueRef llvm_chan,
1013 LLVMValueRef attr_number,
1014 LLVMValueRef params,
1015 LLVMValueRef i,
1016 LLVMValueRef j)
1017 {
1018 LLVMValueRef args[6];
1019 LLVMValueRef p1;
1020
1021 args[0] = i;
1022 args[1] = llvm_chan;
1023 args[2] = attr_number;
1024 args[3] = ctx->i1false;
1025 args[4] = params;
1026
1027 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1.f16",
1028 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
1029
1030 args[0] = p1;
1031 args[1] = j;
1032 args[2] = llvm_chan;
1033 args[3] = attr_number;
1034 args[4] = ctx->i1false;
1035 args[5] = params;
1036
1037 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2.f16",
1038 ctx->f16, args, 6, AC_FUNC_ATTR_READNONE);
1039 }
1040
1041 LLVMValueRef
1042 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
1043 LLVMValueRef parameter,
1044 LLVMValueRef llvm_chan,
1045 LLVMValueRef attr_number,
1046 LLVMValueRef params)
1047 {
1048 LLVMValueRef args[4];
1049
1050 args[0] = parameter;
1051 args[1] = llvm_chan;
1052 args[2] = attr_number;
1053 args[3] = params;
1054
1055 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.mov",
1056 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
1057 }
1058
1059 LLVMValueRef
1060 ac_build_gep_ptr(struct ac_llvm_context *ctx,
1061 LLVMValueRef base_ptr,
1062 LLVMValueRef index)
1063 {
1064 return LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
1065 }
1066
1067 LLVMValueRef
1068 ac_build_gep0(struct ac_llvm_context *ctx,
1069 LLVMValueRef base_ptr,
1070 LLVMValueRef index)
1071 {
1072 LLVMValueRef indices[2] = {
1073 ctx->i32_0,
1074 index,
1075 };
1076 return LLVMBuildGEP(ctx->builder, base_ptr, indices, 2, "");
1077 }
1078
1079 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
1080 LLVMValueRef index)
1081 {
1082 return LLVMBuildPointerCast(ctx->builder,
1083 LLVMBuildGEP(ctx->builder, ptr, &index, 1, ""),
1084 LLVMTypeOf(ptr), "");
1085 }
1086
1087 void
1088 ac_build_indexed_store(struct ac_llvm_context *ctx,
1089 LLVMValueRef base_ptr, LLVMValueRef index,
1090 LLVMValueRef value)
1091 {
1092 LLVMBuildStore(ctx->builder, value,
1093 ac_build_gep0(ctx, base_ptr, index));
1094 }
1095
1096 /**
1097 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
1098 * It's equivalent to doing a load from &base_ptr[index].
1099 *
1100 * \param base_ptr Where the array starts.
1101 * \param index The element index into the array.
1102 * \param uniform Whether the base_ptr and index can be assumed to be
1103 * dynamically uniform (i.e. load to an SGPR)
1104 * \param invariant Whether the load is invariant (no other opcodes affect it)
1105 * \param no_unsigned_wraparound
1106 * For all possible re-associations and re-distributions of an expression
1107 * "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
1108 * without inbounds in base_ptr), this parameter is true if "addr + offset"
1109 * does not result in an unsigned integer wraparound. This is used for
1110 * optimal code generation of 32-bit pointer arithmetic.
1111 *
1112 * For example, a 32-bit immediate offset that causes a 32-bit unsigned
1113 * integer wraparound can't be an imm offset in s_load_dword, because
1114 * the instruction performs "addr + offset" in 64 bits.
1115 *
1116 * Expected usage for bindless textures by chaining GEPs:
1117 * // possible unsigned wraparound, don't use InBounds:
1118 * ptr1 = LLVMBuildGEP(base_ptr, index);
1119 * image = load(ptr1); // becomes "s_load ptr1, 0"
1120 *
1121 * ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1122 * sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1123 */
1124 static LLVMValueRef
1125 ac_build_load_custom(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1126 LLVMValueRef index, bool uniform, bool invariant,
1127 bool no_unsigned_wraparound)
1128 {
1129 LLVMValueRef pointer, result;
1130
1131 if (no_unsigned_wraparound &&
1132 LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr)) == AC_ADDR_SPACE_CONST_32BIT)
1133 pointer = LLVMBuildInBoundsGEP(ctx->builder, base_ptr, &index, 1, "");
1134 else
1135 pointer = LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
1136
1137 if (uniform)
1138 LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
1139 result = LLVMBuildLoad(ctx->builder, pointer, "");
1140 if (invariant)
1141 LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
1142 return result;
1143 }
1144
1145 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1146 LLVMValueRef index)
1147 {
1148 return ac_build_load_custom(ctx, base_ptr, index, false, false, false);
1149 }
1150
1151 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
1152 LLVMValueRef base_ptr, LLVMValueRef index)
1153 {
1154 return ac_build_load_custom(ctx, base_ptr, index, false, true, false);
1155 }
1156
1157 /* This assumes that there is no unsigned integer wraparound during the address
1158 * computation, excluding all GEPs within base_ptr. */
1159 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
1160 LLVMValueRef base_ptr, LLVMValueRef index)
1161 {
1162 return ac_build_load_custom(ctx, base_ptr, index, true, true, true);
1163 }
1164
1165 /* See ac_build_load_custom() documentation. */
1166 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
1167 LLVMValueRef base_ptr, LLVMValueRef index)
1168 {
1169 return ac_build_load_custom(ctx, base_ptr, index, true, true, false);
1170 }
1171
1172 static unsigned get_load_cache_policy(struct ac_llvm_context *ctx,
1173 unsigned cache_policy)
1174 {
1175 return cache_policy |
1176 (ctx->chip_class >= GFX10 && cache_policy & ac_glc ? ac_dlc : 0);
1177 }
1178
1179 static void
1180 ac_build_buffer_store_common(struct ac_llvm_context *ctx,
1181 LLVMValueRef rsrc,
1182 LLVMValueRef data,
1183 LLVMValueRef vindex,
1184 LLVMValueRef voffset,
1185 LLVMValueRef soffset,
1186 unsigned cache_policy,
1187 bool use_format,
1188 bool structurized)
1189 {
1190 LLVMValueRef args[6];
1191 int idx = 0;
1192 args[idx++] = data;
1193 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1194 if (structurized)
1195 args[idx++] = vindex ? vindex : ctx->i32_0;
1196 args[idx++] = voffset ? voffset : ctx->i32_0;
1197 args[idx++] = soffset ? soffset : ctx->i32_0;
1198 args[idx++] = LLVMConstInt(ctx->i32, cache_policy, 0);
1199 const char *indexing_kind = structurized ? "struct" : "raw";
1200 char name[256], type_name[8];
1201
1202 ac_build_type_name_for_intr(LLVMTypeOf(data), type_name, sizeof(type_name));
1203
1204 if (use_format) {
1205 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.format.%s",
1206 indexing_kind, type_name);
1207 } else {
1208 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.%s",
1209 indexing_kind, type_name);
1210 }
1211
1212 ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
1213 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
1214 }
1215
1216 void
1217 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
1218 LLVMValueRef rsrc,
1219 LLVMValueRef data,
1220 LLVMValueRef vindex,
1221 LLVMValueRef voffset,
1222 unsigned cache_policy)
1223 {
1224 ac_build_buffer_store_common(ctx, rsrc, data, vindex, voffset, NULL,
1225 cache_policy, true, true);
1226 }
1227
1228 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1229 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1230 * or v4i32 (num_channels=3,4).
1231 */
1232 void
1233 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
1234 LLVMValueRef rsrc,
1235 LLVMValueRef vdata,
1236 unsigned num_channels,
1237 LLVMValueRef voffset,
1238 LLVMValueRef soffset,
1239 unsigned inst_offset,
1240 unsigned cache_policy)
1241 {
1242 /* Split 3 channel stores, because only LLVM 9+ support 3-channel
1243 * intrinsics. */
1244 if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false)) {
1245 LLVMValueRef v[3], v01;
1246
1247 for (int i = 0; i < 3; i++) {
1248 v[i] = LLVMBuildExtractElement(ctx->builder, vdata,
1249 LLVMConstInt(ctx->i32, i, 0), "");
1250 }
1251 v01 = ac_build_gather_values(ctx, v, 2);
1252
1253 ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset,
1254 soffset, inst_offset, cache_policy);
1255 ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset,
1256 soffset, inst_offset + 8,
1257 cache_policy);
1258 return;
1259 }
1260
1261 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1262 * (voffset is swizzled, but soffset isn't swizzled).
1263 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1264 */
1265 if (!(cache_policy & ac_swizzled)) {
1266 LLVMValueRef offset = soffset;
1267
1268 if (inst_offset)
1269 offset = LLVMBuildAdd(ctx->builder, offset,
1270 LLVMConstInt(ctx->i32, inst_offset, 0), "");
1271
1272 ac_build_buffer_store_common(ctx, rsrc, ac_to_float(ctx, vdata),
1273 ctx->i32_0, voffset, offset,
1274 cache_policy, false, false);
1275 return;
1276 }
1277
1278 static const unsigned dfmts[] = {
1279 V_008F0C_BUF_DATA_FORMAT_32,
1280 V_008F0C_BUF_DATA_FORMAT_32_32,
1281 V_008F0C_BUF_DATA_FORMAT_32_32_32,
1282 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1283 };
1284 unsigned dfmt = dfmts[num_channels - 1];
1285 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1286 LLVMValueRef immoffset = LLVMConstInt(ctx->i32, inst_offset, 0);
1287
1288 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1289 immoffset, num_channels, dfmt, nfmt, cache_policy);
1290 }
1291
1292 static LLVMValueRef
1293 ac_build_buffer_load_common(struct ac_llvm_context *ctx,
1294 LLVMValueRef rsrc,
1295 LLVMValueRef vindex,
1296 LLVMValueRef voffset,
1297 LLVMValueRef soffset,
1298 unsigned num_channels,
1299 LLVMTypeRef channel_type,
1300 unsigned cache_policy,
1301 bool can_speculate,
1302 bool use_format,
1303 bool structurized)
1304 {
1305 LLVMValueRef args[5];
1306 int idx = 0;
1307 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1308 if (structurized)
1309 args[idx++] = vindex ? vindex : ctx->i32_0;
1310 args[idx++] = voffset ? voffset : ctx->i32_0;
1311 args[idx++] = soffset ? soffset : ctx->i32_0;
1312 args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
1313 unsigned func = !ac_has_vec3_support(ctx->chip_class, use_format) && num_channels == 3 ? 4 : num_channels;
1314 const char *indexing_kind = structurized ? "struct" : "raw";
1315 char name[256], type_name[8];
1316
1317 /* D16 is only supported on gfx8+ */
1318 assert((channel_type != ctx->f16 && channel_type != ctx->i16) ||
1319 ctx->chip_class >= GFX8);
1320
1321 LLVMTypeRef type = func > 1 ? LLVMVectorType(channel_type, func) : channel_type;
1322 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1323
1324 if (use_format) {
1325 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.format.%s",
1326 indexing_kind, type_name);
1327 } else {
1328 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.%s",
1329 indexing_kind, type_name);
1330 }
1331
1332 return ac_build_intrinsic(ctx, name, type, args, idx,
1333 ac_get_load_intr_attribs(can_speculate));
1334 }
1335
1336 LLVMValueRef
1337 ac_build_buffer_load(struct ac_llvm_context *ctx,
1338 LLVMValueRef rsrc,
1339 int num_channels,
1340 LLVMValueRef vindex,
1341 LLVMValueRef voffset,
1342 LLVMValueRef soffset,
1343 unsigned inst_offset,
1344 unsigned cache_policy,
1345 bool can_speculate,
1346 bool allow_smem)
1347 {
1348 LLVMValueRef offset = LLVMConstInt(ctx->i32, inst_offset, 0);
1349 if (voffset)
1350 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1351 if (soffset)
1352 offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
1353
1354 if (allow_smem && !(cache_policy & ac_slc) &&
1355 (!(cache_policy & ac_glc) || ctx->chip_class >= GFX8)) {
1356 assert(vindex == NULL);
1357
1358 LLVMValueRef result[8];
1359
1360 for (int i = 0; i < num_channels; i++) {
1361 if (i) {
1362 offset = LLVMBuildAdd(ctx->builder, offset,
1363 LLVMConstInt(ctx->i32, 4, 0), "");
1364 }
1365 LLVMValueRef args[3] = {
1366 rsrc,
1367 offset,
1368 LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0),
1369 };
1370 result[i] = ac_build_intrinsic(ctx,
1371 "llvm.amdgcn.s.buffer.load.f32",
1372 ctx->f32, args, 3,
1373 AC_FUNC_ATTR_READNONE);
1374 }
1375 if (num_channels == 1)
1376 return result[0];
1377
1378 if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false))
1379 result[num_channels++] = LLVMGetUndef(ctx->f32);
1380 return ac_build_gather_values(ctx, result, num_channels);
1381 }
1382
1383 return ac_build_buffer_load_common(ctx, rsrc, vindex,
1384 offset, ctx->i32_0,
1385 num_channels, ctx->f32,
1386 cache_policy,
1387 can_speculate, false, false);
1388 }
1389
1390 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
1391 LLVMValueRef rsrc,
1392 LLVMValueRef vindex,
1393 LLVMValueRef voffset,
1394 unsigned num_channels,
1395 unsigned cache_policy,
1396 bool can_speculate,
1397 bool d16)
1398 {
1399 return ac_build_buffer_load_common(ctx, rsrc, vindex, voffset,
1400 ctx->i32_0, num_channels,
1401 d16 ? ctx->f16 : ctx->f32,
1402 cache_policy, can_speculate,
1403 true, true);
1404 }
1405
1406 static LLVMValueRef
1407 ac_build_tbuffer_load(struct ac_llvm_context *ctx,
1408 LLVMValueRef rsrc,
1409 LLVMValueRef vindex,
1410 LLVMValueRef voffset,
1411 LLVMValueRef soffset,
1412 LLVMValueRef immoffset,
1413 unsigned num_channels,
1414 unsigned dfmt,
1415 unsigned nfmt,
1416 unsigned cache_policy,
1417 bool can_speculate,
1418 bool structurized)
1419 {
1420 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1421
1422 LLVMValueRef args[6];
1423 int idx = 0;
1424 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1425 if (structurized)
1426 args[idx++] = vindex ? vindex : ctx->i32_0;
1427 args[idx++] = voffset ? voffset : ctx->i32_0;
1428 args[idx++] = soffset ? soffset : ctx->i32_0;
1429 args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx->chip_class, dfmt, nfmt), 0);
1430 args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
1431 unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
1432 const char *indexing_kind = structurized ? "struct" : "raw";
1433 char name[256], type_name[8];
1434
1435 LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
1436 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1437
1438 snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.load.%s",
1439 indexing_kind, type_name);
1440
1441 return ac_build_intrinsic(ctx, name, type, args, idx,
1442 ac_get_load_intr_attribs(can_speculate));
1443 }
1444
1445 LLVMValueRef
1446 ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
1447 LLVMValueRef rsrc,
1448 LLVMValueRef vindex,
1449 LLVMValueRef voffset,
1450 LLVMValueRef soffset,
1451 LLVMValueRef immoffset,
1452 unsigned num_channels,
1453 unsigned dfmt,
1454 unsigned nfmt,
1455 unsigned cache_policy,
1456 bool can_speculate)
1457 {
1458 return ac_build_tbuffer_load(ctx, rsrc, vindex, voffset, soffset,
1459 immoffset, num_channels, dfmt, nfmt,
1460 cache_policy, can_speculate, true);
1461 }
1462
1463 LLVMValueRef
1464 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
1465 LLVMValueRef rsrc,
1466 LLVMValueRef voffset,
1467 LLVMValueRef soffset,
1468 LLVMValueRef immoffset,
1469 unsigned num_channels,
1470 unsigned dfmt,
1471 unsigned nfmt,
1472 unsigned cache_policy,
1473 bool can_speculate)
1474 {
1475 return ac_build_tbuffer_load(ctx, rsrc, NULL, voffset, soffset,
1476 immoffset, num_channels, dfmt, nfmt,
1477 cache_policy, can_speculate, false);
1478 }
1479
1480 LLVMValueRef
1481 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
1482 LLVMValueRef rsrc,
1483 LLVMValueRef voffset,
1484 LLVMValueRef soffset,
1485 LLVMValueRef immoffset,
1486 unsigned cache_policy)
1487 {
1488 LLVMValueRef res;
1489
1490 if (LLVM_VERSION_MAJOR >= 9) {
1491 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1492
1493 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1494 res = ac_build_buffer_load_common(ctx, rsrc, NULL,
1495 voffset, soffset,
1496 1, ctx->i16, cache_policy,
1497 false, false, false);
1498 } else {
1499 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
1500 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1501
1502 res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
1503 immoffset, 1, dfmt, nfmt, cache_policy,
1504 false);
1505
1506 res = LLVMBuildTrunc(ctx->builder, res, ctx->i16, "");
1507 }
1508
1509 return res;
1510 }
1511
1512 LLVMValueRef
1513 ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
1514 LLVMValueRef rsrc,
1515 LLVMValueRef voffset,
1516 LLVMValueRef soffset,
1517 LLVMValueRef immoffset,
1518 unsigned cache_policy)
1519 {
1520 LLVMValueRef res;
1521
1522 if (LLVM_VERSION_MAJOR >= 9) {
1523 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1524
1525 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1526 res = ac_build_buffer_load_common(ctx, rsrc, NULL,
1527 voffset, soffset,
1528 1, ctx->i8, cache_policy,
1529 false, false, false);
1530 } else {
1531 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
1532 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1533
1534 res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
1535 immoffset, 1, dfmt, nfmt, cache_policy,
1536 false);
1537
1538 res = LLVMBuildTrunc(ctx->builder, res, ctx->i8, "");
1539 }
1540
1541 return res;
1542 }
1543
1544 /**
1545 * Convert an 11- or 10-bit unsigned floating point number to an f32.
1546 *
1547 * The input exponent is expected to be biased analogous to IEEE-754, i.e. by
1548 * 2^(exp_bits-1) - 1 (as defined in OpenGL and other graphics APIs).
1549 */
1550 static LLVMValueRef
1551 ac_ufN_to_float(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned exp_bits, unsigned mant_bits)
1552 {
1553 assert(LLVMTypeOf(src) == ctx->i32);
1554
1555 LLVMValueRef tmp;
1556 LLVMValueRef mantissa;
1557 mantissa = LLVMBuildAnd(ctx->builder, src, LLVMConstInt(ctx->i32, (1 << mant_bits) - 1, false), "");
1558
1559 /* Converting normal numbers is just a shift + correcting the exponent bias */
1560 unsigned normal_shift = 23 - mant_bits;
1561 unsigned bias_shift = 127 - ((1 << (exp_bits - 1)) - 1);
1562 LLVMValueRef shifted, normal;
1563
1564 shifted = LLVMBuildShl(ctx->builder, src, LLVMConstInt(ctx->i32, normal_shift, false), "");
1565 normal = LLVMBuildAdd(ctx->builder, shifted, LLVMConstInt(ctx->i32, bias_shift << 23, false), "");
1566
1567 /* Converting nan/inf numbers is the same, but with a different exponent update */
1568 LLVMValueRef naninf;
1569 naninf = LLVMBuildOr(ctx->builder, normal, LLVMConstInt(ctx->i32, 0xff << 23, false), "");
1570
1571 /* Converting denormals is the complex case: determine the leading zeros of the
1572 * mantissa to obtain the correct shift for the mantissa and exponent correction.
1573 */
1574 LLVMValueRef denormal;
1575 LLVMValueRef params[2] = {
1576 mantissa,
1577 ctx->i1true, /* result can be undef when arg is 0 */
1578 };
1579 LLVMValueRef ctlz = ac_build_intrinsic(ctx, "llvm.ctlz.i32", ctx->i32,
1580 params, 2, AC_FUNC_ATTR_READNONE);
1581
1582 /* Shift such that the leading 1 ends up as the LSB of the exponent field. */
1583 tmp = LLVMBuildSub(ctx->builder, ctlz, LLVMConstInt(ctx->i32, 8, false), "");
1584 denormal = LLVMBuildShl(ctx->builder, mantissa, tmp, "");
1585
1586 unsigned denormal_exp = bias_shift + (32 - mant_bits) - 1;
1587 tmp = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, denormal_exp, false), ctlz, "");
1588 tmp = LLVMBuildShl(ctx->builder, tmp, LLVMConstInt(ctx->i32, 23, false), "");
1589 denormal = LLVMBuildAdd(ctx->builder, denormal, tmp, "");
1590
1591 /* Select the final result. */
1592 LLVMValueRef result;
1593
1594 tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src,
1595 LLVMConstInt(ctx->i32, ((1 << exp_bits) - 1) << mant_bits, false), "");
1596 result = LLVMBuildSelect(ctx->builder, tmp, naninf, normal, "");
1597
1598 tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src,
1599 LLVMConstInt(ctx->i32, 1 << mant_bits, false), "");
1600 result = LLVMBuildSelect(ctx->builder, tmp, result, denormal, "");
1601
1602 tmp = LLVMBuildICmp(ctx->builder, LLVMIntNE, src, ctx->i32_0, "");
1603 result = LLVMBuildSelect(ctx->builder, tmp, result, ctx->i32_0, "");
1604
1605 return ac_to_float(ctx, result);
1606 }
1607
1608 /**
1609 * Generate a fully general open coded buffer format fetch with all required
1610 * fixups suitable for vertex fetch, using non-format buffer loads.
1611 *
1612 * Some combinations of argument values have special interpretations:
1613 * - size = 8 bytes, format = fixed indicates PIPE_FORMAT_R11G11B10_FLOAT
1614 * - size = 8 bytes, format != {float,fixed} indicates a 2_10_10_10 data format
1615 *
1616 * \param log_size log(size of channel in bytes)
1617 * \param num_channels number of channels (1 to 4)
1618 * \param format AC_FETCH_FORMAT_xxx value
1619 * \param reverse whether XYZ channels are reversed
1620 * \param known_aligned whether the source is known to be aligned to hardware's
1621 * effective element size for loading the given format
1622 * (note: this means dword alignment for 8_8_8_8, 16_16, etc.)
1623 * \param rsrc buffer resource descriptor
1624 * \return the resulting vector of floats or integers bitcast to <4 x i32>
1625 */
1626 LLVMValueRef
1627 ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
1628 unsigned log_size,
1629 unsigned num_channels,
1630 unsigned format,
1631 bool reverse,
1632 bool known_aligned,
1633 LLVMValueRef rsrc,
1634 LLVMValueRef vindex,
1635 LLVMValueRef voffset,
1636 LLVMValueRef soffset,
1637 unsigned cache_policy,
1638 bool can_speculate)
1639 {
1640 LLVMValueRef tmp;
1641 unsigned load_log_size = log_size;
1642 unsigned load_num_channels = num_channels;
1643 if (log_size == 3) {
1644 load_log_size = 2;
1645 if (format == AC_FETCH_FORMAT_FLOAT) {
1646 load_num_channels = 2 * num_channels;
1647 } else {
1648 load_num_channels = 1; /* 10_11_11 or 2_10_10_10 */
1649 }
1650 }
1651
1652 int log_recombine = 0;
1653 if (ctx->chip_class == GFX6 && !known_aligned) {
1654 /* Avoid alignment restrictions by loading one byte at a time. */
1655 load_num_channels <<= load_log_size;
1656 log_recombine = load_log_size;
1657 load_log_size = 0;
1658 } else if (load_num_channels == 2 || load_num_channels == 4) {
1659 log_recombine = -util_logbase2(load_num_channels);
1660 load_num_channels = 1;
1661 load_log_size += -log_recombine;
1662 }
1663
1664 assert(load_log_size >= 2 || LLVM_VERSION_MAJOR >= 9);
1665
1666 LLVMValueRef loads[32]; /* up to 32 bytes */
1667 for (unsigned i = 0; i < load_num_channels; ++i) {
1668 tmp = LLVMBuildAdd(ctx->builder, soffset,
1669 LLVMConstInt(ctx->i32, i << load_log_size, false), "");
1670 LLVMTypeRef channel_type = load_log_size == 0 ? ctx->i8 :
1671 load_log_size == 1 ? ctx->i16 : ctx->i32;
1672 unsigned num_channels = 1 << (MAX2(load_log_size, 2) - 2);
1673 loads[i] = ac_build_buffer_load_common(
1674 ctx, rsrc, vindex, voffset, tmp,
1675 num_channels, channel_type, cache_policy,
1676 can_speculate, false, true);
1677 if (load_log_size >= 2)
1678 loads[i] = ac_to_integer(ctx, loads[i]);
1679 }
1680
1681 if (log_recombine > 0) {
1682 /* Recombine bytes if necessary (GFX6 only) */
1683 LLVMTypeRef dst_type = log_recombine == 2 ? ctx->i32 : ctx->i16;
1684
1685 for (unsigned src = 0, dst = 0; src < load_num_channels; ++dst) {
1686 LLVMValueRef accum = NULL;
1687 for (unsigned i = 0; i < (1 << log_recombine); ++i, ++src) {
1688 tmp = LLVMBuildZExt(ctx->builder, loads[src], dst_type, "");
1689 if (i == 0) {
1690 accum = tmp;
1691 } else {
1692 tmp = LLVMBuildShl(ctx->builder, tmp,
1693 LLVMConstInt(dst_type, 8 * i, false), "");
1694 accum = LLVMBuildOr(ctx->builder, accum, tmp, "");
1695 }
1696 }
1697 loads[dst] = accum;
1698 }
1699 } else if (log_recombine < 0) {
1700 /* Split vectors of dwords */
1701 if (load_log_size > 2) {
1702 assert(load_num_channels == 1);
1703 LLVMValueRef loaded = loads[0];
1704 unsigned log_split = load_log_size - 2;
1705 log_recombine += log_split;
1706 load_num_channels = 1 << log_split;
1707 load_log_size = 2;
1708 for (unsigned i = 0; i < load_num_channels; ++i) {
1709 tmp = LLVMConstInt(ctx->i32, i, false);
1710 loads[i] = LLVMBuildExtractElement(ctx->builder, loaded, tmp, "");
1711 }
1712 }
1713
1714 /* Further split dwords and shorts if required */
1715 if (log_recombine < 0) {
1716 for (unsigned src = load_num_channels,
1717 dst = load_num_channels << -log_recombine;
1718 src > 0; --src) {
1719 unsigned dst_bits = 1 << (3 + load_log_size + log_recombine);
1720 LLVMTypeRef dst_type = LLVMIntTypeInContext(ctx->context, dst_bits);
1721 LLVMValueRef loaded = loads[src - 1];
1722 LLVMTypeRef loaded_type = LLVMTypeOf(loaded);
1723 for (unsigned i = 1 << -log_recombine; i > 0; --i, --dst) {
1724 tmp = LLVMConstInt(loaded_type, dst_bits * (i - 1), false);
1725 tmp = LLVMBuildLShr(ctx->builder, loaded, tmp, "");
1726 loads[dst - 1] = LLVMBuildTrunc(ctx->builder, tmp, dst_type, "");
1727 }
1728 }
1729 }
1730 }
1731
1732 if (log_size == 3) {
1733 if (format == AC_FETCH_FORMAT_FLOAT) {
1734 for (unsigned i = 0; i < num_channels; ++i) {
1735 tmp = ac_build_gather_values(ctx, &loads[2 * i], 2);
1736 loads[i] = LLVMBuildBitCast(ctx->builder, tmp, ctx->f64, "");
1737 }
1738 } else if (format == AC_FETCH_FORMAT_FIXED) {
1739 /* 10_11_11_FLOAT */
1740 LLVMValueRef data = loads[0];
1741 LLVMValueRef i32_2047 = LLVMConstInt(ctx->i32, 2047, false);
1742 LLVMValueRef r = LLVMBuildAnd(ctx->builder, data, i32_2047, "");
1743 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 11, false), "");
1744 LLVMValueRef g = LLVMBuildAnd(ctx->builder, tmp, i32_2047, "");
1745 LLVMValueRef b = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 22, false), "");
1746
1747 loads[0] = ac_to_integer(ctx, ac_ufN_to_float(ctx, r, 5, 6));
1748 loads[1] = ac_to_integer(ctx, ac_ufN_to_float(ctx, g, 5, 6));
1749 loads[2] = ac_to_integer(ctx, ac_ufN_to_float(ctx, b, 5, 5));
1750
1751 num_channels = 3;
1752 log_size = 2;
1753 format = AC_FETCH_FORMAT_FLOAT;
1754 } else {
1755 /* 2_10_10_10 data formats */
1756 LLVMValueRef data = loads[0];
1757 LLVMTypeRef i10 = LLVMIntTypeInContext(ctx->context, 10);
1758 LLVMTypeRef i2 = LLVMIntTypeInContext(ctx->context, 2);
1759 loads[0] = LLVMBuildTrunc(ctx->builder, data, i10, "");
1760 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 10, false), "");
1761 loads[1] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
1762 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 20, false), "");
1763 loads[2] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
1764 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 30, false), "");
1765 loads[3] = LLVMBuildTrunc(ctx->builder, tmp, i2, "");
1766
1767 num_channels = 4;
1768 }
1769 }
1770
1771 if (format == AC_FETCH_FORMAT_FLOAT) {
1772 if (log_size != 2) {
1773 for (unsigned chan = 0; chan < num_channels; ++chan) {
1774 tmp = ac_to_float(ctx, loads[chan]);
1775 if (log_size == 3)
1776 tmp = LLVMBuildFPTrunc(ctx->builder, tmp, ctx->f32, "");
1777 else if (log_size == 1)
1778 tmp = LLVMBuildFPExt(ctx->builder, tmp, ctx->f32, "");
1779 loads[chan] = ac_to_integer(ctx, tmp);
1780 }
1781 }
1782 } else if (format == AC_FETCH_FORMAT_UINT) {
1783 if (log_size != 2) {
1784 for (unsigned chan = 0; chan < num_channels; ++chan)
1785 loads[chan] = LLVMBuildZExt(ctx->builder, loads[chan], ctx->i32, "");
1786 }
1787 } else if (format == AC_FETCH_FORMAT_SINT) {
1788 if (log_size != 2) {
1789 for (unsigned chan = 0; chan < num_channels; ++chan)
1790 loads[chan] = LLVMBuildSExt(ctx->builder, loads[chan], ctx->i32, "");
1791 }
1792 } else {
1793 bool unsign = format == AC_FETCH_FORMAT_UNORM ||
1794 format == AC_FETCH_FORMAT_USCALED ||
1795 format == AC_FETCH_FORMAT_UINT;
1796
1797 for (unsigned chan = 0; chan < num_channels; ++chan) {
1798 if (unsign) {
1799 tmp = LLVMBuildUIToFP(ctx->builder, loads[chan], ctx->f32, "");
1800 } else {
1801 tmp = LLVMBuildSIToFP(ctx->builder, loads[chan], ctx->f32, "");
1802 }
1803
1804 LLVMValueRef scale = NULL;
1805 if (format == AC_FETCH_FORMAT_FIXED) {
1806 assert(log_size == 2);
1807 scale = LLVMConstReal(ctx->f32, 1.0 / 0x10000);
1808 } else if (format == AC_FETCH_FORMAT_UNORM) {
1809 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
1810 scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << bits) - 1));
1811 } else if (format == AC_FETCH_FORMAT_SNORM) {
1812 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
1813 scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << (bits - 1)) - 1));
1814 }
1815 if (scale)
1816 tmp = LLVMBuildFMul(ctx->builder, tmp, scale, "");
1817
1818 if (format == AC_FETCH_FORMAT_SNORM) {
1819 /* Clamp to [-1, 1] */
1820 LLVMValueRef neg_one = LLVMConstReal(ctx->f32, -1.0);
1821 LLVMValueRef clamp =
1822 LLVMBuildFCmp(ctx->builder, LLVMRealULT, tmp, neg_one, "");
1823 tmp = LLVMBuildSelect(ctx->builder, clamp, neg_one, tmp, "");
1824 }
1825
1826 loads[chan] = ac_to_integer(ctx, tmp);
1827 }
1828 }
1829
1830 while (num_channels < 4) {
1831 if (format == AC_FETCH_FORMAT_UINT || format == AC_FETCH_FORMAT_SINT) {
1832 loads[num_channels] = num_channels == 3 ? ctx->i32_1 : ctx->i32_0;
1833 } else {
1834 loads[num_channels] = ac_to_integer(ctx, num_channels == 3 ? ctx->f32_1 : ctx->f32_0);
1835 }
1836 num_channels++;
1837 }
1838
1839 if (reverse) {
1840 tmp = loads[0];
1841 loads[0] = loads[2];
1842 loads[2] = tmp;
1843 }
1844
1845 return ac_build_gather_values(ctx, loads, 4);
1846 }
1847
1848 static void
1849 ac_build_tbuffer_store(struct ac_llvm_context *ctx,
1850 LLVMValueRef rsrc,
1851 LLVMValueRef vdata,
1852 LLVMValueRef vindex,
1853 LLVMValueRef voffset,
1854 LLVMValueRef soffset,
1855 LLVMValueRef immoffset,
1856 unsigned num_channels,
1857 unsigned dfmt,
1858 unsigned nfmt,
1859 unsigned cache_policy,
1860 bool structurized)
1861 {
1862 voffset = LLVMBuildAdd(ctx->builder, voffset ? voffset : ctx->i32_0,
1863 immoffset, "");
1864
1865 LLVMValueRef args[7];
1866 int idx = 0;
1867 args[idx++] = vdata;
1868 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1869 if (structurized)
1870 args[idx++] = vindex ? vindex : ctx->i32_0;
1871 args[idx++] = voffset ? voffset : ctx->i32_0;
1872 args[idx++] = soffset ? soffset : ctx->i32_0;
1873 args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx->chip_class, dfmt, nfmt), 0);
1874 args[idx++] = LLVMConstInt(ctx->i32, cache_policy, 0);
1875 unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
1876 const char *indexing_kind = structurized ? "struct" : "raw";
1877 char name[256], type_name[8];
1878
1879 LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
1880 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1881
1882 snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.store.%s",
1883 indexing_kind, type_name);
1884
1885 ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
1886 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
1887 }
1888
1889 void
1890 ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
1891 LLVMValueRef rsrc,
1892 LLVMValueRef vdata,
1893 LLVMValueRef vindex,
1894 LLVMValueRef voffset,
1895 LLVMValueRef soffset,
1896 LLVMValueRef immoffset,
1897 unsigned num_channels,
1898 unsigned dfmt,
1899 unsigned nfmt,
1900 unsigned cache_policy)
1901 {
1902 ac_build_tbuffer_store(ctx, rsrc, vdata, vindex, voffset, soffset,
1903 immoffset, num_channels, dfmt, nfmt, cache_policy,
1904 true);
1905 }
1906
1907 void
1908 ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
1909 LLVMValueRef rsrc,
1910 LLVMValueRef vdata,
1911 LLVMValueRef voffset,
1912 LLVMValueRef soffset,
1913 LLVMValueRef immoffset,
1914 unsigned num_channels,
1915 unsigned dfmt,
1916 unsigned nfmt,
1917 unsigned cache_policy)
1918 {
1919 ac_build_tbuffer_store(ctx, rsrc, vdata, NULL, voffset, soffset,
1920 immoffset, num_channels, dfmt, nfmt, cache_policy,
1921 false);
1922 }
1923
1924 void
1925 ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
1926 LLVMValueRef rsrc,
1927 LLVMValueRef vdata,
1928 LLVMValueRef voffset,
1929 LLVMValueRef soffset,
1930 unsigned cache_policy)
1931 {
1932 vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i16, "");
1933
1934 if (LLVM_VERSION_MAJOR >= 9) {
1935 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1936 ac_build_buffer_store_common(ctx, rsrc, vdata, NULL,
1937 voffset, soffset, cache_policy,
1938 false, false);
1939 } else {
1940 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
1941 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1942
1943 vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
1944
1945 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1946 ctx->i32_0, 1, dfmt, nfmt, cache_policy);
1947 }
1948 }
1949
1950 void
1951 ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
1952 LLVMValueRef rsrc,
1953 LLVMValueRef vdata,
1954 LLVMValueRef voffset,
1955 LLVMValueRef soffset,
1956 unsigned cache_policy)
1957 {
1958 vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i8, "");
1959
1960 if (LLVM_VERSION_MAJOR >= 9) {
1961 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1962 ac_build_buffer_store_common(ctx, rsrc, vdata, NULL,
1963 voffset, soffset, cache_policy,
1964 false, false);
1965 } else {
1966 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
1967 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1968
1969 vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
1970
1971 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1972 ctx->i32_0, 1, dfmt, nfmt, cache_policy);
1973 }
1974 }
1975 /**
1976 * Set range metadata on an instruction. This can only be used on load and
1977 * call instructions. If you know an instruction can only produce the values
1978 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1979 * \p lo is the minimum value inclusive.
1980 * \p hi is the maximum value exclusive.
1981 */
1982 static void set_range_metadata(struct ac_llvm_context *ctx,
1983 LLVMValueRef value, unsigned lo, unsigned hi)
1984 {
1985 LLVMValueRef range_md, md_args[2];
1986 LLVMTypeRef type = LLVMTypeOf(value);
1987 LLVMContextRef context = LLVMGetTypeContext(type);
1988
1989 md_args[0] = LLVMConstInt(type, lo, false);
1990 md_args[1] = LLVMConstInt(type, hi, false);
1991 range_md = LLVMMDNodeInContext(context, md_args, 2);
1992 LLVMSetMetadata(value, ctx->range_md_kind, range_md);
1993 }
1994
1995 LLVMValueRef
1996 ac_get_thread_id(struct ac_llvm_context *ctx)
1997 {
1998 LLVMValueRef tid;
1999
2000 LLVMValueRef tid_args[2];
2001 tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
2002 tid_args[1] = ctx->i32_0;
2003 tid_args[1] = ac_build_intrinsic(ctx,
2004 "llvm.amdgcn.mbcnt.lo", ctx->i32,
2005 tid_args, 2, AC_FUNC_ATTR_READNONE);
2006
2007 if (ctx->wave_size == 32) {
2008 tid = tid_args[1];
2009 } else {
2010 tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
2011 ctx->i32, tid_args,
2012 2, AC_FUNC_ATTR_READNONE);
2013 }
2014 set_range_metadata(ctx, tid, 0, ctx->wave_size);
2015 return tid;
2016 }
2017
2018 /*
2019 * AMD GCN implements derivatives using the local data store (LDS)
2020 * All writes to the LDS happen in all executing threads at
2021 * the same time. TID is the Thread ID for the current
2022 * thread and is a value between 0 and 63, representing
2023 * the thread's position in the wavefront.
2024 *
2025 * For the pixel shader threads are grouped into quads of four pixels.
2026 * The TIDs of the pixels of a quad are:
2027 *
2028 * +------+------+
2029 * |4n + 0|4n + 1|
2030 * +------+------+
2031 * |4n + 2|4n + 3|
2032 * +------+------+
2033 *
2034 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
2035 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
2036 * the current pixel's column, and masking with 0xfffffffe yields the TID
2037 * of the left pixel of the current pixel's row.
2038 *
2039 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
2040 * adding 2 yields the TID of the pixel below the top pixel.
2041 */
2042 LLVMValueRef
2043 ac_build_ddxy(struct ac_llvm_context *ctx,
2044 uint32_t mask,
2045 int idx,
2046 LLVMValueRef val)
2047 {
2048 unsigned tl_lanes[4], trbl_lanes[4];
2049 char name[32], type[8];
2050 LLVMValueRef tl, trbl;
2051 LLVMTypeRef result_type;
2052 LLVMValueRef result;
2053
2054 result_type = ac_to_float_type(ctx, LLVMTypeOf(val));
2055
2056 if (result_type == ctx->f16)
2057 val = LLVMBuildZExt(ctx->builder, val, ctx->i32, "");
2058
2059 for (unsigned i = 0; i < 4; ++i) {
2060 tl_lanes[i] = i & mask;
2061 trbl_lanes[i] = (i & mask) + idx;
2062 }
2063
2064 tl = ac_build_quad_swizzle(ctx, val,
2065 tl_lanes[0], tl_lanes[1],
2066 tl_lanes[2], tl_lanes[3]);
2067 trbl = ac_build_quad_swizzle(ctx, val,
2068 trbl_lanes[0], trbl_lanes[1],
2069 trbl_lanes[2], trbl_lanes[3]);
2070
2071 if (result_type == ctx->f16) {
2072 tl = LLVMBuildTrunc(ctx->builder, tl, ctx->i16, "");
2073 trbl = LLVMBuildTrunc(ctx->builder, trbl, ctx->i16, "");
2074 }
2075
2076 tl = LLVMBuildBitCast(ctx->builder, tl, result_type, "");
2077 trbl = LLVMBuildBitCast(ctx->builder, trbl, result_type, "");
2078 result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
2079
2080 ac_build_type_name_for_intr(result_type, type, sizeof(type));
2081 snprintf(name, sizeof(name), "llvm.amdgcn.wqm.%s", type);
2082
2083 return ac_build_intrinsic(ctx, name, result_type, &result, 1, 0);
2084 }
2085
2086 void
2087 ac_build_sendmsg(struct ac_llvm_context *ctx,
2088 uint32_t msg,
2089 LLVMValueRef wave_id)
2090 {
2091 LLVMValueRef args[2];
2092 args[0] = LLVMConstInt(ctx->i32, msg, false);
2093 args[1] = wave_id;
2094 ac_build_intrinsic(ctx, "llvm.amdgcn.s.sendmsg", ctx->voidt, args, 2, 0);
2095 }
2096
2097 LLVMValueRef
2098 ac_build_imsb(struct ac_llvm_context *ctx,
2099 LLVMValueRef arg,
2100 LLVMTypeRef dst_type)
2101 {
2102 LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.amdgcn.sffbh.i32",
2103 dst_type, &arg, 1,
2104 AC_FUNC_ATTR_READNONE);
2105
2106 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
2107 * the index from LSB. Invert it by doing "31 - msb". */
2108 msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
2109 msb, "");
2110
2111 LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
2112 LLVMValueRef cond = LLVMBuildOr(ctx->builder,
2113 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
2114 arg, ctx->i32_0, ""),
2115 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
2116 arg, all_ones, ""), "");
2117
2118 return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
2119 }
2120
2121 LLVMValueRef
2122 ac_build_umsb(struct ac_llvm_context *ctx,
2123 LLVMValueRef arg,
2124 LLVMTypeRef dst_type)
2125 {
2126 const char *intrin_name;
2127 LLVMTypeRef type;
2128 LLVMValueRef highest_bit;
2129 LLVMValueRef zero;
2130 unsigned bitsize;
2131
2132 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(arg));
2133 switch (bitsize) {
2134 case 64:
2135 intrin_name = "llvm.ctlz.i64";
2136 type = ctx->i64;
2137 highest_bit = LLVMConstInt(ctx->i64, 63, false);
2138 zero = ctx->i64_0;
2139 break;
2140 case 32:
2141 intrin_name = "llvm.ctlz.i32";
2142 type = ctx->i32;
2143 highest_bit = LLVMConstInt(ctx->i32, 31, false);
2144 zero = ctx->i32_0;
2145 break;
2146 case 16:
2147 intrin_name = "llvm.ctlz.i16";
2148 type = ctx->i16;
2149 highest_bit = LLVMConstInt(ctx->i16, 15, false);
2150 zero = ctx->i16_0;
2151 break;
2152 case 8:
2153 intrin_name = "llvm.ctlz.i8";
2154 type = ctx->i8;
2155 highest_bit = LLVMConstInt(ctx->i8, 7, false);
2156 zero = ctx->i8_0;
2157 break;
2158 default:
2159 unreachable(!"invalid bitsize");
2160 break;
2161 }
2162
2163 LLVMValueRef params[2] = {
2164 arg,
2165 ctx->i1true,
2166 };
2167
2168 LLVMValueRef msb = ac_build_intrinsic(ctx, intrin_name, type,
2169 params, 2,
2170 AC_FUNC_ATTR_READNONE);
2171
2172 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
2173 * the index from LSB. Invert it by doing "31 - msb". */
2174 msb = LLVMBuildSub(ctx->builder, highest_bit, msb, "");
2175
2176 if (bitsize == 64) {
2177 msb = LLVMBuildTrunc(ctx->builder, msb, ctx->i32, "");
2178 } else if (bitsize < 32) {
2179 msb = LLVMBuildSExt(ctx->builder, msb, ctx->i32, "");
2180 }
2181
2182 /* check for zero */
2183 return LLVMBuildSelect(ctx->builder,
2184 LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg, zero, ""),
2185 LLVMConstInt(ctx->i32, -1, true), msb, "");
2186 }
2187
2188 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
2189 LLVMValueRef b)
2190 {
2191 char name[64], type[64];
2192
2193 ac_build_type_name_for_intr(LLVMTypeOf(a), type, sizeof(type));
2194 snprintf(name, sizeof(name), "llvm.minnum.%s", type);
2195 LLVMValueRef args[2] = {a, b};
2196 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
2197 AC_FUNC_ATTR_READNONE);
2198 }
2199
2200 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
2201 LLVMValueRef b)
2202 {
2203 char name[64], type[64];
2204
2205 ac_build_type_name_for_intr(LLVMTypeOf(a), type, sizeof(type));
2206 snprintf(name, sizeof(name), "llvm.maxnum.%s", type);
2207 LLVMValueRef args[2] = {a, b};
2208 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
2209 AC_FUNC_ATTR_READNONE);
2210 }
2211
2212 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
2213 LLVMValueRef b)
2214 {
2215 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSLE, a, b, "");
2216 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2217 }
2218
2219 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
2220 LLVMValueRef b)
2221 {
2222 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, a, b, "");
2223 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2224 }
2225
2226 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a,
2227 LLVMValueRef b)
2228 {
2229 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntULE, a, b, "");
2230 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2231 }
2232
2233 LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a,
2234 LLVMValueRef b)
2235 {
2236 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, a, b, "");
2237 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2238 }
2239
2240 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value)
2241 {
2242 LLVMTypeRef t = LLVMTypeOf(value);
2243 return ac_build_fmin(ctx, ac_build_fmax(ctx, value, LLVMConstReal(t, 0.0)),
2244 LLVMConstReal(t, 1.0));
2245 }
2246
2247 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a)
2248 {
2249 LLVMValueRef args[9];
2250
2251 args[0] = LLVMConstInt(ctx->i32, a->target, 0);
2252 args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
2253
2254 if (a->compr) {
2255 args[2] = LLVMBuildBitCast(ctx->builder, a->out[0],
2256 ctx->v2i16, "");
2257 args[3] = LLVMBuildBitCast(ctx->builder, a->out[1],
2258 ctx->v2i16, "");
2259 args[4] = LLVMConstInt(ctx->i1, a->done, 0);
2260 args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
2261
2262 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16",
2263 ctx->voidt, args, 6, 0);
2264 } else {
2265 args[2] = a->out[0];
2266 args[3] = a->out[1];
2267 args[4] = a->out[2];
2268 args[5] = a->out[3];
2269 args[6] = LLVMConstInt(ctx->i1, a->done, 0);
2270 args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
2271
2272 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.f32",
2273 ctx->voidt, args, 8, 0);
2274 }
2275 }
2276
2277 void ac_build_export_null(struct ac_llvm_context *ctx)
2278 {
2279 struct ac_export_args args;
2280
2281 args.enabled_channels = 0x0; /* enabled channels */
2282 args.valid_mask = 1; /* whether the EXEC mask is valid */
2283 args.done = 1; /* DONE bit */
2284 args.target = V_008DFC_SQ_EXP_NULL;
2285 args.compr = 0; /* COMPR flag (0 = 32-bit export) */
2286 args.out[0] = LLVMGetUndef(ctx->f32); /* R */
2287 args.out[1] = LLVMGetUndef(ctx->f32); /* G */
2288 args.out[2] = LLVMGetUndef(ctx->f32); /* B */
2289 args.out[3] = LLVMGetUndef(ctx->f32); /* A */
2290
2291 ac_build_export(ctx, &args);
2292 }
2293
2294 static unsigned ac_num_coords(enum ac_image_dim dim)
2295 {
2296 switch (dim) {
2297 case ac_image_1d:
2298 return 1;
2299 case ac_image_2d:
2300 case ac_image_1darray:
2301 return 2;
2302 case ac_image_3d:
2303 case ac_image_cube:
2304 case ac_image_2darray:
2305 case ac_image_2dmsaa:
2306 return 3;
2307 case ac_image_2darraymsaa:
2308 return 4;
2309 default:
2310 unreachable("ac_num_coords: bad dim");
2311 }
2312 }
2313
2314 static unsigned ac_num_derivs(enum ac_image_dim dim)
2315 {
2316 switch (dim) {
2317 case ac_image_1d:
2318 case ac_image_1darray:
2319 return 2;
2320 case ac_image_2d:
2321 case ac_image_2darray:
2322 case ac_image_cube:
2323 return 4;
2324 case ac_image_3d:
2325 return 6;
2326 case ac_image_2dmsaa:
2327 case ac_image_2darraymsaa:
2328 default:
2329 unreachable("derivatives not supported");
2330 }
2331 }
2332
2333 static const char *get_atomic_name(enum ac_atomic_op op)
2334 {
2335 switch (op) {
2336 case ac_atomic_swap: return "swap";
2337 case ac_atomic_add: return "add";
2338 case ac_atomic_sub: return "sub";
2339 case ac_atomic_smin: return "smin";
2340 case ac_atomic_umin: return "umin";
2341 case ac_atomic_smax: return "smax";
2342 case ac_atomic_umax: return "umax";
2343 case ac_atomic_and: return "and";
2344 case ac_atomic_or: return "or";
2345 case ac_atomic_xor: return "xor";
2346 case ac_atomic_inc_wrap: return "inc";
2347 case ac_atomic_dec_wrap: return "dec";
2348 }
2349 unreachable("bad atomic op");
2350 }
2351
2352 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
2353 struct ac_image_args *a)
2354 {
2355 const char *overload[3] = { "", "", "" };
2356 unsigned num_overloads = 0;
2357 LLVMValueRef args[18];
2358 unsigned num_args = 0;
2359 enum ac_image_dim dim = a->dim;
2360
2361 assert(!a->lod || a->lod == ctx->i32_0 || a->lod == ctx->f32_0 ||
2362 !a->level_zero);
2363 assert((a->opcode != ac_image_get_resinfo && a->opcode != ac_image_load_mip &&
2364 a->opcode != ac_image_store_mip) ||
2365 a->lod);
2366 assert(a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
2367 (!a->compare && !a->offset));
2368 assert((a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
2369 a->opcode == ac_image_get_lod) ||
2370 !a->bias);
2371 assert((a->bias ? 1 : 0) +
2372 (a->lod ? 1 : 0) +
2373 (a->level_zero ? 1 : 0) +
2374 (a->derivs[0] ? 1 : 0) <= 1);
2375 assert((a->min_lod ? 1 : 0) +
2376 (a->lod ? 1 : 0) +
2377 (a->level_zero ? 1 : 0) <= 1);
2378
2379 if (a->opcode == ac_image_get_lod) {
2380 switch (dim) {
2381 case ac_image_1darray:
2382 dim = ac_image_1d;
2383 break;
2384 case ac_image_2darray:
2385 case ac_image_cube:
2386 dim = ac_image_2d;
2387 break;
2388 default:
2389 break;
2390 }
2391 }
2392
2393 bool sample = a->opcode == ac_image_sample ||
2394 a->opcode == ac_image_gather4 ||
2395 a->opcode == ac_image_get_lod;
2396 bool atomic = a->opcode == ac_image_atomic ||
2397 a->opcode == ac_image_atomic_cmpswap;
2398 bool load = a->opcode == ac_image_sample ||
2399 a->opcode == ac_image_gather4 ||
2400 a->opcode == ac_image_load ||
2401 a->opcode == ac_image_load_mip;
2402 LLVMTypeRef coord_type = sample ? ctx->f32 : ctx->i32;
2403
2404 if (atomic || a->opcode == ac_image_store || a->opcode == ac_image_store_mip) {
2405 args[num_args++] = a->data[0];
2406 if (a->opcode == ac_image_atomic_cmpswap)
2407 args[num_args++] = a->data[1];
2408 }
2409
2410 if (!atomic)
2411 args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, false);
2412
2413 if (a->offset)
2414 args[num_args++] = ac_to_integer(ctx, a->offset);
2415 if (a->bias) {
2416 args[num_args++] = ac_to_float(ctx, a->bias);
2417 overload[num_overloads++] = ".f32";
2418 }
2419 if (a->compare)
2420 args[num_args++] = ac_to_float(ctx, a->compare);
2421 if (a->derivs[0]) {
2422 unsigned count = ac_num_derivs(dim);
2423 for (unsigned i = 0; i < count; ++i)
2424 args[num_args++] = ac_to_float(ctx, a->derivs[i]);
2425 overload[num_overloads++] = ".f32";
2426 }
2427 unsigned num_coords =
2428 a->opcode != ac_image_get_resinfo ? ac_num_coords(dim) : 0;
2429 for (unsigned i = 0; i < num_coords; ++i)
2430 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->coords[i], coord_type, "");
2431 if (a->lod)
2432 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->lod, coord_type, "");
2433 if (a->min_lod)
2434 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->min_lod, coord_type, "");
2435
2436 overload[num_overloads++] = sample ? ".f32" : ".i32";
2437
2438 args[num_args++] = a->resource;
2439 if (sample) {
2440 args[num_args++] = a->sampler;
2441 args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, false);
2442 }
2443
2444 args[num_args++] = ctx->i32_0; /* texfailctrl */
2445 args[num_args++] = LLVMConstInt(ctx->i32,
2446 load ? get_load_cache_policy(ctx, a->cache_policy) :
2447 a->cache_policy, false);
2448
2449 const char *name;
2450 const char *atomic_subop = "";
2451 switch (a->opcode) {
2452 case ac_image_sample: name = "sample"; break;
2453 case ac_image_gather4: name = "gather4"; break;
2454 case ac_image_load: name = "load"; break;
2455 case ac_image_load_mip: name = "load.mip"; break;
2456 case ac_image_store: name = "store"; break;
2457 case ac_image_store_mip: name = "store.mip"; break;
2458 case ac_image_atomic:
2459 name = "atomic.";
2460 atomic_subop = get_atomic_name(a->atomic);
2461 break;
2462 case ac_image_atomic_cmpswap:
2463 name = "atomic.";
2464 atomic_subop = "cmpswap";
2465 break;
2466 case ac_image_get_lod: name = "getlod"; break;
2467 case ac_image_get_resinfo: name = "getresinfo"; break;
2468 default: unreachable("invalid image opcode");
2469 }
2470
2471 const char *dimname;
2472 switch (dim) {
2473 case ac_image_1d: dimname = "1d"; break;
2474 case ac_image_2d: dimname = "2d"; break;
2475 case ac_image_3d: dimname = "3d"; break;
2476 case ac_image_cube: dimname = "cube"; break;
2477 case ac_image_1darray: dimname = "1darray"; break;
2478 case ac_image_2darray: dimname = "2darray"; break;
2479 case ac_image_2dmsaa: dimname = "2dmsaa"; break;
2480 case ac_image_2darraymsaa: dimname = "2darraymsaa"; break;
2481 default: unreachable("invalid dim");
2482 }
2483
2484 bool lod_suffix =
2485 a->lod && (a->opcode == ac_image_sample || a->opcode == ac_image_gather4);
2486 char intr_name[96];
2487 snprintf(intr_name, sizeof(intr_name),
2488 "llvm.amdgcn.image.%s%s" /* base name */
2489 "%s%s%s%s" /* sample/gather modifiers */
2490 ".%s.%s%s%s%s", /* dimension and type overloads */
2491 name, atomic_subop,
2492 a->compare ? ".c" : "",
2493 a->bias ? ".b" :
2494 lod_suffix ? ".l" :
2495 a->derivs[0] ? ".d" :
2496 a->level_zero ? ".lz" : "",
2497 a->min_lod ? ".cl" : "",
2498 a->offset ? ".o" : "",
2499 dimname,
2500 atomic ? "i32" : "v4f32",
2501 overload[0], overload[1], overload[2]);
2502
2503 LLVMTypeRef retty;
2504 if (atomic)
2505 retty = ctx->i32;
2506 else if (a->opcode == ac_image_store || a->opcode == ac_image_store_mip)
2507 retty = ctx->voidt;
2508 else
2509 retty = ctx->v4f32;
2510
2511 LLVMValueRef result =
2512 ac_build_intrinsic(ctx, intr_name, retty, args, num_args,
2513 a->attributes);
2514 if (!sample && retty == ctx->v4f32) {
2515 result = LLVMBuildBitCast(ctx->builder, result,
2516 ctx->v4i32, "");
2517 }
2518 return result;
2519 }
2520
2521 LLVMValueRef ac_build_image_get_sample_count(struct ac_llvm_context *ctx,
2522 LLVMValueRef rsrc)
2523 {
2524 LLVMValueRef samples;
2525
2526 /* Read the samples from the descriptor directly.
2527 * Hardware doesn't have any instruction for this.
2528 */
2529 samples = LLVMBuildExtractElement(ctx->builder, rsrc,
2530 LLVMConstInt(ctx->i32, 3, 0), "");
2531 samples = LLVMBuildLShr(ctx->builder, samples,
2532 LLVMConstInt(ctx->i32, 16, 0), "");
2533 samples = LLVMBuildAnd(ctx->builder, samples,
2534 LLVMConstInt(ctx->i32, 0xf, 0), "");
2535 samples = LLVMBuildShl(ctx->builder, ctx->i32_1,
2536 samples, "");
2537 return samples;
2538 }
2539
2540 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
2541 LLVMValueRef args[2])
2542 {
2543 return ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz", ctx->v2f16,
2544 args, 2, AC_FUNC_ATTR_READNONE);
2545 }
2546
2547 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
2548 LLVMValueRef args[2])
2549 {
2550 LLVMValueRef res =
2551 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16",
2552 ctx->v2i16, args, 2,
2553 AC_FUNC_ATTR_READNONE);
2554 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2555 }
2556
2557 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
2558 LLVMValueRef args[2])
2559 {
2560 LLVMValueRef res =
2561 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16",
2562 ctx->v2i16, args, 2,
2563 AC_FUNC_ATTR_READNONE);
2564 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2565 }
2566
2567 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2568 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
2569 LLVMValueRef args[2], unsigned bits, bool hi)
2570 {
2571 assert(bits == 8 || bits == 10 || bits == 16);
2572
2573 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2574 bits == 8 ? 127 : bits == 10 ? 511 : 32767, 0);
2575 LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
2576 bits == 8 ? -128 : bits == 10 ? -512 : -32768, 0);
2577 LLVMValueRef max_alpha =
2578 bits != 10 ? max_rgb : ctx->i32_1;
2579 LLVMValueRef min_alpha =
2580 bits != 10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
2581
2582 /* Clamp. */
2583 if (bits != 16) {
2584 for (int i = 0; i < 2; i++) {
2585 bool alpha = hi && i == 1;
2586 args[i] = ac_build_imin(ctx, args[i],
2587 alpha ? max_alpha : max_rgb);
2588 args[i] = ac_build_imax(ctx, args[i],
2589 alpha ? min_alpha : min_rgb);
2590 }
2591 }
2592
2593 LLVMValueRef res =
2594 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16",
2595 ctx->v2i16, args, 2,
2596 AC_FUNC_ATTR_READNONE);
2597 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2598 }
2599
2600 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2601 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
2602 LLVMValueRef args[2], unsigned bits, bool hi)
2603 {
2604 assert(bits == 8 || bits == 10 || bits == 16);
2605
2606 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2607 bits == 8 ? 255 : bits == 10 ? 1023 : 65535, 0);
2608 LLVMValueRef max_alpha =
2609 bits != 10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
2610
2611 /* Clamp. */
2612 if (bits != 16) {
2613 for (int i = 0; i < 2; i++) {
2614 bool alpha = hi && i == 1;
2615 args[i] = ac_build_umin(ctx, args[i],
2616 alpha ? max_alpha : max_rgb);
2617 }
2618 }
2619
2620 LLVMValueRef res =
2621 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16",
2622 ctx->v2i16, args, 2,
2623 AC_FUNC_ATTR_READNONE);
2624 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2625 }
2626
2627 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1)
2628 {
2629 return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1,
2630 &i1, 1, AC_FUNC_ATTR_READNONE);
2631 }
2632
2633 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1)
2634 {
2635 ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt,
2636 &i1, 1, 0);
2637 }
2638
2639 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
2640 LLVMValueRef offset, LLVMValueRef width,
2641 bool is_signed)
2642 {
2643 LLVMValueRef args[] = {
2644 input,
2645 offset,
2646 width,
2647 };
2648
2649 return ac_build_intrinsic(ctx, is_signed ? "llvm.amdgcn.sbfe.i32" :
2650 "llvm.amdgcn.ubfe.i32",
2651 ctx->i32, args, 3, AC_FUNC_ATTR_READNONE);
2652
2653 }
2654
2655 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2656 LLVMValueRef s1, LLVMValueRef s2)
2657 {
2658 return LLVMBuildAdd(ctx->builder,
2659 LLVMBuildMul(ctx->builder, s0, s1, ""), s2, "");
2660 }
2661
2662 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2663 LLVMValueRef s1, LLVMValueRef s2)
2664 {
2665 /* FMA is better on GFX10, because it has FMA units instead of MUL-ADD units. */
2666 if (ctx->chip_class >= GFX10) {
2667 return ac_build_intrinsic(ctx, "llvm.fma.f32", ctx->f32,
2668 (LLVMValueRef []) {s0, s1, s2}, 3,
2669 AC_FUNC_ATTR_READNONE);
2670 }
2671
2672 return LLVMBuildFAdd(ctx->builder,
2673 LLVMBuildFMul(ctx->builder, s0, s1, ""), s2, "");
2674 }
2675
2676 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags)
2677 {
2678 if (!wait_flags)
2679 return;
2680
2681 unsigned lgkmcnt = 63;
2682 unsigned vmcnt = ctx->chip_class >= GFX9 ? 63 : 15;
2683 unsigned vscnt = 63;
2684
2685 if (wait_flags & AC_WAIT_LGKM)
2686 lgkmcnt = 0;
2687 if (wait_flags & AC_WAIT_VLOAD)
2688 vmcnt = 0;
2689
2690 if (wait_flags & AC_WAIT_VSTORE) {
2691 if (ctx->chip_class >= GFX10)
2692 vscnt = 0;
2693 else
2694 vmcnt = 0;
2695 }
2696
2697 /* There is no intrinsic for vscnt(0), so use a fence. */
2698 if ((wait_flags & AC_WAIT_LGKM &&
2699 wait_flags & AC_WAIT_VLOAD &&
2700 wait_flags & AC_WAIT_VSTORE) ||
2701 vscnt == 0) {
2702 LLVMBuildFence(ctx->builder, LLVMAtomicOrderingRelease, false, "");
2703 return;
2704 }
2705
2706 unsigned simm16 = (lgkmcnt << 8) |
2707 (7 << 4) | /* expcnt */
2708 (vmcnt & 0xf) |
2709 ((vmcnt >> 4) << 14);
2710
2711 LLVMValueRef args[1] = {
2712 LLVMConstInt(ctx->i32, simm16, false),
2713 };
2714 ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt",
2715 ctx->voidt, args, 1, 0);
2716 }
2717
2718 LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
2719 LLVMValueRef src1, LLVMValueRef src2,
2720 unsigned bitsize)
2721 {
2722 LLVMValueRef result;
2723
2724 if (bitsize == 64 || (bitsize == 16 && ctx->chip_class <= GFX8)) {
2725 /* Lower 64-bit fmed because LLVM doesn't expose an intrinsic,
2726 * or lower 16-bit fmed because it's only supported on GFX9+.
2727 */
2728 LLVMValueRef min1, min2, max1;
2729
2730 min1 = ac_build_fmin(ctx, src0, src1);
2731 max1 = ac_build_fmax(ctx, src0, src1);
2732 min2 = ac_build_fmin(ctx, max1, src2);
2733
2734 result = ac_build_fmax(ctx, min2, min1);
2735 } else {
2736 LLVMTypeRef type;
2737 char *intr;
2738
2739 if (bitsize == 16) {
2740 intr = "llvm.amdgcn.fmed3.f16";
2741 type = ctx->f16;
2742 } else {
2743 assert(bitsize == 32);
2744 intr = "llvm.amdgcn.fmed3.f32";
2745 type = ctx->f32;
2746 }
2747
2748 LLVMValueRef params[] = {
2749 src0,
2750 src1,
2751 src2,
2752 };
2753
2754 result = ac_build_intrinsic(ctx, intr, type, params, 3,
2755 AC_FUNC_ATTR_READNONE);
2756 }
2757
2758 if (ctx->chip_class < GFX9 && bitsize == 32) {
2759 /* Only pre-GFX9 chips do not flush denorms. */
2760 result = ac_build_canonicalize(ctx, result, bitsize);
2761 }
2762
2763 return result;
2764 }
2765
2766 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
2767 unsigned bitsize)
2768 {
2769 LLVMTypeRef type;
2770 char *intr;
2771
2772 if (bitsize == 16) {
2773 intr = "llvm.amdgcn.fract.f16";
2774 type = ctx->f16;
2775 } else if (bitsize == 32) {
2776 intr = "llvm.amdgcn.fract.f32";
2777 type = ctx->f32;
2778 } else {
2779 intr = "llvm.amdgcn.fract.f64";
2780 type = ctx->f64;
2781 }
2782
2783 LLVMValueRef params[] = {
2784 src0,
2785 };
2786 return ac_build_intrinsic(ctx, intr, type, params, 1,
2787 AC_FUNC_ATTR_READNONE);
2788 }
2789
2790 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2791 unsigned bitsize)
2792 {
2793 LLVMTypeRef type = LLVMIntTypeInContext(ctx->context, bitsize);
2794 LLVMValueRef zero = LLVMConstInt(type, 0, false);
2795 LLVMValueRef one = LLVMConstInt(type, 1, false);
2796
2797 LLVMValueRef cmp, val;
2798 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, src0, zero, "");
2799 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2800 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGE, val, zero, "");
2801 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstInt(type, -1, true), "");
2802 return val;
2803 }
2804
2805 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2806 unsigned bitsize)
2807 {
2808 LLVMValueRef cmp, val, zero, one;
2809 LLVMTypeRef type;
2810
2811 if (bitsize == 16) {
2812 type = ctx->f16;
2813 zero = ctx->f16_0;
2814 one = ctx->f16_1;
2815 } else if (bitsize == 32) {
2816 type = ctx->f32;
2817 zero = ctx->f32_0;
2818 one = ctx->f32_1;
2819 } else {
2820 type = ctx->f64;
2821 zero = ctx->f64_0;
2822 one = ctx->f64_1;
2823 }
2824
2825 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src0, zero, "");
2826 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2827 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGE, val, zero, "");
2828 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstReal(type, -1.0), "");
2829 return val;
2830 }
2831
2832 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0)
2833 {
2834 LLVMValueRef result;
2835 unsigned bitsize;
2836
2837 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2838
2839 switch (bitsize) {
2840 case 128:
2841 result = ac_build_intrinsic(ctx, "llvm.ctpop.i128", ctx->i128,
2842 (LLVMValueRef []) { src0 }, 1,
2843 AC_FUNC_ATTR_READNONE);
2844 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
2845 break;
2846 case 64:
2847 result = ac_build_intrinsic(ctx, "llvm.ctpop.i64", ctx->i64,
2848 (LLVMValueRef []) { src0 }, 1,
2849 AC_FUNC_ATTR_READNONE);
2850
2851 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
2852 break;
2853 case 32:
2854 result = ac_build_intrinsic(ctx, "llvm.ctpop.i32", ctx->i32,
2855 (LLVMValueRef []) { src0 }, 1,
2856 AC_FUNC_ATTR_READNONE);
2857 break;
2858 case 16:
2859 result = ac_build_intrinsic(ctx, "llvm.ctpop.i16", ctx->i16,
2860 (LLVMValueRef []) { src0 }, 1,
2861 AC_FUNC_ATTR_READNONE);
2862
2863 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2864 break;
2865 case 8:
2866 result = ac_build_intrinsic(ctx, "llvm.ctpop.i8", ctx->i8,
2867 (LLVMValueRef []) { src0 }, 1,
2868 AC_FUNC_ATTR_READNONE);
2869
2870 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2871 break;
2872 default:
2873 unreachable(!"invalid bitsize");
2874 break;
2875 }
2876
2877 return result;
2878 }
2879
2880 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
2881 LLVMValueRef src0)
2882 {
2883 LLVMValueRef result;
2884 unsigned bitsize;
2885
2886 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2887
2888 switch (bitsize) {
2889 case 64:
2890 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i64", ctx->i64,
2891 (LLVMValueRef []) { src0 }, 1,
2892 AC_FUNC_ATTR_READNONE);
2893
2894 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
2895 break;
2896 case 32:
2897 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i32", ctx->i32,
2898 (LLVMValueRef []) { src0 }, 1,
2899 AC_FUNC_ATTR_READNONE);
2900 break;
2901 case 16:
2902 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i16", ctx->i16,
2903 (LLVMValueRef []) { src0 }, 1,
2904 AC_FUNC_ATTR_READNONE);
2905
2906 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2907 break;
2908 case 8:
2909 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i8", ctx->i8,
2910 (LLVMValueRef []) { src0 }, 1,
2911 AC_FUNC_ATTR_READNONE);
2912
2913 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
2914 break;
2915 default:
2916 unreachable(!"invalid bitsize");
2917 break;
2918 }
2919
2920 return result;
2921 }
2922
2923 #define AC_EXP_TARGET 0
2924 #define AC_EXP_ENABLED_CHANNELS 1
2925 #define AC_EXP_OUT0 2
2926
2927 enum ac_ir_type {
2928 AC_IR_UNDEF,
2929 AC_IR_CONST,
2930 AC_IR_VALUE,
2931 };
2932
2933 struct ac_vs_exp_chan
2934 {
2935 LLVMValueRef value;
2936 float const_float;
2937 enum ac_ir_type type;
2938 };
2939
2940 struct ac_vs_exp_inst {
2941 unsigned offset;
2942 LLVMValueRef inst;
2943 struct ac_vs_exp_chan chan[4];
2944 };
2945
2946 struct ac_vs_exports {
2947 unsigned num;
2948 struct ac_vs_exp_inst exp[VARYING_SLOT_MAX];
2949 };
2950
2951 /* Return true if the PARAM export has been eliminated. */
2952 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset,
2953 uint32_t num_outputs,
2954 struct ac_vs_exp_inst *exp)
2955 {
2956 unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
2957 bool is_zero[4] = {}, is_one[4] = {};
2958
2959 for (i = 0; i < 4; i++) {
2960 /* It's a constant expression. Undef outputs are eliminated too. */
2961 if (exp->chan[i].type == AC_IR_UNDEF) {
2962 is_zero[i] = true;
2963 is_one[i] = true;
2964 } else if (exp->chan[i].type == AC_IR_CONST) {
2965 if (exp->chan[i].const_float == 0)
2966 is_zero[i] = true;
2967 else if (exp->chan[i].const_float == 1)
2968 is_one[i] = true;
2969 else
2970 return false; /* other constant */
2971 } else
2972 return false;
2973 }
2974
2975 /* Only certain combinations of 0 and 1 can be eliminated. */
2976 if (is_zero[0] && is_zero[1] && is_zero[2])
2977 default_val = is_zero[3] ? 0 : 1;
2978 else if (is_one[0] && is_one[1] && is_one[2])
2979 default_val = is_zero