projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
ac: allow control flow statements in NIR callbacks
[mesa.git] / src / amd / common / 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
30 #include "c11/threads.h"
31
32 #include <assert.h>
33 #include <stdio.h>
34
35 #include "ac_llvm_util.h"
36 #include "ac_exp_param.h"
37 #include "util/bitscan.h"
38 #include "util/macros.h"
39 #include "util/u_atomic.h"
40 #include "util/u_math.h"
41 #include "sid.h"
42
43 #include "shader_enums.h"
44
45 #define AC_LLVM_INITIAL_CF_DEPTH 4
46
47 /* Data for if/else/endif and bgnloop/endloop control flow structures.
48 */
49 struct ac_llvm_flow {
50 /* Loop exit or next part of if/else/endif. */
51 LLVMBasicBlockRef next_block;
52 LLVMBasicBlockRef loop_entry_block;
53 };
54
55 /* Initialize module-independent parts of the context.
56 *
57 * The caller is responsible for initializing ctx::module and ctx::builder.
58 */
59 void
60 ac_llvm_context_init(struct ac_llvm_context *ctx,
61 struct ac_llvm_compiler *compiler,
62 enum chip_class chip_class, enum radeon_family family,
63 enum ac_float_mode float_mode, unsigned wave_size)
64 {
65 LLVMValueRef args[1];
66
67 ctx->context = LLVMContextCreate();
68
69 ctx->chip_class = chip_class;
70 ctx->family = family;
71 ctx->wave_size = wave_size;
72 ctx->module = ac_create_module(wave_size == 32 ? compiler->tm_wave32
73 : compiler->tm,
74 ctx->context);
75 ctx->builder = ac_create_builder(ctx->context, float_mode);
76
77 ctx->voidt = LLVMVoidTypeInContext(ctx->context);
78 ctx->i1 = LLVMInt1TypeInContext(ctx->context);
79 ctx->i8 = LLVMInt8TypeInContext(ctx->context);
80 ctx->i16 = LLVMIntTypeInContext(ctx->context, 16);
81 ctx->i32 = LLVMIntTypeInContext(ctx->context, 32);
82 ctx->i64 = LLVMIntTypeInContext(ctx->context, 64);
83 ctx->intptr = ctx->i32;
84 ctx->f16 = LLVMHalfTypeInContext(ctx->context);
85 ctx->f32 = LLVMFloatTypeInContext(ctx->context);
86 ctx->f64 = LLVMDoubleTypeInContext(ctx->context);
87 ctx->v2i16 = LLVMVectorType(ctx->i16, 2);
88 ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
89 ctx->v3i32 = LLVMVectorType(ctx->i32, 3);
90 ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
91 ctx->v2f32 = LLVMVectorType(ctx->f32, 2);
92 ctx->v3f32 = LLVMVectorType(ctx->f32, 3);
93 ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
94 ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
95 ctx->iN_wavemask = LLVMIntTypeInContext(ctx->context, ctx->wave_size);
96
97 ctx->i8_0 = LLVMConstInt(ctx->i8, 0, false);
98 ctx->i8_1 = LLVMConstInt(ctx->i8, 1, false);
99 ctx->i16_0 = LLVMConstInt(ctx->i16, 0, false);
100 ctx->i16_1 = LLVMConstInt(ctx->i16, 1, false);
101 ctx->i32_0 = LLVMConstInt(ctx->i32, 0, false);
102 ctx->i32_1 = LLVMConstInt(ctx->i32, 1, false);
103 ctx->i64_0 = LLVMConstInt(ctx->i64, 0, false);
104 ctx->i64_1 = LLVMConstInt(ctx->i64, 1, false);
105 ctx->f16_0 = LLVMConstReal(ctx->f16, 0.0);
106 ctx->f16_1 = LLVMConstReal(ctx->f16, 1.0);
107 ctx->f32_0 = LLVMConstReal(ctx->f32, 0.0);
108 ctx->f32_1 = LLVMConstReal(ctx->f32, 1.0);
109 ctx->f64_0 = LLVMConstReal(ctx->f64, 0.0);
110 ctx->f64_1 = LLVMConstReal(ctx->f64, 1.0);
111
112 ctx->i1false = LLVMConstInt(ctx->i1, 0, false);
113 ctx->i1true = LLVMConstInt(ctx->i1, 1, false);
114
115 ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context,
116 "range", 5);
117
118 ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context,
119 "invariant.load", 14);
120
121 ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->context, "fpmath", 6);
122
123 args[0] = LLVMConstReal(ctx->f32, 2.5);
124 ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->context, args, 1);
125
126 ctx->uniform_md_kind = LLVMGetMDKindIDInContext(ctx->context,
127 "amdgpu.uniform", 14);
128
129 ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
130 ctx->flow = calloc(1, sizeof(*ctx->flow));
131 }
132
133 void
134 ac_llvm_context_dispose(struct ac_llvm_context *ctx)
135 {
136 free(ctx->flow->stack);
137 free(ctx->flow);
138 ctx->flow = NULL;
139 }
140
141 int
142 ac_get_llvm_num_components(LLVMValueRef value)
143 {
144 LLVMTypeRef type = LLVMTypeOf(value);
145 unsigned num_components = LLVMGetTypeKind(type) == LLVMVectorTypeKind
146 ? LLVMGetVectorSize(type)
147 : 1;
148 return num_components;
149 }
150
151 LLVMValueRef
152 ac_llvm_extract_elem(struct ac_llvm_context *ac,
153 LLVMValueRef value,
154 int index)
155 {
156 if (LLVMGetTypeKind(LLVMTypeOf(value)) != LLVMVectorTypeKind) {
157 assert(index == 0);
158 return value;
159 }
160
161 return LLVMBuildExtractElement(ac->builder, value,
162 LLVMConstInt(ac->i32, index, false), "");
163 }
164
165 int
166 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type)
167 {
168 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
169 type = LLVMGetElementType(type);
170
171 if (LLVMGetTypeKind(type) == LLVMIntegerTypeKind)
172 return LLVMGetIntTypeWidth(type);
173
174 if (type == ctx->f16)
175 return 16;
176 if (type == ctx->f32)
177 return 32;
178 if (type == ctx->f64)
179 return 64;
180
181 unreachable("Unhandled type kind in get_elem_bits");
182 }
183
184 unsigned
185 ac_get_type_size(LLVMTypeRef type)
186 {
187 LLVMTypeKind kind = LLVMGetTypeKind(type);
188
189 switch (kind) {
190 case LLVMIntegerTypeKind:
191 return LLVMGetIntTypeWidth(type) / 8;
192 case LLVMHalfTypeKind:
193 return 2;
194 case LLVMFloatTypeKind:
195 return 4;
196 case LLVMDoubleTypeKind:
197 return 8;
198 case LLVMPointerTypeKind:
199 if (LLVMGetPointerAddressSpace(type) == AC_ADDR_SPACE_CONST_32BIT)
200 return 4;
201 return 8;
202 case LLVMVectorTypeKind:
203 return LLVMGetVectorSize(type) *
204 ac_get_type_size(LLVMGetElementType(type));
205 case LLVMArrayTypeKind:
206 return LLVMGetArrayLength(type) *
207 ac_get_type_size(LLVMGetElementType(type));
208 default:
209 assert(0);
210 return 0;
211 }
212 }
213
214 static LLVMTypeRef to_integer_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
215 {
216 if (t == ctx->i8)
217 return ctx->i8;
218 else if (t == ctx->f16 || t == ctx->i16)
219 return ctx->i16;
220 else if (t == ctx->f32 || t == ctx->i32)
221 return ctx->i32;
222 else if (t == ctx->f64 || t == ctx->i64)
223 return ctx->i64;
224 else
225 unreachable("Unhandled integer size");
226 }
227
228 LLVMTypeRef
229 ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
230 {
231 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
232 LLVMTypeRef elem_type = LLVMGetElementType(t);
233 return LLVMVectorType(to_integer_type_scalar(ctx, elem_type),
234 LLVMGetVectorSize(t));
235 }
236 if (LLVMGetTypeKind(t) == LLVMPointerTypeKind) {
237 switch (LLVMGetPointerAddressSpace(t)) {
238 case AC_ADDR_SPACE_GLOBAL:
239 return ctx->i64;
240 case AC_ADDR_SPACE_LDS:
241 return ctx->i32;
242 default:
243 unreachable("unhandled address space");
244 }
245 }
246 return to_integer_type_scalar(ctx, t);
247 }
248
249 LLVMValueRef
250 ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v)
251 {
252 LLVMTypeRef type = LLVMTypeOf(v);
253 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
254 return LLVMBuildPtrToInt(ctx->builder, v, ac_to_integer_type(ctx, type), "");
255 }
256 return LLVMBuildBitCast(ctx->builder, v, ac_to_integer_type(ctx, type), "");
257 }
258
259 LLVMValueRef
260 ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v)
261 {
262 LLVMTypeRef type = LLVMTypeOf(v);
263 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind)
264 return v;
265 return ac_to_integer(ctx, v);
266 }
267
268 static LLVMTypeRef to_float_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
269 {
270 if (t == ctx->i8)
271 return ctx->i8;
272 else if (t == ctx->i16 || t == ctx->f16)
273 return ctx->f16;
274 else if (t == ctx->i32 || t == ctx->f32)
275 return ctx->f32;
276 else if (t == ctx->i64 || t == ctx->f64)
277 return ctx->f64;
278 else
279 unreachable("Unhandled float size");
280 }
281
282 LLVMTypeRef
283 ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
284 {
285 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
286 LLVMTypeRef elem_type = LLVMGetElementType(t);
287 return LLVMVectorType(to_float_type_scalar(ctx, elem_type),
288 LLVMGetVectorSize(t));
289 }
290 return to_float_type_scalar(ctx, t);
291 }
292
293 LLVMValueRef
294 ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v)
295 {
296 LLVMTypeRef type = LLVMTypeOf(v);
297 return LLVMBuildBitCast(ctx->builder, v, ac_to_float_type(ctx, type), "");
298 }
299
300
301 LLVMValueRef
302 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
303 LLVMTypeRef return_type, LLVMValueRef *params,
304 unsigned param_count, unsigned attrib_mask)
305 {
306 LLVMValueRef function, call;
307 bool set_callsite_attrs = !(attrib_mask & AC_FUNC_ATTR_LEGACY);
308
309 function = LLVMGetNamedFunction(ctx->module, name);
310 if (!function) {
311 LLVMTypeRef param_types[32], function_type;
312 unsigned i;
313
314 assert(param_count <= 32);
315
316 for (i = 0; i < param_count; ++i) {
317 assert(params[i]);
318 param_types[i] = LLVMTypeOf(params[i]);
319 }
320 function_type =
321 LLVMFunctionType(return_type, param_types, param_count, 0);
322 function = LLVMAddFunction(ctx->module, name, function_type);
323
324 LLVMSetFunctionCallConv(function, LLVMCCallConv);
325 LLVMSetLinkage(function, LLVMExternalLinkage);
326
327 if (!set_callsite_attrs)
328 ac_add_func_attributes(ctx->context, function, attrib_mask);
329 }
330
331 call = LLVMBuildCall(ctx->builder, function, params, param_count, "");
332 if (set_callsite_attrs)
333 ac_add_func_attributes(ctx->context, call, attrib_mask);
334 return call;
335 }
336
337 /**
338 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
339 * intrinsic names).
340 */
341 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize)
342 {
343 LLVMTypeRef elem_type = type;
344
345 assert(bufsize >= 8);
346
347 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
348 int ret = snprintf(buf, bufsize, "v%u",
349 LLVMGetVectorSize(type));
350 if (ret < 0) {
351 char *type_name = LLVMPrintTypeToString(type);
352 fprintf(stderr, "Error building type name for: %s\n",
353 type_name);
354 return;
355 }
356 elem_type = LLVMGetElementType(type);
357 buf += ret;
358 bufsize -= ret;
359 }
360 switch (LLVMGetTypeKind(elem_type)) {
361 default: break;
362 case LLVMIntegerTypeKind:
363 snprintf(buf, bufsize, "i%d", LLVMGetIntTypeWidth(elem_type));
364 break;
365 case LLVMHalfTypeKind:
366 snprintf(buf, bufsize, "f16");
367 break;
368 case LLVMFloatTypeKind:
369 snprintf(buf, bufsize, "f32");
370 break;
371 case LLVMDoubleTypeKind:
372 snprintf(buf, bufsize, "f64");
373 break;
374 }
375 }
376
377 /**
378 * Helper function that builds an LLVM IR PHI node and immediately adds
379 * incoming edges.
380 */
381 LLVMValueRef
382 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
383 unsigned count_incoming, LLVMValueRef *values,
384 LLVMBasicBlockRef *blocks)
385 {
386 LLVMValueRef phi = LLVMBuildPhi(ctx->builder, type, "");
387 LLVMAddIncoming(phi, values, blocks, count_incoming);
388 return phi;
389 }
390
391 void ac_build_s_barrier(struct ac_llvm_context *ctx)
392 {
393 ac_build_intrinsic(ctx, "llvm.amdgcn.s.barrier", ctx->voidt, NULL,
394 0, AC_FUNC_ATTR_CONVERGENT);
395 }
396
397 /* Prevent optimizations (at least of memory accesses) across the current
398 * point in the program by emitting empty inline assembly that is marked as
399 * having side effects.
400 *
401 * Optionally, a value can be passed through the inline assembly to prevent
402 * LLVM from hoisting calls to ReadNone functions.
403 */
404 void
405 ac_build_optimization_barrier(struct ac_llvm_context *ctx,
406 LLVMValueRef *pvgpr)
407 {
408 static int counter = 0;
409
410 LLVMBuilderRef builder = ctx->builder;
411 char code[16];
412
413 snprintf(code, sizeof(code), "; %d", p_atomic_inc_return(&counter));
414
415 if (!pvgpr) {
416 LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
417 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "", true, false);
418 LLVMBuildCall(builder, inlineasm, NULL, 0, "");
419 } else {
420 LLVMTypeRef ftype = LLVMFunctionType(ctx->i32, &ctx->i32, 1, false);
421 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "=v,0", true, false);
422 LLVMValueRef vgpr = *pvgpr;
423 LLVMTypeRef vgpr_type = LLVMTypeOf(vgpr);
424 unsigned vgpr_size = ac_get_type_size(vgpr_type);
425 LLVMValueRef vgpr0;
426
427 assert(vgpr_size % 4 == 0);
428
429 vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
430 vgpr0 = LLVMBuildExtractElement(builder, vgpr, ctx->i32_0, "");
431 vgpr0 = LLVMBuildCall(builder, inlineasm, &vgpr0, 1, "");
432 vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
433 vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
434
435 *pvgpr = vgpr;
436 }
437 }
438
439 LLVMValueRef
440 ac_build_shader_clock(struct ac_llvm_context *ctx)
441 {
442 const char *intr = HAVE_LLVM >= 0x0900 && ctx->chip_class >= GFX8 ?
443 "llvm.amdgcn.s.memrealtime" : "llvm.readcyclecounter";
444 LLVMValueRef tmp = ac_build_intrinsic(ctx, intr, ctx->i64, NULL, 0, 0);
445 return LLVMBuildBitCast(ctx->builder, tmp, ctx->v2i32, "");
446 }
447
448 LLVMValueRef
449 ac_build_ballot(struct ac_llvm_context *ctx,
450 LLVMValueRef value)
451 {
452 const char *name;
453
454 if (HAVE_LLVM >= 0x900) {
455 if (ctx->wave_size == 64)
456 name = "llvm.amdgcn.icmp.i64.i32";
457 else
458 name = "llvm.amdgcn.icmp.i32.i32";
459 } else {
460 name = "llvm.amdgcn.icmp.i32";
461 }
462 LLVMValueRef args[3] = {
463 value,
464 ctx->i32_0,
465 LLVMConstInt(ctx->i32, LLVMIntNE, 0)
466 };
467
468 /* We currently have no other way to prevent LLVM from lifting the icmp
469 * calls to a dominating basic block.
470 */
471 ac_build_optimization_barrier(ctx, &args[0]);
472
473 args[0] = ac_to_integer(ctx, args[0]);
474
475 return ac_build_intrinsic(ctx, name, ctx->iN_wavemask, args, 3,
476 AC_FUNC_ATTR_NOUNWIND |
477 AC_FUNC_ATTR_READNONE |
478 AC_FUNC_ATTR_CONVERGENT);
479 }
480
481 LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
482 LLVMValueRef value)
483 {
484 const char *name = HAVE_LLVM >= 0x900 ? "llvm.amdgcn.icmp.i64.i1" : "llvm.amdgcn.icmp.i1";
485 LLVMValueRef args[3] = {
486 value,
487 ctx->i1false,
488 LLVMConstInt(ctx->i32, LLVMIntNE, 0),
489 };
490
491 assert(HAVE_LLVM >= 0x0800);
492 return ac_build_intrinsic(ctx, name, ctx->i64, args, 3,
493 AC_FUNC_ATTR_NOUNWIND |
494 AC_FUNC_ATTR_READNONE |
495 AC_FUNC_ATTR_CONVERGENT);
496 }
497
498 LLVMValueRef
499 ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value)
500 {
501 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
502 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
503 return LLVMBuildICmp(ctx->builder, LLVMIntEQ, vote_set, active_set, "");
504 }
505
506 LLVMValueRef
507 ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value)
508 {
509 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
510 return LLVMBuildICmp(ctx->builder, LLVMIntNE, vote_set,
511 LLVMConstInt(ctx->iN_wavemask, 0, 0), "");
512 }
513
514 LLVMValueRef
515 ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value)
516 {
517 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
518 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
519
520 LLVMValueRef all = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
521 vote_set, active_set, "");
522 LLVMValueRef none = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
523 vote_set,
524 LLVMConstInt(ctx->iN_wavemask, 0, 0), "");
525 return LLVMBuildOr(ctx->builder, all, none, "");
526 }
527
528 LLVMValueRef
529 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
530 unsigned value_count, unsigned component)
531 {
532 LLVMValueRef vec = NULL;
533
534 if (value_count == 1) {
535 return values[component];
536 } else if (!value_count)
537 unreachable("value_count is 0");
538
539 for (unsigned i = component; i < value_count + component; i++) {
540 LLVMValueRef value = values[i];
541
542 if (i == component)
543 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
544 LLVMValueRef index = LLVMConstInt(ctx->i32, i - component, false);
545 vec = LLVMBuildInsertElement(ctx->builder, vec, value, index, "");
546 }
547 return vec;
548 }
549
550 LLVMValueRef
551 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
552 LLVMValueRef *values,
553 unsigned value_count,
554 unsigned value_stride,
555 bool load,
556 bool always_vector)
557 {
558 LLVMBuilderRef builder = ctx->builder;
559 LLVMValueRef vec = NULL;
560 unsigned i;
561
562 if (value_count == 1 && !always_vector) {
563 if (load)
564 return LLVMBuildLoad(builder, values[0], "");
565 return values[0];
566 } else if (!value_count)
567 unreachable("value_count is 0");
568
569 for (i = 0; i < value_count; i++) {
570 LLVMValueRef value = values[i * value_stride];
571 if (load)
572 value = LLVMBuildLoad(builder, value, "");
573
574 if (!i)
575 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
576 LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
577 vec = LLVMBuildInsertElement(builder, vec, value, index, "");
578 }
579 return vec;
580 }
581
582 LLVMValueRef
583 ac_build_gather_values(struct ac_llvm_context *ctx,
584 LLVMValueRef *values,
585 unsigned value_count)
586 {
587 return ac_build_gather_values_extended(ctx, values, value_count, 1, false, false);
588 }
589
590 /* Expand a scalar or vector to <dst_channels x type> by filling the remaining
591 * channels with undef. Extract at most src_channels components from the input.
592 */
593 static LLVMValueRef
594 ac_build_expand(struct ac_llvm_context *ctx,
595 LLVMValueRef value,
596 unsigned src_channels,
597 unsigned dst_channels)
598 {
599 LLVMTypeRef elemtype;
600 LLVMValueRef chan[dst_channels];
601
602 if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMVectorTypeKind) {
603 unsigned vec_size = LLVMGetVectorSize(LLVMTypeOf(value));
604
605 if (src_channels == dst_channels && vec_size == dst_channels)
606 return value;
607
608 src_channels = MIN2(src_channels, vec_size);
609
610 for (unsigned i = 0; i < src_channels; i++)
611 chan[i] = ac_llvm_extract_elem(ctx, value, i);
612
613 elemtype = LLVMGetElementType(LLVMTypeOf(value));
614 } else {
615 if (src_channels) {
616 assert(src_channels == 1);
617 chan[0] = value;
618 }
619 elemtype = LLVMTypeOf(value);
620 }
621
622 for (unsigned i = src_channels; i < dst_channels; i++)
623 chan[i] = LLVMGetUndef(elemtype);
624
625 return ac_build_gather_values(ctx, chan, dst_channels);
626 }
627
628 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
629 * with undef. Extract at most num_channels components from the input.
630 */
631 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
632 LLVMValueRef value,
633 unsigned num_channels)
634 {
635 return ac_build_expand(ctx, value, num_channels, 4);
636 }
637
638 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value)
639 {
640 unsigned type_size = ac_get_type_size(LLVMTypeOf(value));
641 const char *name;
642
643 if (type_size == 2)
644 name = "llvm.rint.f16";
645 else if (type_size == 4)
646 name = "llvm.rint.f32";
647 else
648 name = "llvm.rint.f64";
649
650 return ac_build_intrinsic(ctx, name, LLVMTypeOf(value), &value, 1,
651 AC_FUNC_ATTR_READNONE);
652 }
653
654 LLVMValueRef
655 ac_build_fdiv(struct ac_llvm_context *ctx,
656 LLVMValueRef num,
657 LLVMValueRef den)
658 {
659 /* If we do (num / den), LLVM >= 7.0 does:
660 * return num * v_rcp_f32(den * (fabs(den) > 0x1.0p+96f ? 0x1.0p-32f : 1.0f));
661 *
662 * If we do (num * (1 / den)), LLVM does:
663 * return num * v_rcp_f32(den);
664 */
665 LLVMValueRef one = LLVMConstReal(LLVMTypeOf(num), 1.0);
666 LLVMValueRef rcp = LLVMBuildFDiv(ctx->builder, one, den, "");
667 LLVMValueRef ret = LLVMBuildFMul(ctx->builder, num, rcp, "");
668
669 /* Use v_rcp_f32 instead of precise division. */
670 if (!LLVMIsConstant(ret))
671 LLVMSetMetadata(ret, ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
672 return ret;
673 }
674
675 /* See fast_idiv_by_const.h. */
676 /* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
677 LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
678 LLVMValueRef num,
679 LLVMValueRef multiplier,
680 LLVMValueRef pre_shift,
681 LLVMValueRef post_shift,
682 LLVMValueRef increment)
683 {
684 LLVMBuilderRef builder = ctx->builder;
685
686 num = LLVMBuildLShr(builder, num, pre_shift, "");
687 num = LLVMBuildMul(builder,
688 LLVMBuildZExt(builder, num, ctx->i64, ""),
689 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
690 num = LLVMBuildAdd(builder, num,
691 LLVMBuildZExt(builder, increment, ctx->i64, ""), "");
692 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
693 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
694 return LLVMBuildLShr(builder, num, post_shift, "");
695 }
696
697 /* See fast_idiv_by_const.h. */
698 /* If num != UINT_MAX, this more efficient version can be used. */
699 /* Set: increment = util_fast_udiv_info::increment; */
700 LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
701 LLVMValueRef num,
702 LLVMValueRef multiplier,
703 LLVMValueRef pre_shift,
704 LLVMValueRef post_shift,
705 LLVMValueRef increment)
706 {
707 LLVMBuilderRef builder = ctx->builder;
708
709 num = LLVMBuildLShr(builder, num, pre_shift, "");
710 num = LLVMBuildNUWAdd(builder, num, increment, "");
711 num = LLVMBuildMul(builder,
712 LLVMBuildZExt(builder, num, ctx->i64, ""),
713 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
714 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
715 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
716 return LLVMBuildLShr(builder, num, post_shift, "");
717 }
718
719 /* See fast_idiv_by_const.h. */
720 /* Both operands must fit in 31 bits and the divisor must not be 1. */
721 LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
722 LLVMValueRef num,
723 LLVMValueRef multiplier,
724 LLVMValueRef post_shift)
725 {
726 LLVMBuilderRef builder = ctx->builder;
727
728 num = LLVMBuildMul(builder,
729 LLVMBuildZExt(builder, num, ctx->i64, ""),
730 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
731 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
732 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
733 return LLVMBuildLShr(builder, num, post_shift, "");
734 }
735
736 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
737 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
738 * already multiplied by two. id is the cube face number.
739 */
740 struct cube_selection_coords {
741 LLVMValueRef stc[2];
742 LLVMValueRef ma;
743 LLVMValueRef id;
744 };
745
746 static void
747 build_cube_intrinsic(struct ac_llvm_context *ctx,
748 LLVMValueRef in[3],
749 struct cube_selection_coords *out)
750 {
751 LLVMTypeRef f32 = ctx->f32;
752
753 out->stc[1] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubetc",
754 f32, in, 3, AC_FUNC_ATTR_READNONE);
755 out->stc[0] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubesc",
756 f32, in, 3, AC_FUNC_ATTR_READNONE);
757 out->ma = ac_build_intrinsic(ctx, "llvm.amdgcn.cubema",
758 f32, in, 3, AC_FUNC_ATTR_READNONE);
759 out->id = ac_build_intrinsic(ctx, "llvm.amdgcn.cubeid",
760 f32, in, 3, AC_FUNC_ATTR_READNONE);
761 }
762
763 /**
764 * Build a manual selection sequence for cube face sc/tc coordinates and
765 * major axis vector (multiplied by 2 for consistency) for the given
766 * vec3 \p coords, for the face implied by \p selcoords.
767 *
768 * For the major axis, we always adjust the sign to be in the direction of
769 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
770 * the selcoords major axis.
771 */
772 static void build_cube_select(struct ac_llvm_context *ctx,
773 const struct cube_selection_coords *selcoords,
774 const LLVMValueRef *coords,
775 LLVMValueRef *out_st,
776 LLVMValueRef *out_ma)
777 {
778 LLVMBuilderRef builder = ctx->builder;
779 LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
780 LLVMValueRef is_ma_positive;
781 LLVMValueRef sgn_ma;
782 LLVMValueRef is_ma_z, is_not_ma_z;
783 LLVMValueRef is_ma_y;
784 LLVMValueRef is_ma_x;
785 LLVMValueRef sgn;
786 LLVMValueRef tmp;
787
788 is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
789 selcoords->ma, LLVMConstReal(f32, 0.0), "");
790 sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
791 LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
792
793 is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
794 is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
795 is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
796 LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
797 is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
798
799 /* Select sc */
800 tmp = LLVMBuildSelect(builder, is_ma_x, coords[2], coords[0], "");
801 sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
802 LLVMBuildSelect(builder, is_ma_z, sgn_ma,
803 LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
804 out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
805
806 /* Select tc */
807 tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
808 sgn = LLVMBuildSelect(builder, is_ma_y, sgn_ma,
809 LLVMConstReal(f32, -1.0), "");
810 out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
811
812 /* Select ma */
813 tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
814 LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
815 tmp = ac_build_intrinsic(ctx, "llvm.fabs.f32",
816 ctx->f32, &tmp, 1, AC_FUNC_ATTR_READNONE);
817 *out_ma = LLVMBuildFMul(builder, tmp, LLVMConstReal(f32, 2.0), "");
818 }
819
820 void
821 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
822 bool is_deriv, bool is_array, bool is_lod,
823 LLVMValueRef *coords_arg,
824 LLVMValueRef *derivs_arg)
825 {
826
827 LLVMBuilderRef builder = ctx->builder;
828 struct cube_selection_coords selcoords;
829 LLVMValueRef coords[3];
830 LLVMValueRef invma;
831
832 if (is_array && !is_lod) {
833 LLVMValueRef tmp = ac_build_round(ctx, coords_arg[3]);
834
835 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
836 *
837 * "For Array forms, the array layer used will be
838 *
839 * max(0, min(d−1, floor(layer+0.5)))
840 *
841 * where d is the depth of the texture array and layer
842 * comes from the component indicated in the tables below.
843 * Workaroudn for an issue where the layer is taken from a
844 * helper invocation which happens to fall on a different
845 * layer due to extrapolation."
846 *
847 * GFX8 and earlier attempt to implement this in hardware by
848 * clamping the value of coords[2] = (8 * layer) + face.
849 * Unfortunately, this means that the we end up with the wrong
850 * face when clamping occurs.
851 *
852 * Clamp the layer earlier to work around the issue.
853 */
854 if (ctx->chip_class <= GFX8) {
855 LLVMValueRef ge0;
856 ge0 = LLVMBuildFCmp(builder, LLVMRealOGE, tmp, ctx->f32_0, "");
857 tmp = LLVMBuildSelect(builder, ge0, tmp, ctx->f32_0, "");
858 }
859
860 coords_arg[3] = tmp;
861 }
862
863 build_cube_intrinsic(ctx, coords_arg, &selcoords);
864
865 invma = ac_build_intrinsic(ctx, "llvm.fabs.f32",
866 ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
867 invma = ac_build_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
868
869 for (int i = 0; i < 2; ++i)
870 coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
871
872 coords[2] = selcoords.id;
873
874 if (is_deriv && derivs_arg) {
875 LLVMValueRef derivs[4];
876 int axis;
877
878 /* Convert cube derivatives to 2D derivatives. */
879 for (axis = 0; axis < 2; axis++) {
880 LLVMValueRef deriv_st[2];
881 LLVMValueRef deriv_ma;
882
883 /* Transform the derivative alongside the texture
884 * coordinate. Mathematically, the correct formula is
885 * as follows. Assume we're projecting onto the +Z face
886 * and denote by dx/dh the derivative of the (original)
887 * X texture coordinate with respect to horizontal
888 * window coordinates. The projection onto the +Z face
889 * plane is:
890 *
891 * f(x,z) = x/z
892 *
893 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
894 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
895 *
896 * This motivatives the implementation below.
897 *
898 * Whether this actually gives the expected results for
899 * apps that might feed in derivatives obtained via
900 * finite differences is anyone's guess. The OpenGL spec
901 * seems awfully quiet about how textureGrad for cube
902 * maps should be handled.
903 */
904 build_cube_select(ctx, &selcoords, &derivs_arg[axis * 3],
905 deriv_st, &deriv_ma);
906
907 deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
908
909 for (int i = 0; i < 2; ++i)
910 derivs[axis * 2 + i] =
911 LLVMBuildFSub(builder,
912 LLVMBuildFMul(builder, deriv_st[i], invma, ""),
913 LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
914 }
915
916 memcpy(derivs_arg, derivs, sizeof(derivs));
917 }
918
919 /* Shift the texture coordinate. This must be applied after the
920 * derivative calculation.
921 */
922 for (int i = 0; i < 2; ++i)
923 coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
924
925 if (is_array) {
926 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
927 /* coords_arg.w component - array_index for cube arrays */
928 coords[2] = ac_build_fmad(ctx, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), coords[2]);
929 }
930
931 memcpy(coords_arg, coords, sizeof(coords));
932 }
933
934
935 LLVMValueRef
936 ac_build_fs_interp(struct ac_llvm_context *ctx,
937 LLVMValueRef llvm_chan,
938 LLVMValueRef attr_number,
939 LLVMValueRef params,
940 LLVMValueRef i,
941 LLVMValueRef j)
942 {
943 LLVMValueRef args[5];
944 LLVMValueRef p1;
945
946 args[0] = i;
947 args[1] = llvm_chan;
948 args[2] = attr_number;
949 args[3] = params;
950
951 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1",
952 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
953
954 args[0] = p1;
955 args[1] = j;
956 args[2] = llvm_chan;
957 args[3] = attr_number;
958 args[4] = params;
959
960 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2",
961 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
962 }
963
964 LLVMValueRef
965 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
966 LLVMValueRef llvm_chan,
967 LLVMValueRef attr_number,
968 LLVMValueRef params,
969 LLVMValueRef i,
970 LLVMValueRef j)
971 {
972 LLVMValueRef args[6];
973 LLVMValueRef p1;
974
975 args[0] = i;
976 args[1] = llvm_chan;
977 args[2] = attr_number;
978 args[3] = ctx->i1false;
979 args[4] = params;
980
981 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1.f16",
982 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
983
984 args[0] = p1;
985 args[1] = j;
986 args[2] = llvm_chan;
987 args[3] = attr_number;
988 args[4] = ctx->i1false;
989 args[5] = params;
990
991 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2.f16",
992 ctx->f16, args, 6, AC_FUNC_ATTR_READNONE);
993 }
994
995 LLVMValueRef
996 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
997 LLVMValueRef parameter,
998 LLVMValueRef llvm_chan,
999 LLVMValueRef attr_number,
1000 LLVMValueRef params)
1001 {
1002 LLVMValueRef args[4];
1003
1004 args[0] = parameter;
1005 args[1] = llvm_chan;
1006 args[2] = attr_number;
1007 args[3] = params;
1008
1009 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.mov",
1010 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
1011 }
1012
1013 LLVMValueRef
1014 ac_build_gep_ptr(struct ac_llvm_context *ctx,
1015 LLVMValueRef base_ptr,
1016 LLVMValueRef index)
1017 {
1018 return LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
1019 }
1020
1021 LLVMValueRef
1022 ac_build_gep0(struct ac_llvm_context *ctx,
1023 LLVMValueRef base_ptr,
1024 LLVMValueRef index)
1025 {
1026 LLVMValueRef indices[2] = {
1027 ctx->i32_0,
1028 index,
1029 };
1030 return LLVMBuildGEP(ctx->builder, base_ptr, indices, 2, "");
1031 }
1032
1033 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
1034 LLVMValueRef index)
1035 {
1036 return LLVMBuildPointerCast(ctx->builder,
1037 LLVMBuildGEP(ctx->builder, ptr, &index, 1, ""),
1038 LLVMTypeOf(ptr), "");
1039 }
1040
1041 void
1042 ac_build_indexed_store(struct ac_llvm_context *ctx,
1043 LLVMValueRef base_ptr, LLVMValueRef index,
1044 LLVMValueRef value)
1045 {
1046 LLVMBuildStore(ctx->builder, value,
1047 ac_build_gep0(ctx, base_ptr, index));
1048 }
1049
1050 /**
1051 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
1052 * It's equivalent to doing a load from &base_ptr[index].
1053 *
1054 * \param base_ptr Where the array starts.
1055 * \param index The element index into the array.
1056 * \param uniform Whether the base_ptr and index can be assumed to be
1057 * dynamically uniform (i.e. load to an SGPR)
1058 * \param invariant Whether the load is invariant (no other opcodes affect it)
1059 * \param no_unsigned_wraparound
1060 * For all possible re-associations and re-distributions of an expression
1061 * "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
1062 * without inbounds in base_ptr), this parameter is true if "addr + offset"
1063 * does not result in an unsigned integer wraparound. This is used for
1064 * optimal code generation of 32-bit pointer arithmetic.
1065 *
1066 * For example, a 32-bit immediate offset that causes a 32-bit unsigned
1067 * integer wraparound can't be an imm offset in s_load_dword, because
1068 * the instruction performs "addr + offset" in 64 bits.
1069 *
1070 * Expected usage for bindless textures by chaining GEPs:
1071 * // possible unsigned wraparound, don't use InBounds:
1072 * ptr1 = LLVMBuildGEP(base_ptr, index);
1073 * image = load(ptr1); // becomes "s_load ptr1, 0"
1074 *
1075 * ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1076 * sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1077 */
1078 static LLVMValueRef
1079 ac_build_load_custom(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1080 LLVMValueRef index, bool uniform, bool invariant,
1081 bool no_unsigned_wraparound)
1082 {
1083 LLVMValueRef pointer, result;
1084
1085 if (no_unsigned_wraparound &&
1086 LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr)) == AC_ADDR_SPACE_CONST_32BIT)
1087 pointer = LLVMBuildInBoundsGEP(ctx->builder, base_ptr, &index, 1, "");
1088 else
1089 pointer = LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
1090
1091 if (uniform)
1092 LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
1093 result = LLVMBuildLoad(ctx->builder, pointer, "");
1094 if (invariant)
1095 LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
1096 return result;
1097 }
1098
1099 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1100 LLVMValueRef index)
1101 {
1102 return ac_build_load_custom(ctx, base_ptr, index, false, false, false);
1103 }
1104
1105 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
1106 LLVMValueRef base_ptr, LLVMValueRef index)
1107 {
1108 return ac_build_load_custom(ctx, base_ptr, index, false, true, false);
1109 }
1110
1111 /* This assumes that there is no unsigned integer wraparound during the address
1112 * computation, excluding all GEPs within base_ptr. */
1113 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
1114 LLVMValueRef base_ptr, LLVMValueRef index)
1115 {
1116 return ac_build_load_custom(ctx, base_ptr, index, true, true, true);
1117 }
1118
1119 /* See ac_build_load_custom() documentation. */
1120 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
1121 LLVMValueRef base_ptr, LLVMValueRef index)
1122 {
1123 return ac_build_load_custom(ctx, base_ptr, index, true, true, false);
1124 }
1125
1126 static unsigned get_load_cache_policy(struct ac_llvm_context *ctx,
1127 unsigned cache_policy)
1128 {
1129 return cache_policy |
1130 (ctx->chip_class >= GFX10 && cache_policy & ac_glc ? ac_dlc : 0);
1131 }
1132
1133 static void
1134 ac_build_llvm7_buffer_store_common(struct ac_llvm_context *ctx,
1135 LLVMValueRef rsrc,
1136 LLVMValueRef data,
1137 LLVMValueRef vindex,
1138 LLVMValueRef voffset,
1139 unsigned num_channels,
1140 unsigned cache_policy,
1141 bool use_format)
1142 {
1143 LLVMValueRef args[] = {
1144 data,
1145 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
1146 vindex ? vindex : ctx->i32_0,
1147 voffset,
1148 LLVMConstInt(ctx->i1, !!(cache_policy & ac_glc), 0),
1149 LLVMConstInt(ctx->i1, !!(cache_policy & ac_slc), 0)
1150 };
1151 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1152
1153 const char *type_names[] = {"f32", "v2f32", "v4f32"};
1154 char name[256];
1155
1156 if (use_format) {
1157 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.store.format.%s",
1158 type_names[func]);
1159 } else {
1160 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.store.%s",
1161 type_names[func]);
1162 }
1163
1164 ac_build_intrinsic(ctx, name, ctx->voidt, args, ARRAY_SIZE(args),
1165 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
1166 }
1167
1168 static void
1169 ac_build_llvm8_buffer_store_common(struct ac_llvm_context *ctx,
1170 LLVMValueRef rsrc,
1171 LLVMValueRef data,
1172 LLVMValueRef vindex,
1173 LLVMValueRef voffset,
1174 LLVMValueRef soffset,
1175 unsigned num_channels,
1176 LLVMTypeRef return_channel_type,
1177 unsigned cache_policy,
1178 bool use_format,
1179 bool structurized)
1180 {
1181 LLVMValueRef args[6];
1182 int idx = 0;
1183 args[idx++] = data;
1184 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1185 if (structurized)
1186 args[idx++] = vindex ? vindex : ctx->i32_0;
1187 args[idx++] = voffset ? voffset : ctx->i32_0;
1188 args[idx++] = soffset ? soffset : ctx->i32_0;
1189 args[idx++] = LLVMConstInt(ctx->i32, cache_policy, 0);
1190 unsigned func = !ac_has_vec3_support(ctx->chip_class, use_format) && num_channels == 3 ? 4 : num_channels;
1191 const char *indexing_kind = structurized ? "struct" : "raw";
1192 char name[256], type_name[8];
1193
1194 LLVMTypeRef type = func > 1 ? LLVMVectorType(return_channel_type, func) : return_channel_type;
1195 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1196
1197 if (use_format) {
1198 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.format.%s",
1199 indexing_kind, type_name);
1200 } else {
1201 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.%s",
1202 indexing_kind, type_name);
1203 }
1204
1205 ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
1206 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
1207 }
1208
1209 void
1210 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
1211 LLVMValueRef rsrc,
1212 LLVMValueRef data,
1213 LLVMValueRef vindex,
1214 LLVMValueRef voffset,
1215 unsigned num_channels,
1216 unsigned cache_policy)
1217 {
1218 if (HAVE_LLVM >= 0x800) {
1219 ac_build_llvm8_buffer_store_common(ctx, rsrc, data, vindex,
1220 voffset, NULL, num_channels,
1221 ctx->f32, cache_policy,
1222 true, true);
1223 } else {
1224 ac_build_llvm7_buffer_store_common(ctx, rsrc, data, vindex, voffset,
1225 num_channels, cache_policy,
1226 true);
1227 }
1228 }
1229
1230 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1231 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1232 * or v4i32 (num_channels=3,4).
1233 */
1234 void
1235 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
1236 LLVMValueRef rsrc,
1237 LLVMValueRef vdata,
1238 unsigned num_channels,
1239 LLVMValueRef voffset,
1240 LLVMValueRef soffset,
1241 unsigned inst_offset,
1242 unsigned cache_policy,
1243 bool swizzle_enable_hint)
1244 {
1245 /* Split 3 channel stores, because only LLVM 9+ support 3-channel
1246 * intrinsics. */
1247 if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false)) {
1248 LLVMValueRef v[3], v01;
1249
1250 for (int i = 0; i < 3; i++) {
1251 v[i] = LLVMBuildExtractElement(ctx->builder, vdata,
1252 LLVMConstInt(ctx->i32, i, 0), "");
1253 }
1254 v01 = ac_build_gather_values(ctx, v, 2);
1255
1256 ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset,
1257 soffset, inst_offset, cache_policy,
1258 swizzle_enable_hint);
1259 ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset,
1260 soffset, inst_offset + 8,
1261 cache_policy,
1262 swizzle_enable_hint);
1263 return;
1264 }
1265
1266 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1267 * (voffset is swizzled, but soffset isn't swizzled).
1268 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1269 */
1270 if (!swizzle_enable_hint) {
1271 LLVMValueRef offset = soffset;
1272
1273 if (inst_offset)
1274 offset = LLVMBuildAdd(ctx->builder, offset,
1275 LLVMConstInt(ctx->i32, inst_offset, 0), "");
1276
1277 if (HAVE_LLVM >= 0x800) {
1278 ac_build_llvm8_buffer_store_common(ctx, rsrc,
1279 ac_to_float(ctx, vdata),
1280 ctx->i32_0,
1281 voffset, offset,
1282 num_channels,
1283 ctx->f32,
1284 cache_policy,
1285 false, false);
1286 } else {
1287 if (voffset)
1288 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1289
1290 ac_build_llvm7_buffer_store_common(ctx, rsrc,
1291 ac_to_float(ctx, vdata),
1292 ctx->i32_0, offset,
1293 num_channels, cache_policy,
1294 false);
1295 }
1296 return;
1297 }
1298
1299 static const unsigned dfmts[] = {
1300 V_008F0C_BUF_DATA_FORMAT_32,
1301 V_008F0C_BUF_DATA_FORMAT_32_32,
1302 V_008F0C_BUF_DATA_FORMAT_32_32_32,
1303 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1304 };
1305 unsigned dfmt = dfmts[num_channels - 1];
1306 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1307 LLVMValueRef immoffset = LLVMConstInt(ctx->i32, inst_offset, 0);
1308
1309 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1310 immoffset, num_channels, dfmt, nfmt, cache_policy);
1311 }
1312
1313 static LLVMValueRef
1314 ac_build_llvm7_buffer_load_common(struct ac_llvm_context *ctx,
1315 LLVMValueRef rsrc,
1316 LLVMValueRef vindex,
1317 LLVMValueRef voffset,
1318 unsigned num_channels,
1319 unsigned cache_policy,
1320 bool can_speculate,
1321 bool use_format)
1322 {
1323 LLVMValueRef args[] = {
1324 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
1325 vindex ? vindex : ctx->i32_0,
1326 voffset,
1327 LLVMConstInt(ctx->i1, !!(cache_policy & ac_glc), 0),
1328 LLVMConstInt(ctx->i1, !!(cache_policy & ac_slc), 0)
1329 };
1330 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1331
1332 LLVMTypeRef types[] = {ctx->f32, ctx->v2f32, ctx->v4f32};
1333 const char *type_names[] = {"f32", "v2f32", "v4f32"};
1334 char name[256];
1335
1336 if (use_format) {
1337 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.format.%s",
1338 type_names[func]);
1339 } else {
1340 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.%s",
1341 type_names[func]);
1342 }
1343
1344 return ac_build_intrinsic(ctx, name, types[func], args,
1345 ARRAY_SIZE(args),
1346 ac_get_load_intr_attribs(can_speculate));
1347 }
1348
1349 static LLVMValueRef
1350 ac_build_llvm8_buffer_load_common(struct ac_llvm_context *ctx,
1351 LLVMValueRef rsrc,
1352 LLVMValueRef vindex,
1353 LLVMValueRef voffset,
1354 LLVMValueRef soffset,
1355 unsigned num_channels,
1356 LLVMTypeRef channel_type,
1357 unsigned cache_policy,
1358 bool can_speculate,
1359 bool use_format,
1360 bool structurized)
1361 {
1362 LLVMValueRef args[5];
1363 int idx = 0;
1364 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1365 if (structurized)
1366 args[idx++] = vindex ? vindex : ctx->i32_0;
1367 args[idx++] = voffset ? voffset : ctx->i32_0;
1368 args[idx++] = soffset ? soffset : ctx->i32_0;
1369 args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
1370 unsigned func = !ac_has_vec3_support(ctx->chip_class, use_format) && num_channels == 3 ? 4 : num_channels;
1371 const char *indexing_kind = structurized ? "struct" : "raw";
1372 char name[256], type_name[8];
1373
1374 LLVMTypeRef type = func > 1 ? LLVMVectorType(channel_type, func) : channel_type;
1375 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1376
1377 if (use_format) {
1378 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.format.%s",
1379 indexing_kind, type_name);
1380 } else {
1381 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.%s",
1382 indexing_kind, type_name);
1383 }
1384
1385 return ac_build_intrinsic(ctx, name, type, args, idx,
1386 ac_get_load_intr_attribs(can_speculate));
1387 }
1388
1389 LLVMValueRef
1390 ac_build_buffer_load(struct ac_llvm_context *ctx,
1391 LLVMValueRef rsrc,
1392 int num_channels,
1393 LLVMValueRef vindex,
1394 LLVMValueRef voffset,
1395 LLVMValueRef soffset,
1396 unsigned inst_offset,
1397 unsigned cache_policy,
1398 bool can_speculate,
1399 bool allow_smem)
1400 {
1401 LLVMValueRef offset = LLVMConstInt(ctx->i32, inst_offset, 0);
1402 if (voffset)
1403 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1404 if (soffset)
1405 offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
1406
1407 if (allow_smem && !(cache_policy & ac_slc) &&
1408 (!(cache_policy & ac_glc) || (HAVE_LLVM >= 0x0800 && ctx->chip_class >= GFX8))) {
1409 assert(vindex == NULL);
1410
1411 LLVMValueRef result[8];
1412
1413 for (int i = 0; i < num_channels; i++) {
1414 if (i) {
1415 offset = LLVMBuildAdd(ctx->builder, offset,
1416 LLVMConstInt(ctx->i32, 4, 0), "");
1417 }
1418 const char *intrname =
1419 HAVE_LLVM >= 0x0800 ? "llvm.amdgcn.s.buffer.load.f32"
1420 : "llvm.SI.load.const.v4i32";
1421 unsigned num_args = HAVE_LLVM >= 0x0800 ? 3 : 2;
1422 LLVMValueRef args[3] = {
1423 rsrc,
1424 offset,
1425 LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0),
1426 };
1427 result[i] = ac_build_intrinsic(ctx, intrname,
1428 ctx->f32, args, num_args,
1429 AC_FUNC_ATTR_READNONE |
1430 (HAVE_LLVM < 0x0800 ? AC_FUNC_ATTR_LEGACY : 0));
1431 }
1432 if (num_channels == 1)
1433 return result[0];
1434
1435 if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false))
1436 result[num_channels++] = LLVMGetUndef(ctx->f32);
1437 return ac_build_gather_values(ctx, result, num_channels);
1438 }
1439
1440 if (HAVE_LLVM >= 0x0800) {
1441 return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex,
1442 offset, ctx->i32_0,
1443 num_channels, ctx->f32,
1444 cache_policy,
1445 can_speculate, false,
1446 false);
1447 }
1448
1449 return ac_build_llvm7_buffer_load_common(ctx, rsrc, vindex, offset,
1450 num_channels, cache_policy,
1451 can_speculate, false);
1452 }
1453
1454 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
1455 LLVMValueRef rsrc,
1456 LLVMValueRef vindex,
1457 LLVMValueRef voffset,
1458 unsigned num_channels,
1459 unsigned cache_policy,
1460 bool can_speculate)
1461 {
1462 if (HAVE_LLVM >= 0x800) {
1463 return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
1464 num_channels, ctx->f32,
1465 cache_policy, can_speculate, true, true);
1466 }
1467 return ac_build_llvm7_buffer_load_common(ctx, rsrc, vindex, voffset,
1468 num_channels, cache_policy,
1469 can_speculate, true);
1470 }
1471
1472 LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
1473 LLVMValueRef rsrc,
1474 LLVMValueRef vindex,
1475 LLVMValueRef voffset,
1476 unsigned num_channels,
1477 unsigned cache_policy,
1478 bool can_speculate)
1479 {
1480 if (HAVE_LLVM >= 0x800) {
1481 return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
1482 num_channels, ctx->f32,
1483 cache_policy, can_speculate, true, true);
1484 }
1485
1486 LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->builder, rsrc, LLVMConstInt(ctx->i32, 2, 0), "");
1487 LLVMValueRef stride = LLVMBuildExtractElement(ctx->builder, rsrc, ctx->i32_1, "");
1488 stride = LLVMBuildLShr(ctx->builder, stride, LLVMConstInt(ctx->i32, 16, 0), "");
1489
1490 LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->builder,
1491 LLVMBuildICmp(ctx->builder, LLVMIntUGT, elem_count, stride, ""),
1492 elem_count, stride, "");
1493
1494 LLVMValueRef new_rsrc = LLVMBuildInsertElement(ctx->builder, rsrc, new_elem_count,
1495 LLVMConstInt(ctx->i32, 2, 0), "");
1496
1497 return ac_build_llvm7_buffer_load_common(ctx, new_rsrc, vindex, voffset,
1498 num_channels, cache_policy,
1499 can_speculate, true);
1500 }
1501
1502 /// Translate a (dfmt, nfmt) pair into a chip-appropriate combined format
1503 /// value for LLVM8+ tbuffer intrinsics.
1504 static unsigned
1505 ac_get_tbuffer_format(struct ac_llvm_context *ctx,
1506 unsigned dfmt, unsigned nfmt)
1507 {
1508 if (ctx->chip_class >= GFX10) {
1509 unsigned format;
1510 switch (dfmt) {
1511 default: unreachable("bad dfmt");
1512 case V_008F0C_BUF_DATA_FORMAT_INVALID: format = V_008F0C_IMG_FORMAT_INVALID; break;
1513 case V_008F0C_BUF_DATA_FORMAT_8: format = V_008F0C_IMG_FORMAT_8_UINT; break;
1514 case V_008F0C_BUF_DATA_FORMAT_8_8: format = V_008F0C_IMG_FORMAT_8_8_UINT; break;
1515 case V_008F0C_BUF_DATA_FORMAT_8_8_8_8: format = V_008F0C_IMG_FORMAT_8_8_8_8_UINT; break;
1516 case V_008F0C_BUF_DATA_FORMAT_16: format = V_008F0C_IMG_FORMAT_16_UINT; break;
1517 case V_008F0C_BUF_DATA_FORMAT_16_16: format = V_008F0C_IMG_FORMAT_16_16_UINT; break;
1518 case V_008F0C_BUF_DATA_FORMAT_16_16_16_16: format = V_008F0C_IMG_FORMAT_16_16_16_16_UINT; break;
1519 case V_008F0C_BUF_DATA_FORMAT_32: format = V_008F0C_IMG_FORMAT_32_UINT; break;
1520 case V_008F0C_BUF_DATA_FORMAT_32_32: format = V_008F0C_IMG_FORMAT_32_32_UINT; break;
1521 case V_008F0C_BUF_DATA_FORMAT_32_32_32: format = V_008F0C_IMG_FORMAT_32_32_32_UINT; break;
1522 case V_008F0C_BUF_DATA_FORMAT_32_32_32_32: format = V_008F0C_IMG_FORMAT_32_32_32_32_UINT; break;
1523 case V_008F0C_BUF_DATA_FORMAT_2_10_10_10: format = V_008F0C_IMG_FORMAT_2_10_10_10_UINT; break;
1524 }
1525
1526 // Use the regularity properties of the combined format enum.
1527 //
1528 // Note: float is incompatible with 8-bit data formats,
1529 // [us]{norm,scaled} are incomparible with 32-bit data formats.
1530 // [us]scaled are not writable.
1531 switch (nfmt) {
1532 case V_008F0C_BUF_NUM_FORMAT_UNORM: format -= 4; break;
1533 case V_008F0C_BUF_NUM_FORMAT_SNORM: format -= 3; break;
1534 case V_008F0C_BUF_NUM_FORMAT_USCALED: format -= 2; break;
1535 case V_008F0C_BUF_NUM_FORMAT_SSCALED: format -= 1; break;
1536 default: unreachable("bad nfmt");
1537 case V_008F0C_BUF_NUM_FORMAT_UINT: break;
1538 case V_008F0C_BUF_NUM_FORMAT_SINT: format += 1; break;
1539 case V_008F0C_BUF_NUM_FORMAT_FLOAT: format += 2; break;
1540 }
1541
1542 return format;
1543 } else {
1544 return dfmt | (nfmt << 4);
1545 }
1546 }
1547
1548 static LLVMValueRef
1549 ac_build_llvm8_tbuffer_load(struct ac_llvm_context *ctx,
1550 LLVMValueRef rsrc,
1551 LLVMValueRef vindex,
1552 LLVMValueRef voffset,
1553 LLVMValueRef soffset,
1554 unsigned num_channels,
1555 unsigned dfmt,
1556 unsigned nfmt,
1557 unsigned cache_policy,
1558 bool can_speculate,
1559 bool structurized)
1560 {
1561 LLVMValueRef args[6];
1562 int idx = 0;
1563 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1564 if (structurized)
1565 args[idx++] = vindex ? vindex : ctx->i32_0;
1566 args[idx++] = voffset ? voffset : ctx->i32_0;
1567 args[idx++] = soffset ? soffset : ctx->i32_0;
1568 args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx, dfmt, nfmt), 0);
1569 args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
1570 unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
1571 const char *indexing_kind = structurized ? "struct" : "raw";
1572 char name[256], type_name[8];
1573
1574 LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
1575 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1576
1577 snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.load.%s",
1578 indexing_kind, type_name);
1579
1580 return ac_build_intrinsic(ctx, name, type, args, idx,
1581 ac_get_load_intr_attribs(can_speculate));
1582 }
1583
1584 static LLVMValueRef
1585 ac_build_tbuffer_load(struct ac_llvm_context *ctx,
1586 LLVMValueRef rsrc,
1587 LLVMValueRef vindex,
1588 LLVMValueRef voffset,
1589 LLVMValueRef soffset,
1590 LLVMValueRef immoffset,
1591 unsigned num_channels,
1592 unsigned dfmt,
1593 unsigned nfmt,
1594 unsigned cache_policy,
1595 bool can_speculate,
1596 bool structurized) /* only matters for LLVM 8+ */
1597 {
1598 if (HAVE_LLVM >= 0x800) {
1599 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1600
1601 return ac_build_llvm8_tbuffer_load(ctx, rsrc, vindex, voffset,
1602 soffset, num_channels,
1603 dfmt, nfmt, cache_policy,
1604 can_speculate, structurized);
1605 }
1606
1607 LLVMValueRef args[] = {
1608 rsrc,
1609 vindex ? vindex : ctx->i32_0,
1610 voffset,
1611 soffset,
1612 immoffset,
1613 LLVMConstInt(ctx->i32, dfmt, false),
1614 LLVMConstInt(ctx->i32, nfmt, false),
1615 LLVMConstInt(ctx->i1, !!(cache_policy & ac_glc), false),
1616 LLVMConstInt(ctx->i1, !!(cache_policy & ac_slc), false),
1617 };
1618 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1619 LLVMTypeRef types[] = {ctx->i32, ctx->v2i32, ctx->v4i32};
1620 const char *type_names[] = {"i32", "v2i32", "v4i32"};
1621 char name[256];
1622
1623 snprintf(name, sizeof(name), "llvm.amdgcn.tbuffer.load.%s",
1624 type_names[func]);
1625
1626 return ac_build_intrinsic(ctx, name, types[func], args, 9,
1627 ac_get_load_intr_attribs(can_speculate));
1628 }
1629
1630 LLVMValueRef
1631 ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
1632 LLVMValueRef rsrc,
1633 LLVMValueRef vindex,
1634 LLVMValueRef voffset,
1635 LLVMValueRef soffset,
1636 LLVMValueRef immoffset,
1637 unsigned num_channels,
1638 unsigned dfmt,
1639 unsigned nfmt,
1640 unsigned cache_policy,
1641 bool can_speculate)
1642 {
1643 return ac_build_tbuffer_load(ctx, rsrc, vindex, voffset, soffset,
1644 immoffset, num_channels, dfmt, nfmt,
1645 cache_policy, can_speculate, true);
1646 }
1647
1648 LLVMValueRef
1649 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
1650 LLVMValueRef rsrc,
1651 LLVMValueRef voffset,
1652 LLVMValueRef soffset,
1653 LLVMValueRef immoffset,
1654 unsigned num_channels,
1655 unsigned dfmt,
1656 unsigned nfmt,
1657 unsigned cache_policy,
1658 bool can_speculate)
1659 {
1660 return ac_build_tbuffer_load(ctx, rsrc, NULL, voffset, soffset,
1661 immoffset, num_channels, dfmt, nfmt,
1662 cache_policy, can_speculate, false);
1663 }
1664
1665 LLVMValueRef
1666 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
1667 LLVMValueRef rsrc,
1668 LLVMValueRef voffset,
1669 LLVMValueRef soffset,
1670 LLVMValueRef immoffset,
1671 unsigned cache_policy)
1672 {
1673 LLVMValueRef res;
1674
1675 if (HAVE_LLVM >= 0x900) {
1676 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1677
1678 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1679 res = ac_build_llvm8_buffer_load_common(ctx, rsrc, NULL,
1680 voffset, soffset,
1681 1, ctx->i16, cache_policy,
1682 false, false, false);
1683 } else {
1684 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
1685 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1686
1687 res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
1688 immoffset, 1, dfmt, nfmt, cache_policy,
1689 false);
1690
1691 res = LLVMBuildTrunc(ctx->builder, res, ctx->i16, "");
1692 }
1693
1694 return res;
1695 }
1696
1697 LLVMValueRef
1698 ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
1699 LLVMValueRef rsrc,
1700 LLVMValueRef voffset,
1701 LLVMValueRef soffset,
1702 LLVMValueRef immoffset,
1703 unsigned cache_policy)
1704 {
1705 LLVMValueRef res;
1706
1707 if (HAVE_LLVM >= 0x900) {
1708 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1709
1710 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1711 res = ac_build_llvm8_buffer_load_common(ctx, rsrc, NULL,
1712 voffset, soffset,
1713 1, ctx->i8, cache_policy,
1714 false, false, false);
1715 } else {
1716 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
1717 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1718
1719 res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
1720 immoffset, 1, dfmt, nfmt, cache_policy,
1721 false);
1722
1723 res = LLVMBuildTrunc(ctx->builder, res, ctx->i8, "");
1724 }
1725
1726 return res;
1727 }
1728
1729 /**
1730 * Convert an 11- or 10-bit unsigned floating point number to an f32.
1731 *
1732 * The input exponent is expected to be biased analogous to IEEE-754, i.e. by
1733 * 2^(exp_bits-1) - 1 (as defined in OpenGL and other graphics APIs).
1734 */
1735 static LLVMValueRef
1736 ac_ufN_to_float(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned exp_bits, unsigned mant_bits)
1737 {
1738 assert(LLVMTypeOf(src) == ctx->i32);
1739
1740 LLVMValueRef tmp;
1741 LLVMValueRef mantissa;
1742 mantissa = LLVMBuildAnd(ctx->builder, src, LLVMConstInt(ctx->i32, (1 << mant_bits) - 1, false), "");
1743
1744 /* Converting normal numbers is just a shift + correcting the exponent bias */
1745 unsigned normal_shift = 23 - mant_bits;
1746 unsigned bias_shift = 127 - ((1 << (exp_bits - 1)) - 1);
1747 LLVMValueRef shifted, normal;
1748
1749 shifted = LLVMBuildShl(ctx->builder, src, LLVMConstInt(ctx->i32, normal_shift, false), "");
1750 normal = LLVMBuildAdd(ctx->builder, shifted, LLVMConstInt(ctx->i32, bias_shift << 23, false), "");
1751
1752 /* Converting nan/inf numbers is the same, but with a different exponent update */
1753 LLVMValueRef naninf;
1754 naninf = LLVMBuildOr(ctx->builder, normal, LLVMConstInt(ctx->i32, 0xff << 23, false), "");
1755
1756 /* Converting denormals is the complex case: determine the leading zeros of the
1757 * mantissa to obtain the correct shift for the mantissa and exponent correction.
1758 */
1759 LLVMValueRef denormal;
1760 LLVMValueRef params[2] = {
1761 mantissa,
1762 ctx->i1true, /* result can be undef when arg is 0 */
1763 };
1764 LLVMValueRef ctlz = ac_build_intrinsic(ctx, "llvm.ctlz.i32", ctx->i32,
1765 params, 2, AC_FUNC_ATTR_READNONE);
1766
1767 /* Shift such that the leading 1 ends up as the LSB of the exponent field. */
1768 tmp = LLVMBuildSub(ctx->builder, ctlz, LLVMConstInt(ctx->i32, 8, false), "");
1769 denormal = LLVMBuildShl(ctx->builder, mantissa, tmp, "");
1770
1771 unsigned denormal_exp = bias_shift + (32 - mant_bits) - 1;
1772 tmp = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, denormal_exp, false), ctlz, "");
1773 tmp = LLVMBuildShl(ctx->builder, tmp, LLVMConstInt(ctx->i32, 23, false), "");
1774 denormal = LLVMBuildAdd(ctx->builder, denormal, tmp, "");
1775
1776 /* Select the final result. */
1777 LLVMValueRef result;
1778
1779 tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src,
1780 LLVMConstInt(ctx->i32, ((1 << exp_bits) - 1) << mant_bits, false), "");
1781 result = LLVMBuildSelect(ctx->builder, tmp, naninf, normal, "");
1782
1783 tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src,
1784 LLVMConstInt(ctx->i32, 1 << mant_bits, false), "");
1785 result = LLVMBuildSelect(ctx->builder, tmp, result, denormal, "");
1786
1787 tmp = LLVMBuildICmp(ctx->builder, LLVMIntNE, src, ctx->i32_0, "");
1788 result = LLVMBuildSelect(ctx->builder, tmp, result, ctx->i32_0, "");
1789
1790 return ac_to_float(ctx, result);
1791 }
1792
1793 /**
1794 * Generate a fully general open coded buffer format fetch with all required
1795 * fixups suitable for vertex fetch, using non-format buffer loads.
1796 *
1797 * Some combinations of argument values have special interpretations:
1798 * - size = 8 bytes, format = fixed indicates PIPE_FORMAT_R11G11B10_FLOAT
1799 * - size = 8 bytes, format != {float,fixed} indicates a 2_10_10_10 data format
1800 *
1801 * \param log_size log(size of channel in bytes)
1802 * \param num_channels number of channels (1 to 4)
1803 * \param format AC_FETCH_FORMAT_xxx value
1804 * \param reverse whether XYZ channels are reversed
1805 * \param known_aligned whether the source is known to be aligned to hardware's
1806 * effective element size for loading the given format
1807 * (note: this means dword alignment for 8_8_8_8, 16_16, etc.)
1808 * \param rsrc buffer resource descriptor
1809 * \return the resulting vector of floats or integers bitcast to <4 x i32>
1810 */
1811 LLVMValueRef
1812 ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
1813 unsigned log_size,
1814 unsigned num_channels,
1815 unsigned format,
1816 bool reverse,
1817 bool known_aligned,
1818 LLVMValueRef rsrc,
1819 LLVMValueRef vindex,
1820 LLVMValueRef voffset,
1821 LLVMValueRef soffset,
1822 unsigned cache_policy,
1823 bool can_speculate)
1824 {
1825 LLVMValueRef tmp;
1826 unsigned load_log_size = log_size;
1827 unsigned load_num_channels = num_channels;
1828 if (log_size == 3) {
1829 load_log_size = 2;
1830 if (format == AC_FETCH_FORMAT_FLOAT) {
1831 load_num_channels = 2 * num_channels;
1832 } else {
1833 load_num_channels = 1; /* 10_11_11 or 2_10_10_10 */
1834 }
1835 }
1836
1837 int log_recombine = 0;
1838 if (ctx->chip_class == GFX6 && !known_aligned) {
1839 /* Avoid alignment restrictions by loading one byte at a time. */
1840 load_num_channels <<= load_log_size;
1841 log_recombine = load_log_size;
1842 load_log_size = 0;
1843 } else if (load_num_channels == 2 || load_num_channels == 4) {
1844 log_recombine = -util_logbase2(load_num_channels);
1845 load_num_channels = 1;
1846 load_log_size += -log_recombine;
1847 }
1848
1849 assert(load_log_size >= 2 || HAVE_LLVM >= 0x0900);
1850
1851 LLVMValueRef loads[32]; /* up to 32 bytes */
1852 for (unsigned i = 0; i < load_num_channels; ++i) {
1853 tmp = LLVMBuildAdd(ctx->builder, soffset,
1854 LLVMConstInt(ctx->i32, i << load_log_size, false), "");
1855 if (HAVE_LLVM >= 0x0800) {
1856 LLVMTypeRef channel_type = load_log_size == 0 ? ctx->i8 :
1857 load_log_size == 1 ? ctx->i16 : ctx->i32;
1858 unsigned num_channels = 1 << (MAX2(load_log_size, 2) - 2);
1859 loads[i] = ac_build_llvm8_buffer_load_common(
1860 ctx, rsrc, vindex, voffset, tmp,
1861 num_channels, channel_type, cache_policy,
1862 can_speculate, false, true);
1863 } else {
1864 tmp = LLVMBuildAdd(ctx->builder, voffset, tmp, "");
1865 loads[i] = ac_build_llvm7_buffer_load_common(
1866 ctx, rsrc, vindex, tmp,
1867 1 << (load_log_size - 2), cache_policy, can_speculate, false);
1868 }
1869 if (load_log_size >= 2)
1870 loads[i] = ac_to_integer(ctx, loads[i]);
1871 }
1872
1873 if (log_recombine > 0) {
1874 /* Recombine bytes if necessary (GFX6 only) */
1875 LLVMTypeRef dst_type = log_recombine == 2 ? ctx->i32 : ctx->i16;
1876
1877 for (unsigned src = 0, dst = 0; src < load_num_channels; ++dst) {
1878 LLVMValueRef accum = NULL;
1879 for (unsigned i = 0; i < (1 << log_recombine); ++i, ++src) {
1880 tmp = LLVMBuildZExt(ctx->builder, loads[src], dst_type, "");
1881 if (i == 0) {
1882 accum = tmp;
1883 } else {
1884 tmp = LLVMBuildShl(ctx->builder, tmp,
1885 LLVMConstInt(dst_type, 8 * i, false), "");
1886 accum = LLVMBuildOr(ctx->builder, accum, tmp, "");
1887 }
1888 }
1889 loads[dst] = accum;
1890 }
1891 } else if (log_recombine < 0) {
1892 /* Split vectors of dwords */
1893 if (load_log_size > 2) {
1894 assert(load_num_channels == 1);
1895 LLVMValueRef loaded = loads[0];
1896 unsigned log_split = load_log_size - 2;
1897 log_recombine += log_split;
1898 load_num_channels = 1 << log_split;
1899 load_log_size = 2;
1900 for (unsigned i = 0; i < load_num_channels; ++i) {
1901 tmp = LLVMConstInt(ctx->i32, i, false);
1902 loads[i] = LLVMBuildExtractElement(ctx->builder, loaded, tmp, "");
1903 }
1904 }
1905
1906 /* Further split dwords and shorts if required */
1907 if (log_recombine < 0) {
1908 for (unsigned src = load_num_channels,
1909 dst = load_num_channels << -log_recombine;
1910 src > 0; --src) {
1911 unsigned dst_bits = 1 << (3 + load_log_size + log_recombine);
1912 LLVMTypeRef dst_type = LLVMIntTypeInContext(ctx->context, dst_bits);
1913 LLVMValueRef loaded = loads[src - 1];
1914 LLVMTypeRef loaded_type = LLVMTypeOf(loaded);
1915 for (unsigned i = 1 << -log_recombine; i > 0; --i, --dst) {
1916 tmp = LLVMConstInt(loaded_type, dst_bits * (i - 1), false);
1917 tmp = LLVMBuildLShr(ctx->builder, loaded, tmp, "");
1918 loads[dst - 1] = LLVMBuildTrunc(ctx->builder, tmp, dst_type, "");
1919 }
1920 }
1921 }
1922 }
1923
1924 if (log_size == 3) {
1925 if (format == AC_FETCH_FORMAT_FLOAT) {
1926 for (unsigned i = 0; i < num_channels; ++i) {
1927 tmp = ac_build_gather_values(ctx, &loads[2 * i], 2);
1928 loads[i] = LLVMBuildBitCast(ctx->builder, tmp, ctx->f64, "");
1929 }
1930 } else if (format == AC_FETCH_FORMAT_FIXED) {
1931 /* 10_11_11_FLOAT */
1932 LLVMValueRef data = loads[0];
1933 LLVMValueRef i32_2047 = LLVMConstInt(ctx->i32, 2047, false);
1934 LLVMValueRef r = LLVMBuildAnd(ctx->builder, data, i32_2047, "");
1935 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 11, false), "");
1936 LLVMValueRef g = LLVMBuildAnd(ctx->builder, tmp, i32_2047, "");
1937 LLVMValueRef b = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 22, false), "");
1938
1939 loads[0] = ac_to_integer(ctx, ac_ufN_to_float(ctx, r, 5, 6));
1940 loads[1] = ac_to_integer(ctx, ac_ufN_to_float(ctx, g, 5, 6));
1941 loads[2] = ac_to_integer(ctx, ac_ufN_to_float(ctx, b, 5, 5));
1942
1943 num_channels = 3;
1944 log_size = 2;
1945 format = AC_FETCH_FORMAT_FLOAT;
1946 } else {
1947 /* 2_10_10_10 data formats */
1948 LLVMValueRef data = loads[0];
1949 LLVMTypeRef i10 = LLVMIntTypeInContext(ctx->context, 10);
1950 LLVMTypeRef i2 = LLVMIntTypeInContext(ctx->context, 2);
1951 loads[0] = LLVMBuildTrunc(ctx->builder, data, i10, "");
1952 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 10, false), "");
1953 loads[1] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
1954 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 20, false), "");
1955 loads[2] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
1956 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 30, false), "");
1957 loads[3] = LLVMBuildTrunc(ctx->builder, tmp, i2, "");
1958
1959 num_channels = 4;
1960 }
1961 }
1962
1963 if (format == AC_FETCH_FORMAT_FLOAT) {
1964 if (log_size != 2) {
1965 for (unsigned chan = 0; chan < num_channels; ++chan) {
1966 tmp = ac_to_float(ctx, loads[chan]);
1967 if (log_size == 3)
1968 tmp = LLVMBuildFPTrunc(ctx->builder, tmp, ctx->f32, "");
1969 else if (log_size == 1)
1970 tmp = LLVMBuildFPExt(ctx->builder, tmp, ctx->f32, "");
1971 loads[chan] = ac_to_integer(ctx, tmp);
1972 }
1973 }
1974 } else if (format == AC_FETCH_FORMAT_UINT) {
1975 if (log_size != 2) {
1976 for (unsigned chan = 0; chan < num_channels; ++chan)
1977 loads[chan] = LLVMBuildZExt(ctx->builder, loads[chan], ctx->i32, "");
1978 }
1979 } else if (format == AC_FETCH_FORMAT_SINT) {
1980 if (log_size != 2) {
1981 for (unsigned chan = 0; chan < num_channels; ++chan)
1982 loads[chan] = LLVMBuildSExt(ctx->builder, loads[chan], ctx->i32, "");
1983 }
1984 } else {
1985 bool unsign = format == AC_FETCH_FORMAT_UNORM ||
1986 format == AC_FETCH_FORMAT_USCALED ||
1987 format == AC_FETCH_FORMAT_UINT;
1988
1989 for (unsigned chan = 0; chan < num_channels; ++chan) {
1990 if (unsign) {
1991 tmp = LLVMBuildUIToFP(ctx->builder, loads[chan], ctx->f32, "");
1992 } else {
1993 tmp = LLVMBuildSIToFP(ctx->builder, loads[chan], ctx->f32, "");
1994 }
1995
1996 LLVMValueRef scale = NULL;
1997 if (format == AC_FETCH_FORMAT_FIXED) {
1998 assert(log_size == 2);
1999 scale = LLVMConstReal(ctx->f32, 1.0 / 0x10000);
2000 } else if (format == AC_FETCH_FORMAT_UNORM) {
2001 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
2002 scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << bits) - 1));
2003 } else if (format == AC_FETCH_FORMAT_SNORM) {
2004 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
2005 scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << (bits - 1)) - 1));
2006 }
2007 if (scale)
2008 tmp = LLVMBuildFMul(ctx->builder, tmp, scale, "");
2009
2010 if (format == AC_FETCH_FORMAT_SNORM) {
2011 /* Clamp to [-1, 1] */
2012 LLVMValueRef neg_one = LLVMConstReal(ctx->f32, -1.0);
2013 LLVMValueRef clamp =
2014 LLVMBuildFCmp(ctx->builder, LLVMRealULT, tmp, neg_one, "");
2015 tmp = LLVMBuildSelect(ctx->builder, clamp, neg_one, tmp, "");
2016 }
2017
2018 loads[chan] = ac_to_integer(ctx, tmp);
2019 }
2020 }
2021
2022 while (num_channels < 4) {
2023 if (format == AC_FETCH_FORMAT_UINT || format == AC_FETCH_FORMAT_SINT) {
2024 loads[num_channels] = num_channels == 3 ? ctx->i32_1 : ctx->i32_0;
2025 } else {
2026 loads[num_channels] = ac_to_integer(ctx, num_channels == 3 ? ctx->f32_1 : ctx->f32_0);
2027 }
2028 num_channels++;
2029 }
2030
2031 if (reverse) {
2032 tmp = loads[0];
2033 loads[0] = loads[2];
2034 loads[2] = tmp;
2035 }
2036
2037 return ac_build_gather_values(ctx, loads, 4);
2038 }
2039
2040 static void
2041 ac_build_llvm8_tbuffer_store(struct ac_llvm_context *ctx,
2042 LLVMValueRef rsrc,
2043 LLVMValueRef vdata,
2044 LLVMValueRef vindex,
2045 LLVMValueRef voffset,
2046 LLVMValueRef soffset,
2047 unsigned num_channels,
2048 unsigned dfmt,
2049 unsigned nfmt,
2050 unsigned cache_policy,
2051 bool structurized)
2052 {
2053 LLVMValueRef args[7];
2054 int idx = 0;
2055 args[idx++] = vdata;
2056 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
2057 if (structurized)
2058 args[idx++] = vindex ? vindex : ctx->i32_0;
2059 args[idx++] = voffset ? voffset : ctx->i32_0;
2060 args[idx++] = soffset ? soffset : ctx->i32_0;
2061 args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx, dfmt, nfmt), 0);
2062 args[idx++] = LLVMConstInt(ctx->i32, cache_policy, 0);
2063 unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
2064 const char *indexing_kind = structurized ? "struct" : "raw";
2065 char name[256], type_name[8];
2066
2067 LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
2068 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
2069
2070 snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.store.%s",
2071 indexing_kind, type_name);
2072
2073 ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
2074 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
2075 }
2076
2077 static void
2078 ac_build_tbuffer_store(struct ac_llvm_context *ctx,
2079 LLVMValueRef rsrc,
2080 LLVMValueRef vdata,
2081 LLVMValueRef vindex,
2082 LLVMValueRef voffset,
2083 LLVMValueRef soffset,
2084 LLVMValueRef immoffset,
2085 unsigned num_channels,
2086 unsigned dfmt,
2087 unsigned nfmt,
2088 unsigned cache_policy,
2089 bool structurized) /* only matters for LLVM 8+ */
2090 {
2091 if (HAVE_LLVM >= 0x800) {
2092 voffset = LLVMBuildAdd(ctx->builder,
2093 voffset ? voffset : ctx->i32_0,
2094 immoffset, "");
2095
2096 ac_build_llvm8_tbuffer_store(ctx, rsrc, vdata, vindex, voffset,
2097 soffset, num_channels, dfmt, nfmt,
2098 cache_policy, structurized);
2099 } else {
2100 LLVMValueRef params[] = {
2101 vdata,
2102 rsrc,
2103 vindex ? vindex : ctx->i32_0,
2104 voffset ? voffset : ctx->i32_0,
2105 soffset ? soffset : ctx->i32_0,
2106 immoffset,
2107 LLVMConstInt(ctx->i32, dfmt, false),
2108 LLVMConstInt(ctx->i32, nfmt, false),
2109 LLVMConstInt(ctx->i1, !!(cache_policy & ac_glc), false),
2110 LLVMConstInt(ctx->i1, !!(cache_policy & ac_slc), false),
2111 };
2112 unsigned func = CLAMP(num_channels, 1, 3) - 1;
2113 const char *type_names[] = {"i32", "v2i32", "v4i32"};
2114 char name[256];
2115
2116 snprintf(name, sizeof(name), "llvm.amdgcn.tbuffer.store.%s",
2117 type_names[func]);
2118
2119 ac_build_intrinsic(ctx, name, ctx->voidt, params, 10,
2120 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
2121 }
2122 }
2123
2124 void
2125 ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
2126 LLVMValueRef rsrc,
2127 LLVMValueRef vdata,
2128 LLVMValueRef vindex,
2129 LLVMValueRef voffset,
2130 LLVMValueRef soffset,
2131 LLVMValueRef immoffset,
2132 unsigned num_channels,
2133 unsigned dfmt,
2134 unsigned nfmt,
2135 unsigned cache_policy)
2136 {
2137 ac_build_tbuffer_store(ctx, rsrc, vdata, vindex, voffset, soffset,
2138 immoffset, num_channels, dfmt, nfmt, cache_policy,
2139 true);
2140 }
2141
2142 void
2143 ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
2144 LLVMValueRef rsrc,
2145 LLVMValueRef vdata,
2146 LLVMValueRef voffset,
2147 LLVMValueRef soffset,
2148 LLVMValueRef immoffset,
2149 unsigned num_channels,
2150 unsigned dfmt,
2151 unsigned nfmt,
2152 unsigned cache_policy)
2153 {
2154 ac_build_tbuffer_store(ctx, rsrc, vdata, NULL, voffset, soffset,
2155 immoffset, num_channels, dfmt, nfmt, cache_policy,
2156 false);
2157 }
2158
2159 void
2160 ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
2161 LLVMValueRef rsrc,
2162 LLVMValueRef vdata,
2163 LLVMValueRef voffset,
2164 LLVMValueRef soffset,
2165 unsigned cache_policy)
2166 {
2167 vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i16, "");
2168
2169 if (HAVE_LLVM >= 0x900) {
2170 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
2171 ac_build_llvm8_buffer_store_common(ctx, rsrc, vdata, NULL,
2172 voffset, soffset, 1,
2173 ctx->i16, cache_policy,
2174 false, false);
2175 } else {
2176 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
2177 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
2178
2179 vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
2180
2181 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
2182 ctx->i32_0, 1, dfmt, nfmt, cache_policy);
2183 }
2184 }
2185
2186 void
2187 ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
2188 LLVMValueRef rsrc,
2189 LLVMValueRef vdata,
2190 LLVMValueRef voffset,
2191 LLVMValueRef soffset,
2192 unsigned cache_policy)
2193 {
2194 vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i8, "");
2195
2196 if (HAVE_LLVM >= 0x900) {
2197 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
2198 ac_build_llvm8_buffer_store_common(ctx, rsrc, vdata, NULL,
2199 voffset, soffset, 1,
2200 ctx->i8, cache_policy,
2201 false, false);
2202 } else {
2203 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
2204 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
2205
2206 vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
2207
2208 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
2209 ctx->i32_0, 1, dfmt, nfmt, cache_policy);
2210 }
2211 }
2212 /**
2213 * Set range metadata on an instruction. This can only be used on load and
2214 * call instructions. If you know an instruction can only produce the values
2215 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
2216 * \p lo is the minimum value inclusive.
2217 * \p hi is the maximum value exclusive.
2218 */
2219 static void set_range_metadata(struct ac_llvm_context *ctx,
2220 LLVMValueRef value, unsigned lo, unsigned hi)
2221 {
2222 LLVMValueRef range_md, md_args[2];
2223 LLVMTypeRef type = LLVMTypeOf(value);
2224 LLVMContextRef context = LLVMGetTypeContext(type);
2225
2226 md_args[0] = LLVMConstInt(type, lo, false);
2227 md_args[1] = LLVMConstInt(type, hi, false);
2228 range_md = LLVMMDNodeInContext(context, md_args, 2);
2229 LLVMSetMetadata(value, ctx->range_md_kind, range_md);
2230 }
2231
2232 LLVMValueRef
2233 ac_get_thread_id(struct ac_llvm_context *ctx)
2234 {
2235 LLVMValueRef tid;
2236
2237 LLVMValueRef tid_args[2];
2238 tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
2239 tid_args[1] = ctx->i32_0;
2240 tid_args[1] = ac_build_intrinsic(ctx,
2241 "llvm.amdgcn.mbcnt.lo", ctx->i32,
2242 tid_args, 2, AC_FUNC_ATTR_READNONE);
2243
2244 if (ctx->wave_size == 32) {
2245 tid = tid_args[1];
2246 } else {
2247 tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
2248 ctx->i32, tid_args,
2249 2, AC_FUNC_ATTR_READNONE);
2250 }
2251 set_range_metadata(ctx, tid, 0, ctx->wave_size);
2252 return tid;
2253 }
2254
2255 /*
2256 * AMD GCN implements derivatives using the local data store (LDS)
2257 * All writes to the LDS happen in all executing threads at
2258 * the same time. TID is the Thread ID for the current
2259 * thread and is a value between 0 and 63, representing
2260 * the thread's position in the wavefront.
2261 *
2262 * For the pixel shader threads are grouped into quads of four pixels.
2263 * The TIDs of the pixels of a quad are:
2264 *
2265 * +------+------+
2266 * |4n + 0|4n + 1|
2267 * +------+------+
2268 * |4n + 2|4n + 3|
2269 * +------+------+
2270 *
2271 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
2272 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
2273 * the current pixel's column, and masking with 0xfffffffe yields the TID
2274 * of the left pixel of the current pixel's row.
2275 *
2276 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
2277 * adding 2 yields the TID of the pixel below the top pixel.
2278 */
2279 LLVMValueRef
2280 ac_build_ddxy(struct ac_llvm_context *ctx,
2281 uint32_t mask,
2282 int idx,
2283 LLVMValueRef val)
2284 {
2285 unsigned tl_lanes[4], trbl_lanes[4];
2286 char name[32], type[8];
2287 LLVMValueRef tl, trbl;
2288 LLVMTypeRef result_type;
2289 LLVMValueRef result;
2290
2291 result_type = ac_to_float_type(ctx, LLVMTypeOf(val));
2292
2293 if (result_type == ctx->f16)
2294 val = LLVMBuildZExt(ctx->builder, val, ctx->i32, "");
2295
2296 for (unsigned i = 0; i < 4; ++i) {
2297 tl_lanes[i] = i & mask;
2298 trbl_lanes[i] = (i & mask) + idx;
2299 }
2300
2301 tl = ac_build_quad_swizzle(ctx, val,
2302 tl_lanes[0], tl_lanes[1],
2303 tl_lanes[2], tl_lanes[3]);
2304 trbl = ac_build_quad_swizzle(ctx, val,
2305 trbl_lanes[0], trbl_lanes[1],
2306 trbl_lanes[2], trbl_lanes[3]);
2307
2308 if (result_type == ctx->f16) {
2309 tl = LLVMBuildTrunc(ctx->builder, tl, ctx->i16, "");
2310 trbl = LLVMBuildTrunc(ctx->builder, trbl, ctx->i16, "");
2311 }
2312
2313 tl = LLVMBuildBitCast(ctx->builder, tl, result_type, "");
2314 trbl = LLVMBuildBitCast(ctx->builder, trbl, result_type, "");
2315 result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
2316
2317 ac_build_type_name_for_intr(result_type, type, sizeof(type));
2318 snprintf(name, sizeof(name), "llvm.amdgcn.wqm.%s", type);
2319
2320 return ac_build_intrinsic(ctx, name, result_type, &result, 1, 0);
2321 }
2322
2323 void
2324 ac_build_sendmsg(struct ac_llvm_context *ctx,
2325 uint32_t msg,
2326 LLVMValueRef wave_id)
2327 {
2328 LLVMValueRef args[2];
2329 args[0] = LLVMConstInt(ctx->i32, msg, false);
2330 args[1] = wave_id;
2331 ac_build_intrinsic(ctx, "llvm.amdgcn.s.sendmsg", ctx->voidt, args, 2, 0);
2332 }
2333
2334 LLVMValueRef
2335 ac_build_imsb(struct ac_llvm_context *ctx,
2336 LLVMValueRef arg,
2337 LLVMTypeRef dst_type)
2338 {
2339 LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.amdgcn.sffbh.i32",
2340 dst_type, &arg, 1,
2341 AC_FUNC_ATTR_READNONE);
2342
2343 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
2344 * the index from LSB. Invert it by doing "31 - msb". */
2345 msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
2346 msb, "");
2347
2348 LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
2349 LLVMValueRef cond = LLVMBuildOr(ctx->builder,
2350 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
2351 arg, ctx->i32_0, ""),
2352 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
2353 arg, all_ones, ""), "");
2354
2355 return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
2356 }
2357
2358 LLVMValueRef
2359 ac_build_umsb(struct ac_llvm_context *ctx,
2360 LLVMValueRef arg,
2361 LLVMTypeRef dst_type)
2362 {
2363 const char *intrin_name;
2364 LLVMTypeRef type;
2365 LLVMValueRef highest_bit;
2366 LLVMValueRef zero;
2367 unsigned bitsize;
2368
2369 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(arg));
2370 switch (bitsize) {
2371 case 64:
2372 intrin_name = "llvm.ctlz.i64";
2373 type = ctx->i64;
2374 highest_bit = LLVMConstInt(ctx->i64, 63, false);
2375 zero = ctx->i64_0;
2376 break;
2377 case 32:
2378 intrin_name = "llvm.ctlz.i32";
2379 type = ctx->i32;
2380 highest_bit = LLVMConstInt(ctx->i32, 31, false);
2381 zero = ctx->i32_0;
2382 break;
2383 case 16:
2384 intrin_name = "llvm.ctlz.i16";
2385 type = ctx->i16;
2386 highest_bit = LLVMConstInt(ctx->i16, 15, false);
2387 zero = ctx->i16_0;
2388 break;
2389 case 8:
2390 intrin_name = "llvm.ctlz.i8";
2391 type = ctx->i8;
2392 highest_bit = LLVMConstInt(ctx->i8, 7, false);
2393 zero = ctx->i8_0;
2394 break;
2395 default:
2396 unreachable(!"invalid bitsize");
2397 break;
2398 }
2399
2400 LLVMValueRef params[2] = {
2401 arg,
2402 ctx->i1true,
2403 };
2404
2405 LLVMValueRef msb = ac_build_intrinsic(ctx, intrin_name, type,
2406 params, 2,
2407 AC_FUNC_ATTR_READNONE);
2408
2409 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
2410 * the index from LSB. Invert it by doing "31 - msb". */
2411 msb = LLVMBuildSub(ctx->builder, highest_bit, msb, "");
2412
2413 if (bitsize == 64) {
2414 msb = LLVMBuildTrunc(ctx->builder, msb, ctx->i32, "");
2415 } else if (bitsize < 32) {
2416 msb = LLVMBuildSExt(ctx->builder, msb, ctx->i32, "");
2417 }
2418
2419 /* check for zero */
2420 return LLVMBuildSelect(ctx->builder,
2421 LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg, zero, ""),
2422 LLVMConstInt(ctx->i32, -1, true), msb, "");
2423 }
2424
2425 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
2426 LLVMValueRef b)
2427 {
2428 char name[64];
2429 snprintf(name, sizeof(name), "llvm.minnum.f%d", ac_get_elem_bits(ctx, LLVMTypeOf(a)));
2430 LLVMValueRef args[2] = {a, b};
2431 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
2432 AC_FUNC_ATTR_READNONE);
2433 }
2434
2435 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
2436 LLVMValueRef b)
2437 {
2438 char name[64];
2439 snprintf(name, sizeof(name), "llvm.maxnum.f%d", ac_get_elem_bits(ctx, LLVMTypeOf(a)));
2440 LLVMValueRef args[2] = {a, b};
2441 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
2442 AC_FUNC_ATTR_READNONE);
2443 }
2444
2445 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
2446 LLVMValueRef b)
2447 {
2448 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSLE, a, b, "");
2449 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2450 }
2451
2452 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
2453 LLVMValueRef b)
2454 {
2455 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, a, b, "");
2456 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2457 }
2458
2459 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a,
2460 LLVMValueRef b)
2461 {
2462 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntULE, a, b, "");
2463 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2464 }
2465
2466 LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a,
2467 LLVMValueRef b)
2468 {
2469 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, a, b, "");
2470 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2471 }
2472
2473 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value)
2474 {
2475 LLVMTypeRef t = LLVMTypeOf(value);
2476 return ac_build_fmin(ctx, ac_build_fmax(ctx, value, LLVMConstReal(t, 0.0)),
2477 LLVMConstReal(t, 1.0));
2478 }
2479
2480 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a)
2481 {
2482 LLVMValueRef args[9];
2483
2484 args[0] = LLVMConstInt(ctx->i32, a->target, 0);
2485 args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
2486
2487 if (a->compr) {
2488 LLVMTypeRef i16 = LLVMInt16TypeInContext(ctx->context);
2489 LLVMTypeRef v2i16 = LLVMVectorType(i16, 2);
2490
2491 args[2] = LLVMBuildBitCast(ctx->builder, a->out[0],
2492 v2i16, "");
2493 args[3] = LLVMBuildBitCast(ctx->builder, a->out[1],
2494 v2i16, "");
2495 args[4] = LLVMConstInt(ctx->i1, a->done, 0);
2496 args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
2497
2498 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16",
2499 ctx->voidt, args, 6, 0);
2500 } else {
2501 args[2] = a->out[0];
2502 args[3] = a->out[1];
2503 args[4] = a->out[2];
2504 args[5] = a->out[3];
2505 args[6] = LLVMConstInt(ctx->i1, a->done, 0);
2506 args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
2507
2508 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.f32",
2509 ctx->voidt, args, 8, 0);
2510 }
2511 }
2512
2513 void ac_build_export_null(struct ac_llvm_context *ctx)
2514 {
2515 struct ac_export_args args;
2516
2517 args.enabled_channels = 0x0; /* enabled channels */
2518 args.valid_mask = 1; /* whether the EXEC mask is valid */
2519 args.done = 1; /* DONE bit */
2520 args.target = V_008DFC_SQ_EXP_NULL;
2521 args.compr = 0; /* COMPR flag (0 = 32-bit export) */
2522 args.out[0] = LLVMGetUndef(ctx->f32); /* R */
2523 args.out[1] = LLVMGetUndef(ctx->f32); /* G */
2524 args.out[2] = LLVMGetUndef(ctx->f32); /* B */
2525 args.out[3] = LLVMGetUndef(ctx->f32); /* A */
2526
2527 ac_build_export(ctx, &args);
2528 }
2529
2530 static unsigned ac_num_coords(enum ac_image_dim dim)
2531 {
2532 switch (dim) {
2533 case ac_image_1d:
2534 return 1;
2535 case ac_image_2d:
2536 case ac_image_1darray:
2537 return 2;
2538 case ac_image_3d:
2539 case ac_image_cube:
2540 case ac_image_2darray:
2541 case ac_image_2dmsaa:
2542 return 3;
2543 case ac_image_2darraymsaa:
2544 return 4;
2545 default:
2546 unreachable("ac_num_coords: bad dim");
2547 }
2548 }
2549
2550 static unsigned ac_num_derivs(enum ac_image_dim dim)
2551 {
2552 switch (dim) {
2553 case ac_image_1d:
2554 case ac_image_1darray:
2555 return 2;
2556 case ac_image_2d:
2557 case ac_image_2darray:
2558 case ac_image_cube:
2559 return 4;
2560 case ac_image_3d:
2561 return 6;
2562 case ac_image_2dmsaa:
2563 case ac_image_2darraymsaa:
2564 default:
2565 unreachable("derivatives not supported");
2566 }
2567 }
2568
2569 static const char *get_atomic_name(enum ac_atomic_op op)
2570 {
2571 switch (op) {
2572 case ac_atomic_swap: return "swap";
2573 case ac_atomic_add: return "add";
2574 case ac_atomic_sub: return "sub";
2575 case ac_atomic_smin: return "smin";
2576 case ac_atomic_umin: return "umin";
2577 case ac_atomic_smax: return "smax";
2578 case ac_atomic_umax: return "umax";
2579 case ac_atomic_and: return "and";
2580 case ac_atomic_or: return "or";
2581 case ac_atomic_xor: return "xor";
2582 }
2583 unreachable("bad atomic op");
2584 }
2585
2586 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
2587 struct ac_image_args *a)
2588 {
2589 const char *overload[3] = { "", "", "" };
2590 unsigned num_overloads = 0;
2591 LLVMValueRef args[18];
2592 unsigned num_args = 0;
2593 enum ac_image_dim dim = a->dim;
2594
2595 assert(!a->lod || a->lod == ctx->i32_0 || a->lod == ctx->f32_0 ||
2596 !a->level_zero);
2597 assert((a->opcode != ac_image_get_resinfo && a->opcode != ac_image_load_mip &&
2598 a->opcode != ac_image_store_mip) ||
2599 a->lod);
2600 assert(a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
2601 (!a->compare && !a->offset));
2602 assert((a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
2603 a->opcode == ac_image_get_lod) ||
2604 !a->bias);
2605 assert((a->bias ? 1 : 0) +
2606 (a->lod ? 1 : 0) +
2607 (a->level_zero ? 1 : 0) +
2608 (a->derivs[0] ? 1 : 0) <= 1);
2609
2610 if (a->opcode == ac_image_get_lod) {
2611 switch (dim) {
2612 case ac_image_1darray:
2613 dim = ac_image_1d;
2614 break;
2615 case ac_image_2darray:
2616 case ac_image_cube:
2617 dim = ac_image_2d;
2618 break;
2619 default:
2620 break;
2621 }
2622 }
2623
2624 bool sample = a->opcode == ac_image_sample ||
2625 a->opcode == ac_image_gather4 ||
2626 a->opcode == ac_image_get_lod;
2627 bool atomic = a->opcode == ac_image_atomic ||
2628 a->opcode == ac_image_atomic_cmpswap;
2629 bool load = a->opcode == ac_image_sample ||
2630 a->opcode == ac_image_gather4 ||
2631 a->opcode == ac_image_load ||
2632 a->opcode == ac_image_load_mip;
2633 LLVMTypeRef coord_type = sample ? ctx->f32 : ctx->i32;
2634
2635 if (atomic || a->opcode == ac_image_store || a->opcode == ac_image_store_mip) {
2636 args[num_args++] = a->data[0];
2637 if (a->opcode == ac_image_atomic_cmpswap)
2638 args[num_args++] = a->data[1];
2639 }
2640
2641 if (!atomic)
2642 args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, false);
2643
2644 if (a->offset)
2645 args[num_args++] = ac_to_integer(ctx, a->offset);
2646 if (a->bias) {
2647 args[num_args++] = ac_to_float(ctx, a->bias);
2648 overload[num_overloads++] = ".f32";
2649 }
2650 if (a->compare)
2651 args[num_args++] = ac_to_float(ctx, a->compare);
2652 if (a->derivs[0]) {
2653 unsigned count = ac_num_derivs(dim);
2654 for (unsigned i = 0; i < count; ++i)
2655 args[num_args++] = ac_to_float(ctx, a->derivs[i]);
2656 overload[num_overloads++] = ".f32";
2657 }
2658 unsigned num_coords =
2659 a->opcode != ac_image_get_resinfo ? ac_num_coords(dim) : 0;
2660 for (unsigned i = 0; i < num_coords; ++i)
2661 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->coords[i], coord_type, "");
2662 if (a->lod)
2663 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->lod, coord_type, "");
2664 overload[num_overloads++] = sample ? ".f32" : ".i32";
2665
2666 args[num_args++] = a->resource;
2667 if (sample) {
2668 args[num_args++] = a->sampler;
2669 args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, false);
2670 }
2671
2672 args[num_args++] = ctx->i32_0; /* texfailctrl */
2673 args[num_args++] = LLVMConstInt(ctx->i32,
2674 load ? get_load_cache_policy(ctx, a->cache_policy) :
2675 a->cache_policy, false);
2676
2677 const char *name;
2678 const char *atomic_subop = "";
2679 switch (a->opcode) {
2680 case ac_image_sample: name = "sample"; break;
2681 case ac_image_gather4: name = "gather4"; break;
2682 case ac_image_load: name = "load"; break;
2683 case ac_image_load_mip: name = "load.mip"; break;
2684 case ac_image_store: name = "store"; break;
2685 case ac_image_store_mip: name = "store.mip"; break;
2686 case ac_image_atomic:
2687 name = "atomic.";
2688 atomic_subop = get_atomic_name(a->atomic);
2689 break;
2690 case ac_image_atomic_cmpswap:
2691 name = "atomic.";
2692 atomic_subop = "cmpswap";
2693 break;
2694 case ac_image_get_lod: name = "getlod"; break;
2695 case ac_image_get_resinfo: name = "getresinfo"; break;
2696 default: unreachable("invalid image opcode");
2697 }
2698
2699 const char *dimname;
2700 switch (dim) {
2701 case ac_image_1d: dimname = "1d"; break;
2702 case ac_image_2d: dimname = "2d"; break;
2703 case ac_image_3d: dimname = "3d"; break;
2704 case ac_image_cube: dimname = "cube"; break;
2705 case ac_image_1darray: dimname = "1darray"; break;
2706 case ac_image_2darray: dimname = "2darray"; break;
2707 case ac_image_2dmsaa: dimname = "2dmsaa"; break;
2708 case ac_image_2darraymsaa: dimname = "2darraymsaa"; break;
2709 default: unreachable("invalid dim");
2710 }
2711
2712 bool lod_suffix =
2713 a->lod && (a->opcode == ac_image_sample || a->opcode == ac_image_gather4);
2714 char intr_name[96];
2715 snprintf(intr_name, sizeof(intr_name),
2716 "llvm.amdgcn.image.%s%s" /* base name */
2717 "%s%s%s" /* sample/gather modifiers */
2718 ".%s.%s%s%s%s", /* dimension and type overloads */
2719 name, atomic_subop,
2720 a->compare ? ".c" : "",
2721 a->bias ? ".b" :
2722 lod_suffix ? ".l" :
2723 a->derivs[0] ? ".d" :
2724 a->level_zero ? ".lz" : "",
2725 a->offset ? ".o" : "",
2726 dimname,
2727 atomic ? "i32" : "v4f32",
2728 overload[0], overload[1], overload[2]);
2729
2730 LLVMTypeRef retty;
2731 if (atomic)
2732 retty = ctx->i32;
2733 else if (a->opcode == ac_image_store || a->opcode == ac_image_store_mip)
2734 retty = ctx->voidt;
2735 else
2736 retty = ctx->v4f32;
2737
2738 LLVMValueRef result =
2739 ac_build_intrinsic(ctx, intr_name, retty, args, num_args,
2740 a->attributes);
2741 if (!sample && retty == ctx->v4f32) {
2742 result = LLVMBuildBitCast(ctx->builder, result,
2743 ctx->v4i32, "");
2744 }
2745 return result;
2746 }
2747
2748 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
2749 LLVMValueRef args[2])
2750 {
2751 LLVMTypeRef v2f16 =
2752 LLVMVectorType(LLVMHalfTypeInContext(ctx->context), 2);
2753
2754 return ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz", v2f16,
2755 args, 2, AC_FUNC_ATTR_READNONE);
2756 }
2757
2758 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
2759 LLVMValueRef args[2])
2760 {
2761 LLVMValueRef res =
2762 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16",
2763 ctx->v2i16, args, 2,
2764 AC_FUNC_ATTR_READNONE);
2765 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2766 }
2767
2768 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
2769 LLVMValueRef args[2])
2770 {
2771 LLVMValueRef res =
2772 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16",
2773 ctx->v2i16, args, 2,
2774 AC_FUNC_ATTR_READNONE);
2775 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2776 }
2777
2778 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2779 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
2780 LLVMValueRef args[2], unsigned bits, bool hi)
2781 {
2782 assert(bits == 8 || bits == 10 || bits == 16);
2783
2784 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2785 bits == 8 ? 127 : bits == 10 ? 511 : 32767, 0);
2786 LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
2787 bits == 8 ? -128 : bits == 10 ? -512 : -32768, 0);
2788 LLVMValueRef max_alpha =
2789 bits != 10 ? max_rgb : ctx->i32_1;
2790 LLVMValueRef min_alpha =
2791 bits != 10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
2792
2793 /* Clamp. */
2794 if (bits != 16) {
2795 for (int i = 0; i < 2; i++) {
2796 bool alpha = hi && i == 1;
2797 args[i] = ac_build_imin(ctx, args[i],
2798 alpha ? max_alpha : max_rgb);
2799 args[i] = ac_build_imax(ctx, args[i],
2800 alpha ? min_alpha : min_rgb);
2801 }
2802 }
2803
2804 LLVMValueRef res =
2805 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16",
2806 ctx->v2i16, args, 2,
2807 AC_FUNC_ATTR_READNONE);
2808 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2809 }
2810
2811 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2812 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
2813 LLVMValueRef args[2], unsigned bits, bool hi)
2814 {
2815 assert(bits == 8 || bits == 10 || bits == 16);
2816
2817 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2818 bits == 8 ? 255 : bits == 10 ? 1023 : 65535, 0);
2819 LLVMValueRef max_alpha =
2820 bits != 10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
2821
2822 /* Clamp. */
2823 if (bits != 16) {
2824 for (int i = 0; i < 2; i++) {
2825 bool alpha = hi && i == 1;
2826 args[i] = ac_build_umin(ctx, args[i],
2827 alpha ? max_alpha : max_rgb);
2828 }
2829 }
2830
2831 LLVMValueRef res =
2832 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16",
2833 ctx->v2i16, args, 2,
2834 AC_FUNC_ATTR_READNONE);
2835 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2836 }
2837
2838 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1)
2839 {
2840 return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1,
2841 &i1, 1, AC_FUNC_ATTR_READNONE);
2842 }
2843
2844 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1)
2845 {
2846 ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt,
2847 &i1, 1, 0);
2848 }
2849
2850 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
2851 LLVMValueRef offset, LLVMValueRef width,
2852 bool is_signed)
2853 {
2854 LLVMValueRef args[] = {
2855 input,
2856 offset,
2857 width,
2858 };
2859
2860 LLVMValueRef result = ac_build_intrinsic(ctx,
2861 is_signed ? "llvm.amdgcn.sbfe.i32" :
2862 "llvm.amdgcn.ubfe.i32",
2863 ctx->i32, args, 3,
2864 AC_FUNC_ATTR_READNONE);
2865
2866 if (HAVE_LLVM < 0x0800) {
2867 /* FIXME: LLVM 7+ returns incorrect result when count is 0.
2868 * https://bugs.freedesktop.org/show_bug.cgi?id=107276
2869 */
2870 LLVMValueRef zero = ctx->i32_0;
2871 LLVMValueRef icond = LLVMBuildICmp(ctx->builder, LLVMIntEQ, width, zero, "");
2872 result = LLVMBuildSelect(ctx->builder, icond, zero, result, "");
2873 }
2874
2875 return result;
2876 }
2877
2878 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2879 LLVMValueRef s1, LLVMValueRef s2)
2880 {
2881 return LLVMBuildAdd(ctx->builder,
2882 LLVMBuildMul(ctx->builder, s0, s1, ""), s2, "");
2883 }
2884
2885 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2886 LLVMValueRef s1, LLVMValueRef s2)
2887 {
2888 return LLVMBuildFAdd(ctx->builder,
2889 LLVMBuildFMul(ctx->builder, s0, s1, ""), s2, "");
2890 }
2891
2892 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags)
2893 {
2894 if (!wait_flags)
2895 return;
2896
2897 unsigned lgkmcnt = 63;
2898 unsigned vmcnt = ctx->chip_class >= GFX9 ? 63 : 15;
2899 unsigned vscnt = 63;
2900
2901 if (wait_flags & AC_WAIT_LGKM)
2902 lgkmcnt = 0;
2903 if (wait_flags & AC_WAIT_VLOAD)
2904 vmcnt = 0;
2905
2906 if (wait_flags & AC_WAIT_VSTORE) {
2907 if (ctx->chip_class >= GFX10)
2908 vscnt = 0;
2909 else
2910 vmcnt = 0;
2911 }
2912
2913 /* There is no intrinsic for vscnt(0), so use a fence. */
2914 if ((wait_flags & AC_WAIT_LGKM &&
2915 wait_flags & AC_WAIT_VLOAD &&
2916 wait_flags & AC_WAIT_VSTORE) ||
2917 vscnt == 0) {
2918 LLVMBuildFence(ctx->builder, LLVMAtomicOrderingRelease, false, "");
2919 return;
2920 }
2921
2922 unsigned simm16 = (lgkmcnt << 8) |
2923 (7 << 4) | /* expcnt */
2924 (vmcnt & 0xf) |
2925 ((vmcnt >> 4) << 14);
2926
2927 LLVMValueRef args[1] = {
2928 LLVMConstInt(ctx->i32, simm16, false),
2929 };
2930 ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt",
2931 ctx->voidt, args, 1, 0);
2932 }
2933
2934 LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
2935 LLVMValueRef src1, LLVMValueRef src2,
2936 unsigned bitsize)
2937 {
2938 LLVMTypeRef type;
2939 char *intr;
2940
2941 if (bitsize == 16) {
2942 intr = "llvm.amdgcn.fmed3.f16";
2943 type = ctx->f16;
2944 } else if (bitsize == 32) {
2945 intr = "llvm.amdgcn.fmed3.f32";
2946 type = ctx->f32;
2947 } else {
2948 intr = "llvm.amdgcn.fmed3.f64";
2949 type = ctx->f64;
2950 }
2951
2952 LLVMValueRef params[] = {
2953 src0,
2954 src1,
2955 src2,
2956 };
2957 return ac_build_intrinsic(ctx, intr, type, params, 3,
2958 AC_FUNC_ATTR_READNONE);
2959 }
2960
2961 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
2962 unsigned bitsize)
2963 {
2964 LLVMTypeRef type;
2965 char *intr;
2966
2967 if (bitsize == 16) {
2968 intr = "llvm.amdgcn.fract.f16";
2969 type = ctx->f16;
2970 } else if (bitsize == 32) {
2971 intr = "llvm.amdgcn.fract.f32";
2972 type = ctx->f32;
2973 } else {
2974 intr = "llvm.amdgcn.fract.f64";
2975 type = ctx->f64;
2976 }
2977
2978 LLVMValueRef params[] = {
2979 src0,
2980 };
2981 return ac_build_intrinsic(ctx, intr, type, params, 1,
2982 AC_FUNC_ATTR_READNONE);
2983 }
2984
2985 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2986 unsigned bitsize)
2987 {
2988 LLVMTypeRef type = LLVMIntTypeInContext(ctx->context, bitsize);
2989 LLVMValueRef zero = LLVMConstInt(type, 0, false);
2990 LLVMValueRef one = LLVMConstInt(type, 1, false);
2991
2992 LLVMValueRef cmp, val;
2993 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, src0, zero, "");
2994 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2995 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGE, val, zero, "");
2996 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstInt(type, -1, true), "");
2997 return val;
2998 }
2999
3000 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
3001 unsigned bitsize)
3002 {
3003 LLVMValueRef cmp, val, zero, one;
3004 LLVMTypeRef type;
3005
3006 if (bitsize == 16) {
3007 type = ctx->f16;
3008 zero = ctx->f16_0;
3009 one = ctx->f16_1;
3010 } else if (bitsize == 32) {
3011 type = ctx->f32;
3012 zero = ctx->f32_0;
3013 one = ctx->f32_1;
3014 } else {
3015 type = ctx->f64;
3016 zero = ctx->f64_0;
3017 one = ctx->f64_1;
3018 }
3019
3020 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src0, zero, "");
3021 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
3022 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGE, val, zero, "");
3023 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstReal(type, -1.0), "");
3024 return val;
3025 }
3026
3027 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0)
3028 {
3029 LLVMValueRef result;
3030 unsigned bitsize;
3031
3032 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
3033
3034 switch (bitsize) {
3035 case 64:
3036 result = ac_build_intrinsic(ctx, "llvm.ctpop.i64", ctx->i64,
3037 (LLVMValueRef []) { src0 }, 1,
3038 AC_FUNC_ATTR_READNONE);
3039
3040 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
3041 break;
3042 case 32:
3043 result = ac_build_intrinsic(ctx, "llvm.ctpop.i32", ctx->i32,
3044 (LLVMValueRef []) { src0 }, 1,
3045 AC_FUNC_ATTR_READNONE);
3046 break;
3047 case 16:
3048 result = ac_build_intrinsic(ctx, "llvm.ctpop.i16", ctx->i16,
3049 (LLVMValueRef []) { src0 }, 1,
3050 AC_FUNC_ATTR_READNONE);
3051
3052 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
3053 break;
3054 case 8:
3055 result = ac_build_intrinsic(ctx, "llvm.ctpop.i8", ctx->i8,
3056 (LLVMValueRef []) { src0 }, 1,
3057 AC_FUNC_ATTR_READNONE);
3058
3059 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
3060 break;
3061 default:
3062 unreachable(!"invalid bitsize");
3063 break;
3064 }
3065
3066 return result;
3067 }
3068
3069 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
3070 LLVMValueRef src0)
3071 {
3072 LLVMValueRef result;
3073 unsigned bitsize;
3074
3075 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
3076
3077 switch (bitsize) {
3078 case 64:
3079 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i64", ctx->i64,
3080 (LLVMValueRef []) { src0 }, 1,
3081 AC_FUNC_ATTR_READNONE);
3082
3083 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
3084 break;
3085 case 32:
3086 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i32", ctx->i32,
3087 (LLVMValueRef []) { src0 }, 1,
3088 AC_FUNC_ATTR_READNONE);
3089 break;
3090 case 16:
3091 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i16", ctx->i16,
3092 (LLVMValueRef []) { src0 }, 1,
3093 AC_FUNC_ATTR_READNONE);
3094
3095 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
3096 break;
3097 case 8:
3098 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i8", ctx->i8,
3099 (LLVMValueRef []) { src0 }, 1,
3100 AC_FUNC_ATTR_READNONE);
3101
3102 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
3103 break;
3104 default:
3105 unreachable(!"invalid bitsize");
3106 break;
3107 }
3108
3109 return result;
3110 }
3111
3112 #define AC_EXP_TARGET 0
3113 #define AC_EXP_ENABLED_CHANNELS 1
3114 #define AC_EXP_OUT0 2
3115
3116 enum ac_ir_type {
3117 AC_IR_UNDEF,
3118 AC_IR_CONST,
3119 AC_IR_VALUE,
3120 };
3121
3122 struct ac_vs_exp_chan
3123 {
3124 LLVMValueRef value;
3125 float const_float;
3126 enum ac_ir_type type;
3127 };
3128
3129 struct ac_vs_exp_inst {
3130 unsigned offset;
3131 LLVMValueRef inst;
3132 struct ac_vs_exp_chan chan[4];
3133 };
3134
3135 struct ac_vs_exports {
3136 unsigned num;
3137 struct ac_vs_exp_inst exp[VARYING_SLOT_MAX];
3138 };
3139
3140 /* Return true if the PARAM export has been eliminated. */
3141 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset,
3142 uint32_t num_outputs,
3143 struct ac_vs_exp_inst *exp)
3144 {
3145 unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
3146 bool is_zero[4] = {}, is_one[4] = {};
3147
3148 for (i = 0; i < 4; i++) {
3149 /* It's a constant expression. Undef outputs are eliminated too. */
3150 if (exp->chan[i].type == AC_IR_UNDEF) {
3151 is_zero[i] = true;
3152 is_one[i] = true;
3153 } else if (exp->chan[i].type == AC_IR_CONST) {
3154 if (exp->chan[i].const_float == 0)
3155 is_zero[i] = true;
3156 else if (exp->chan[i].const_float == 1)
3157 is_one[i] = true;
3158 else
3159 return false; /* other constant */
3160 } else
3161 return false;
3162 }
3163
3164 /* Only certain combinations of 0 and 1 can be eliminated. */
3165 if (is_zero[0] && is_zero[1] && is_zero[2])
3166 default_val = is_zero[3] ? 0 : 1;
3167 else if (is_one[0] && is_one[1] && is_one[2])
3168 default_val = is_zero[3] ? 2 : 3;
3169 else
3170 return false;
3171
3172 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
3173 LLVMInstructionEraseFromParent(exp->inst);
3174
3175 /* Change OFFSET to DEFAULT_VAL. */
3176 for (i = 0; i < num_outputs; i++) {
3177 if (vs_output_param_offset[i] == exp->offset) {
3178 vs_output_param_offset[i] =
3179 AC_EXP_PARAM_DEFAULT_VAL_0000 + default_val;
3180 break;
3181 }
3182 }
3183 return true;
3184 }
3185
3186 static bool ac_eliminate_duplicated_output(struct ac_llvm_context *ctx,
3187 uint8_t *vs_output_param_offset,
3188 uint32_t num_outputs,
3189 struct ac_vs_exports *processed,
3190 struct ac_vs_exp_inst *exp)
3191 {
3192 unsigned p, copy_back_channels = 0;
3193
3194 /* See if the output is already in the list of processed outputs.
3195 * The LLVMValueRef comparison relies on SSA.
3196 */
3197 for (p = 0; p < processed->num; p++) {
3198 bool different = false;
3199
3200 for (unsigned j = 0; j < 4; j++) {
3201 struct ac_vs_exp_chan *c1 = &processed->exp[p].chan[j];
3202 struct ac_vs_exp_chan *c2 = &exp->chan[j];
3203
3204 /* Treat undef as a match. */
3205 if (c2->type == AC_IR_UNDEF)
3206 continue;
3207
3208 /* If c1 is undef but c2 isn't, we can copy c2 to c1
3209 * and consider the instruction duplicated.
3210 */
3211 if (c1->type == AC_IR_UNDEF) {
3212 copy_back_channels |= 1 << j;
3213 continue;
3214 }
3215
3216 /* Test whether the channels are not equal. */
3217 if (c1->type != c2->type ||
3218 (c1->type == AC_IR_CONST &&
3219 c1->const_float != c2->const_float) ||
3220 (c1->type == AC_IR_VALUE &&
3221 c1->value != c2->value)) {
3222 different = true;
3223 break;
3224 }
3225 }
3226 if (!different)
3227 break;
3228
3229 copy_back_channels = 0;
3230 }
3231 if (p == processed->num)
3232 return false;
3233
3234 /* If a match was found, but the matching export has undef where the new
3235 * one has a normal value, copy the normal value to the undef channel.
3236 */
3237 struct ac_vs_exp_inst *match = &processed->exp[p];
3238
3239 /* Get current enabled channels mask. */
3240 LLVMValueRef arg = LLVMGetOperand(match->inst, AC_EXP_ENABLED_CHANNELS);
3241 unsigned enabled_channels = LLVMConstIntGetZExtValue(arg);
3242
3243 while (copy_back_channels) {
3244 unsigned chan = u_bit_scan(&copy_back_channels);
3245
3246 assert(match->chan[chan].type == AC_IR_UNDEF);
3247 LLVMSetOperand(match->inst, AC_EXP_OUT0 + chan,
3248 exp->chan[chan].value);
3249 match->chan[chan] = exp->chan[chan];
3250
3251 /* Update number of enabled channels because the original mask
3252 * is not always 0xf.
3253 */
3254 enabled_channels |= (1 << chan);
3255 LLVMSetOperand(match->inst, AC_EXP_ENABLED_CHANNELS,
3256 LLVMConstInt(ctx->i32, enabled_channels, 0));
3257 }
3258
3259 /* The PARAM export is duplicated. Kill it. */
3260 LLVMInstructionEraseFromParent(exp->inst);
3261
3262 /* Change OFFSET to the matching export. */
3263 for (unsigned i = 0; i < num_outputs; i++) {
3264 if (vs_output_param_offset[i] == exp->offset) {
3265 vs_output_param_offset[i] = match->offset;
3266 break;
3267 }
3268 }
3269 return true;
3270 }
3271
3272 void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
3273 LLVMValueRef main_fn,
3274 uint8_t *vs_output_param_offset,
3275 uint32_t num_outputs,
3276 uint8_t *num_param_exports)
3277 {
3278 LLVMBasicBlockRef bb;
3279 bool removed_any = false;
3280 struct ac_vs_exports exports;
3281
3282 exports.num = 0;
3283
3284 /* Process all LLVM instructions. */
3285 bb = LLVMGetFirstBasicBlock(main_fn);
3286 while (bb) {
3287 LLVMValueRef inst = LLVMGetFirstInstruction(bb);
3288
3289 while (inst) {
3290 LLVMValueRef cur = inst;
3291 inst = LLVMGetNextInstruction(inst);
3292 struct ac_vs_exp_inst exp;
3293
3294 if (LLVMGetInstructionOpcode(cur) != LLVMCall)
3295 continue;
3296
3297 LLVMValueRef callee = ac_llvm_get_called_value(cur);
3298
3299 if (!ac_llvm_is_function(callee))
3300 continue;
3301
3302 const char *name = LLVMGetValueName(callee);
3303 unsigned num_args = LLVMCountParams(callee);
3304
3305 /* Check if this is an export instruction. */
3306 if ((num_args != 9 && num_args != 8) ||
3307 (strcmp(name, "llvm.SI.export") &&
3308 strcmp(name, "llvm.amdgcn.exp.f32")))
3309 continue;
3310
3311 LLVMValueRef arg = LLVMGetOperand(cur, AC_EXP_TARGET);
3312 unsigned target = LLVMConstIntGetZExtValue(arg);
3313
3314 if (target < V_008DFC_SQ_EXP_PARAM)
3315 continue;
3316
3317 target -= V_008DFC_SQ_EXP_PARAM;
3318
3319 /* Parse the instruction. */
3320 memset(&exp, 0, sizeof(exp));
3321 exp.offset = target;
3322 exp.inst = cur;
3323
3324 for (unsigned i = 0; i < 4; i++) {
3325 LLVMValueRef v = LLVMGetOperand(cur, AC_EXP_OUT0 + i);
3326
3327 exp.chan[i].value = v;
3328
3329 if (LLVMIsUndef(v)) {
3330 exp.chan[i].type = AC_IR_UNDEF;
3331 } else if (LLVMIsAConstantFP(v)) {
3332 LLVMBool loses_info;
3333 exp.chan[i].type = AC_IR_CONST;
3334 exp.chan[i].const_float =
3335 LLVMConstRealGetDouble(v, &loses_info);
3336 } else {
3337 exp.chan[i].type = AC_IR_VALUE;
3338 }
3339 }
3340
3341 /* Eliminate constant and duplicated PARAM exports. */
3342 if (ac_eliminate_const_output(vs_output_param_offset,
3343 num_outputs, &exp) ||
3344 ac_eliminate_duplicated_output(ctx,
3345 vs_output_param_offset,
3346 num_outputs, &exports,
3347 &exp)) {
3348 removed_any = true;
3349 } else {
3350 exports.exp[exports.num++] = exp;
3351 }
3352 }
3353 bb = LLVMGetNextBasicBlock(bb);
3354 }
3355
3356 /* Remove holes in export memory due to removed PARAM exports.
3357 * This is done by renumbering all PARAM exports.
3358 */
3359 if (removed_any) {
3360 uint8_t old_offset[VARYING_SLOT_MAX];
3361 unsigned out, i;
3362
3363 /* Make a copy of the offsets. We need the old version while
3364 * we are modifying some of them. */
3365 memcpy(old_offset, vs_output_param_offset,
3366 sizeof(old_offset));
3367
3368 for (i = 0; i < exports.num; i++) {
3369 unsigned offset = exports.exp[i].offset;
3370
3371 /* Update vs_output_param_offset. Multiple outputs can
3372 * have the same offset.
3373 */
3374 for (out = 0; out < num_outputs; out++) {
3375 if (old_offset[out] == offset)
3376 vs_output_param_offset[out] = i;
3377 }
3378
3379 /* Change the PARAM offset in the instruction. */
3380 LLVMSetOperand(exports.exp[i].inst, AC_EXP_TARGET,
3381 LLVMConstInt(ctx->i32,
3382 V_008DFC_SQ_EXP_PARAM + i, 0));
3383 }
3384 *num_param_exports = exports.num;
3385 }
3386 }
3387
3388 void ac_init_exec_full_mask(struct ac_llvm_context *ctx)
3389 {
3390 LLVMValueRef full_mask = LLVMConstInt(ctx->i64, ~0ull, 0);
3391 ac_build_intrinsic(ctx,
3392 "llvm.amdgcn.init.exec", ctx->voidt,
3393 &full_mask, 1, AC_FUNC_ATTR_CONVERGENT);
3394 }
3395
3396 void ac_declare_lds_as_pointer(struct ac_llvm_context *ctx)
3397 {
3398 unsigned lds_size = ctx->chip_class >= GFX7 ? 65536 : 32768;
3399 ctx->lds = LLVMBuildIntToPtr(ctx->builder, ctx->i32_0,
3400 LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), AC_ADDR_SPACE_LDS),
3401 "lds");
3402 }
3403
3404 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
3405 LLVMValueRef dw_addr)
3406 {
3407 return LLVMBuildLoad(ctx->builder, ac_build_gep0(ctx, ctx->lds, dw_addr), "");
3408 }
3409
3410 void ac_lds_store(struct ac_llvm_context *ctx,
3411 LLVMValueRef dw_addr,
3412 LLVMValueRef value)
3413 {
3414 value = ac_to_integer(ctx, value);
3415 ac_build_indexed_store(ctx, ctx->lds,
3416 dw_addr, value);
3417 }
3418
3419 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
3420 LLVMTypeRef dst_type,
3421 LLVMValueRef src0)
3422 {
3423 unsigned src0_bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
3424 const char *intrin_name;
3425 LLVMTypeRef type;
3426 LLVMValueRef zero;
3427
3428 switch (src0_bitsize) {
3429 case 64:
3430 intrin_name = "llvm.cttz.i64";
3431 type = ctx->i64;
3432 zero = ctx->i64_0;
3433 break;
3434 case 32:
3435 intrin_name = "llvm.cttz.i32";
3436 type = ctx->i32;
3437 zero = ctx->i32_0;
3438 break;
3439 case 16:
3440 intrin_name = "llvm.cttz.i16";
3441 type = ctx->i16;
3442 zero = ctx->i16_0;
3443 break;
3444 case 8:
3445 intrin_name = "llvm.cttz.i8";
3446 type = ctx->i8;
3447 zero = ctx->i8_0;
3448 break;
3449 default:
3450 unreachable(!"invalid bitsize");
3451 }
3452
3453 LLVMValueRef params[2] = {
3454 src0,
3455
3456 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
3457 * add special code to check for x=0. The reason is that
3458 * the LLVM behavior for x=0 is different from what we
3459 * need here. However, LLVM also assumes that ffs(x) is
3460 * in [0, 31], but GLSL expects that ffs(0) = -1, so
3461 * a conditional assignment to handle 0 is still required.
3462 *
3463 * The hardware already implements the correct behavior.
3464 */
3465 ctx->i1true,
3466 };
3467
3468 LLVMValueRef lsb = ac_build_intrinsic(ctx, intrin_name, type,
3469 params, 2,
3470 AC_FUNC_ATTR_READNONE);
3471
3472 if (src0_bitsize == 64) {
3473 lsb = LLVMBuildTrunc(ctx->builder, lsb, ctx->i32, "");
3474 } else if (src0_bitsize < 32) {
3475 lsb = LLVMBuildSExt(ctx->builder, lsb, ctx->i32, "");
3476 }
3477
3478 /* TODO: We need an intrinsic to skip this conditional. */
3479 /* Check for zero: */
3480 return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder,
3481 LLVMIntEQ, src0,
3482 zero, ""),
3483 LLVMConstInt(ctx->i32, -1, 0), lsb, "");
3484 }
3485
3486 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type)
3487 {
3488 return LLVMPointerType(elem_type, AC_ADDR_SPACE_CONST);
3489 }
3490
3491 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type)
3492 {
3493 return LLVMPointerType(elem_type, AC_ADDR_SPACE_CONST_32BIT);
3494 }
3495
3496 static struct ac_llvm_flow *
3497 get_current_flow(struct ac_llvm_context *ctx)
3498 {
3499 if (ctx->flow->depth > 0)
3500 return &ctx->flow->stack[ctx->flow->depth - 1];
3501 return NULL;
3502 }
3503
3504 static struct ac_llvm_flow *
3505 get_innermost_loop(struct ac_llvm_context *ctx)
3506 {
3507 for (unsigned i = ctx->flow->depth; i > 0; --i) {
3508 if (ctx->flow->stack[i - 1].loop_entry_block)
3509 return &ctx->flow->stack[i - 1];
3510 }
3511 return NULL;
3512 }
3513
3514 static struct ac_llvm_flow *
3515 push_flow(struct ac_llvm_context *ctx)
3516 {
3517 struct ac_llvm_flow *flow;
3518
3519 if (ctx->flow->depth >= ctx->flow->depth_max) {
3520 unsigned new_max = MAX2(ctx->flow->depth << 1,
3521 AC_LLVM_INITIAL_CF_DEPTH);
3522
3523 ctx->flow->stack = realloc(ctx->flow->stack, new_max * sizeof(*ctx->flow->stack));
3524 ctx->flow->depth_max = new_max;
3525 }
3526
3527 flow = &ctx->flow->stack[ctx->flow->depth];
3528 ctx->flow->depth++;
3529
3530 flow->next_block = NULL;
3531 flow->loop_entry_block = NULL;
3532 return flow;
3533 }
3534
3535 static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base,
3536 int label_id)
3537 {
3538 char buf[32];
3539 snprintf(buf, sizeof(buf), "%s%d", base, label_id);
3540 LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
3541 }
3542
3543 /* Append a basic block at the level of the parent flow.
3544 */
3545 static LLVMBasicBlockRef append_basic_block(struct ac_llvm_context *ctx,
3546 const char *name)
3547 {
3548 assert(ctx->flow->depth >= 1);
3549
3550 if (ctx->flow->depth >= 2) {
3551 struct ac_llvm_flow *flow = &ctx->flow->stack[ctx->flow->depth - 2];
3552
3553 return LLVMInsertBasicBlockInContext(ctx->context,
3554 flow->next_block, name);
3555 }
3556
3557 LLVMValueRef main_fn =
3558 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->builder));
3559 return LLVMAppendBasicBlockInContext(ctx->context, main_fn, name);
3560 }
3561
3562 /* Emit a branch to the given default target for the current block if
3563 * applicable -- that is, if the current block does not already contain a
3564 * branch from a break or continue.
3565 */
3566 static void emit_default_branch(LLVMBuilderRef builder,
3567 LLVMBasicBlockRef target)
3568 {
3569 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
3570 LLVMBuildBr(builder, target);
3571 }
3572
3573 void ac_build_bgnloop(struct ac_llvm_context *ctx, int label_id)
3574 {
3575 struct ac_llvm_flow *flow = push_flow(ctx);
3576 flow->loop_entry_block = append_basic_block(ctx, "LOOP");
3577 flow->next_block = append_basic_block(ctx, "ENDLOOP");
3578 set_basicblock_name(flow->loop_entry_block, "loop", label_id);
3579 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
3580 LLVMPositionBuilderAtEnd(ctx->builder, flow->loop_entry_block);
3581 }
3582
3583 void ac_build_break(struct ac_llvm_context *ctx)
3584 {
3585 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
3586 LLVMBuildBr(ctx->builder, flow->next_block);
3587 }
3588
3589 void ac_build_continue(struct ac_llvm_context *ctx)
3590 {
3591 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
3592 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
3593 }
3594
3595 void ac_build_else(struct ac_llvm_context *ctx, int label_id)
3596 {
3597 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
3598 LLVMBasicBlockRef endif_block;
3599
3600 assert(!current_branch->loop_entry_block);
3601
3602 endif_block = append_basic_block(ctx, "ENDIF");
3603 emit_default_branch(ctx->builder, endif_block);
3604
3605 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
3606 set_basicblock_name(current_branch->next_block, "else", label_id);
3607
3608 current_branch->next_block = endif_block;
3609 }
3610
3611 void ac_build_endif(struct ac_llvm_context *ctx, int label_id)
3612 {
3613 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
3614
3615 assert(!current_branch->loop_entry_block);
3616
3617 emit_default_branch(ctx->builder, current_branch->next_block);
3618 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
3619 set_basicblock_name(current_branch->next_block, "endif", label_id);
3620
3621 ctx->flow->depth--;
3622 }
3623
3624 void ac_build_endloop(struct ac_llvm_context *ctx, int label_id)
3625 {
3626 struct ac_llvm_flow *current_loop = get_current_flow(ctx);
3627
3628 assert(current_loop->loop_entry_block);
3629
3630 emit_default_branch(ctx->builder, current_loop->loop_entry_block);
3631
3632 LLVMPositionBuilderAtEnd(ctx->builder, current_loop->next_block);
3633 set_basicblock_name(current_loop->next_block, "endloop", label_id);
3634 ctx->flow->depth--;
3635 }
3636
3637 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id)
3638 {
3639 struct ac_llvm_flow *flow = push_flow(ctx);
3640 LLVMBasicBlockRef if_block;
3641
3642 if_block = append_basic_block(ctx, "IF");
3643 flow->next_block = append_basic_block(ctx, "ELSE");
3644 set_basicblock_name(if_block, "if", label_id);
3645 LLVMBuildCondBr(ctx->builder, cond, if_block, flow->next_block);
3646 LLVMPositionBuilderAtEnd(ctx->builder, if_block);
3647 }
3648
3649 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
3650 int label_id)
3651 {
3652 LLVMValueRef cond = LLVMBuildFCmp(ctx->builder, LLVMRealUNE,
3653 value, ctx->f32_0, "");
3654 ac_build_ifcc(ctx, cond, label_id);
3655 }
3656
3657 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
3658 int label_id)
3659 {
3660 LLVMValueRef cond = LLVMBuildICmp(ctx->builder, LLVMIntNE,
3661 ac_to_integer(ctx, value),
3662 ctx->i32_0, "");
3663 ac_build_ifcc(ctx, cond, label_id);
3664 }
3665
3666 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
3667 const char *name)
3668 {
3669 LLVMBuilderRef builder = ac->builder;
3670 LLVMBasicBlockRef current_block = LLVMGetInsertBlock(builder);
3671 LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
3672 LLVMBasicBlockRef first_block = LLVMGetEntryBasicBlock(function);
3673 LLVMValueRef first_instr = LLVMGetFirstInstruction(first_block);
3674 LLVMBuilderRef first_builder = LLVMCreateBuilderInContext(ac->context);
3675 LLVMValueRef res;
3676
3677 if (first_instr) {
3678 LLVMPositionBuilderBefore(first_builder, first_instr);
3679 } else {
3680 LLVMPositionBuilderAtEnd(first_builder, first_block);
3681 }
3682
3683 res = LLVMBuildAlloca(first_builder, type, name);
3684 LLVMDisposeBuilder(first_builder);
3685 return res;
3686 }
3687
3688 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac,
3689 LLVMTypeRef type, const char *name)
3690 {
3691 LLVMValueRef ptr = ac_build_alloca_undef(ac, type, name);
3692 LLVMBuildStore(ac->builder, LLVMConstNull(type), ptr);
3693 return ptr;
3694 }
3695
3696 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
3697 LLVMTypeRef type)
3698 {
3699 int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
3700 return LLVMBuildBitCast(ctx->builder, ptr,
3701 LLVMPointerType(type, addr_space), "");
3702 }
3703
3704 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
3705 unsigned count)
3706 {
3707 unsigned num_components = ac_get_llvm_num_components(value);
3708 if (count == num_components)
3709 return value;
3710
3711 LLVMValueRef masks[MAX2(count, 2)];
3712 masks[0] = ctx->i32_0;
3713 masks[1] = ctx->i32_1;
3714 for (unsigned i = 2; i < count; i++)
3715 masks[i] = LLVMConstInt(ctx->i32, i, false);
3716
3717 if (count == 1)
3718 return LLVMBuildExtractElement(ctx->builder, value, masks[0],
3719 "");
3720
3721 LLVMValueRef swizzle = LLVMConstVector(masks, count);
3722 return LLVMBuildShuffleVector(ctx->builder, value, value, swizzle, "");
3723 }
3724
3725 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
3726 unsigned rshift, unsigned bitwidth)
3727 {
3728 LLVMValueRef value = param;
3729 if (rshift)
3730 value = LLVMBuildLShr(ctx->builder, value,
3731 LLVMConstInt(ctx->i32, rshift, false), "");
3732
3733 if (rshift + bitwidth < 32) {
3734 unsigned mask = (1 << bitwidth) - 1;
3735 value = LLVMBuildAnd(ctx->builder, value,
3736 LLVMConstInt(ctx->i32, mask, false), "");
3737 }
3738 return value;
3739 }
3740
3741 /* Adjust the sample index according to FMASK.
3742 *
3743 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3744 * which is the identity mapping. Each nibble says which physical sample
3745 * should be fetched to get that sample.
3746 *
3747 * For example, 0x11111100 means there are only 2 samples stored and
3748 * the second sample covers 3/4 of the pixel. When reading samples 0
3749 * and 1, return physical sample 0 (determined by the first two 0s
3750 * in FMASK), otherwise return physical sample 1.
3751 *
3752 * The sample index should be adjusted as follows:
3753 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
3754 */
3755 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
3756 LLVMValueRef *addr, bool is_array_tex)
3757 {
3758 struct ac_image_args fmask_load = {};
3759 fmask_load.opcode = ac_image_load;
3760 fmask_load.resource = fmask;
3761 fmask_load.dmask = 0xf;
3762 fmask_load.dim = is_array_tex ? ac_image_2darray : ac_image_2d;
3763 fmask_load.attributes = AC_FUNC_ATTR_READNONE;
3764
3765 fmask_load.coords[0] = addr[0];
3766 fmask_load.coords[1] = addr[1];
3767 if (is_array_tex)
3768 fmask_load.coords[2] = addr[2];
3769
3770 LLVMValueRef fmask_value = ac_build_image_opcode(ac, &fmask_load);
3771 fmask_value = LLVMBuildExtractElement(ac->builder, fmask_value,
3772 ac->i32_0, "");
3773
3774 /* Apply the formula. */
3775 unsigned sample_chan = is_array_tex ? 3 : 2;
3776 LLVMValueRef final_sample;
3777 final_sample = LLVMBuildMul(ac->builder, addr[sample_chan],
3778 LLVMConstInt(ac->i32, 4, 0), "");
3779 final_sample = LLVMBuildLShr(ac->builder, fmask_value, final_sample, "");
3780 /* Mask the sample index by 0x7, because 0x8 means an unknown value
3781 * with EQAA, so those will map to 0. */
3782 final_sample = LLVMBuildAnd(ac->builder, final_sample,
3783 LLVMConstInt(ac->i32, 0x7, 0), "");
3784
3785 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3786 * resource descriptor is 0 (invalid).
3787 */
3788 LLVMValueRef tmp;
3789 tmp = LLVMBuildBitCast(ac->builder, fmask, ac->v8i32, "");
3790 tmp = LLVMBuildExtractElement(ac->builder, tmp, ac->i32_1, "");
3791 tmp = LLVMBuildICmp(ac->builder, LLVMIntNE, tmp, ac->i32_0, "");
3792
3793 /* Replace the MSAA sample index. */
3794 addr[sample_chan] = LLVMBuildSelect(ac->builder, tmp, final_sample,
3795 addr[sample_chan], "");
3796 }
3797
3798 static LLVMValueRef
3799 _ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
3800 {
3801 ac_build_optimization_barrier(ctx, &src);
3802 return ac_build_intrinsic(ctx,
3803 lane == NULL ? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
3804 LLVMTypeOf(src), (LLVMValueRef []) {
3805 src, lane },
3806 lane == NULL ? 1 : 2,
3807 AC_FUNC_ATTR_READNONE |
3808 AC_FUNC_ATTR_CONVERGENT);
3809 }
3810
3811 /**
3812 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
3813 * @param ctx
3814 * @param src
3815 * @param lane - id of the lane or NULL for the first active lane
3816 * @return value of the lane
3817 */
3818 LLVMValueRef
3819 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
3820 {
3821 LLVMTypeRef src_type = LLVMTypeOf(src);
3822 src = ac_to_integer(ctx, src);
3823 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3824 LLVMValueRef ret;
3825
3826 if (bits == 32) {
3827 ret = _ac_build_readlane(ctx, src, lane);
3828 } else {
3829 assert(bits % 32 == 0);
3830 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3831 LLVMValueRef src_vector =
3832 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3833 ret = LLVMGetUndef(vec_type);
3834 for (unsigned i = 0; i < bits / 32; i++) {
3835 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3836 LLVMConstInt(ctx->i32, i, 0), "");
3837 LLVMValueRef ret_comp = _ac_build_readlane(ctx, src, lane);
3838 ret = LLVMBuildInsertElement(ctx->builder, ret, ret_comp,
3839 LLVMConstInt(ctx->i32, i, 0), "");
3840 }
3841 }
3842 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3843 }
3844
3845 LLVMValueRef
3846 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane)
3847 {
3848 if (HAVE_LLVM >= 0x0800) {
3849 return ac_build_intrinsic(ctx, "llvm.amdgcn.writelane", ctx->i32,
3850 (LLVMValueRef []) {value, lane, src}, 3,
3851 AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3852 }
3853
3854 LLVMValueRef pred = LLVMBuildICmp(ctx->builder, LLVMIntEQ, lane,
3855 ac_get_thread_id(ctx), "");
3856 return LLVMBuildSelect(ctx->builder, pred, value, src, "");
3857 }
3858
3859 LLVMValueRef
3860 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask)
3861 {
3862 if (ctx->wave_size == 32) {
3863 return ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
3864 (LLVMValueRef []) { mask, ctx->i32_0 },
3865 2, AC_FUNC_ATTR_READNONE);
3866 }
3867 LLVMValueRef mask_vec = LLVMBuildBitCast(ctx->builder, mask,
3868 LLVMVectorType(ctx->i32, 2),
3869 "");
3870 LLVMValueRef mask_lo = LLVMBuildExtractElement(ctx->builder, mask_vec,
3871 ctx->i32_0, "");
3872 LLVMValueRef mask_hi = LLVMBuildExtractElement(ctx->builder, mask_vec,
3873 ctx->i32_1, "");
3874 LLVMValueRef val =
3875 ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
3876 (LLVMValueRef []) { mask_lo, ctx->i32_0 },
3877 2, AC_FUNC_ATTR_READNONE);
3878 val = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi", ctx->i32,
3879 (LLVMValueRef []) { mask_hi, val },
3880 2, AC_FUNC_ATTR_READNONE);
3881 return val;
3882 }
3883
3884 enum dpp_ctrl {
3885 _dpp_quad_perm = 0x000,
3886 _dpp_row_sl = 0x100,
3887 _dpp_row_sr = 0x110,
3888 _dpp_row_rr = 0x120,
3889 dpp_wf_sl1 = 0x130,
3890 dpp_wf_rl1 = 0x134,
3891 dpp_wf_sr1 = 0x138,
3892 dpp_wf_rr1 = 0x13C,
3893 dpp_row_mirror = 0x140,
3894 dpp_row_half_mirror = 0x141,
3895 dpp_row_bcast15 = 0x142,
3896 dpp_row_bcast31 = 0x143
3897 };
3898
3899 static inline enum dpp_ctrl
3900 dpp_quad_perm(unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
3901 {
3902 assert(lane0 < 4 && lane1 < 4 && lane2 < 4 && lane3 < 4);
3903 return _dpp_quad_perm | lane0 | (lane1 << 2) | (lane2 << 4) | (lane3 << 6);
3904 }
3905
3906 static inline enum dpp_ctrl
3907 dpp_row_sl(unsigned amount)
3908 {
3909 assert(amount > 0 && amount < 16);
3910 return _dpp_row_sl | amount;
3911 }
3912
3913 static inline enum dpp_ctrl
3914 dpp_row_sr(unsigned amount)
3915 {
3916 assert(amount > 0 && amount < 16);
3917 return _dpp_row_sr | amount;
3918 }
3919
3920 static LLVMValueRef
3921 _ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3922 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3923 bool bound_ctrl)
3924 {
3925 return ac_build_intrinsic(ctx, "llvm.amdgcn.update.dpp.i32",
3926 LLVMTypeOf(old),
3927 (LLVMValueRef[]) {
3928 old, src,
3929 LLVMConstInt(ctx->i32, dpp_ctrl, 0),
3930 LLVMConstInt(ctx->i32, row_mask, 0),
3931 LLVMConstInt(ctx->i32, bank_mask, 0),
3932 LLVMConstInt(ctx->i1, bound_ctrl, 0) },
3933 6, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3934 }
3935
3936 static LLVMValueRef
3937 ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3938 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3939 bool bound_ctrl)
3940 {
3941 LLVMTypeRef src_type = LLVMTypeOf(src);
3942 src = ac_to_integer(ctx, src);
3943 old = ac_to_integer(ctx, old);
3944 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3945 LLVMValueRef ret;
3946 if (bits == 32) {
3947 ret = _ac_build_dpp(ctx, old, src, dpp_ctrl, row_mask,
3948 bank_mask, bound_ctrl);
3949 } else {
3950 assert(bits % 32 == 0);
3951 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3952 LLVMValueRef src_vector =
3953 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3954 LLVMValueRef old_vector =
3955 LLVMBuildBitCast(ctx->builder, old, vec_type, "");
3956 ret = LLVMGetUndef(vec_type);
3957 for (unsigned i = 0; i < bits / 32; i++) {
3958 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3959 LLVMConstInt(ctx->i32, i,
3960 0), "");
3961 old = LLVMBuildExtractElement(ctx->builder, old_vector,
3962 LLVMConstInt(ctx->i32, i,
3963 0), "");
3964 LLVMValueRef ret_comp = _ac_build_dpp(ctx, old, src,
3965 dpp_ctrl,
3966 row_mask,
3967 bank_mask,
3968 bound_ctrl);
3969 ret = LLVMBuildInsertElement(ctx->builder, ret,
3970 ret_comp,
3971 LLVMConstInt(ctx->i32, i,
3972 0), "");
3973 }
3974 }
3975 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3976 }
3977
3978 static LLVMValueRef
3979 _ac_build_permlane16(struct ac_llvm_context *ctx, LLVMValueRef src, uint64_t sel,
3980 bool exchange_rows, bool bound_ctrl)
3981 {
3982 LLVMValueRef args[6] = {
3983 src,
3984 src,
3985 LLVMConstInt(ctx->i32, sel, false),
3986 LLVMConstInt(ctx->i32, sel >> 32, false),
3987 ctx->i1true, /* fi */
3988 bound_ctrl ? ctx->i1true : ctx->i1false,
3989 };
3990 return ac_build_intrinsic(ctx, exchange_rows ? "llvm.amdgcn.permlanex16"
3991 : "llvm.amdgcn.permlane16",
3992 ctx->i32, args, 6,
3993 AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3994 }
3995
3996 static LLVMValueRef
3997 ac_build_permlane16(struct ac_llvm_context *ctx, LLVMValueRef src, uint64_t sel,
3998 bool exchange_rows, bool bound_ctrl)
3999 {
4000 LLVMTypeRef src_type = LLVMTypeOf(src);
4001 src = ac_to_integer(ctx, src);
4002 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
4003 LLVMValueRef ret;
4004 if (bits == 32) {
4005 ret = _ac_build_permlane16(ctx, src, sel, exchange_rows,
4006 bound_ctrl);
4007 } else {
4008 assert(bits % 32 == 0);
4009 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
4010 LLVMValueRef src_vector =
4011 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
4012 ret = LLVMGetUndef(vec_type);
4013 for (unsigned i = 0; i < bits / 32; i++) {
4014 src = LLVMBuildExtractElement(ctx->builder, src_vector,
4015 LLVMConstInt(ctx->i32, i,
4016 0), "");
4017 LLVMValueRef ret_comp =
4018 _ac_build_permlane16(ctx, src, sel,
4019 exchange_rows,
4020 bound_ctrl);
4021 ret = LLVMBuildInsertElement(ctx->builder, ret,
4022 ret_comp,
4023 LLVMConstInt(ctx->i32, i,
4024 0), "");
4025 }
4026 }
4027 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
4028 }
4029
4030 static inline unsigned
4031 ds_pattern_bitmode(unsigned and_mask, unsigned or_mask, unsigned xor_mask)
4032 {
4033 assert(and_mask < 32 && or_mask < 32 && xor_mask < 32);
4034 return and_mask | (or_mask << 5) | (xor_mask << 10);
4035 }
4036
4037 static LLVMValueRef
4038 _ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
4039 {
4040 return ac_build_intrinsic(ctx, "llvm.amdgcn.ds.swizzle",
4041 LLVMTypeOf(src), (LLVMValueRef []) {
4042 src, LLVMConstInt(ctx->i32, mask, 0) },
4043 2, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
4044 }
4045
4046 LLVMValueRef
4047 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
4048 {
4049 LLVMTypeRef src_type = LLVMTypeOf(src);
4050 src = ac_to_integer(ctx, src);
4051 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
4052 LLVMValueRef ret;
4053 if (bits == 32) {
4054 ret = _ac_build_ds_swizzle(ctx, src, mask);
4055 } else {
4056 assert(bits % 32 == 0);
4057 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
4058 LLVMValueRef src_vector =
4059 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
4060 ret = LLVMGetUndef(vec_type);
4061 for (unsigned i = 0; i < bits / 32; i++) {
4062 src = LLVMBuildExtractElement(ctx->builder, src_vector,
4063 LLVMConstInt(ctx->i32, i,
4064 0), "");
4065 LLVMValueRef ret_comp = _ac_build_ds_swizzle(ctx, src,
4066 mask);
4067 ret = LLVMBuildInsertElement(ctx->builder, ret,
4068 ret_comp,
4069 LLVMConstInt(ctx->i32, i,
4070 0), "");
4071 }
4072 }
4073 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
4074 }
4075
4076 static LLVMValueRef
4077 ac_build_wwm(struct ac_llvm_context *ctx, LLVMValueRef src)
4078 {
4079 char name[32], type[8];
4080 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
4081 snprintf(name, sizeof(name), "llvm.amdgcn.wwm.%s", type);
4082 return ac_build_intrinsic(ctx, name, LLVMTypeOf(src),
4083 (LLVMValueRef []) { src }, 1,
4084 AC_FUNC_ATTR_READNONE);
4085 }
4086
4087 static LLVMValueRef
4088 ac_build_set_inactive(struct ac_llvm_context *ctx, LLVMValueRef src,
4089 LLVMValueRef inactive)
4090 {
4091 char name[33], type[8];
4092 LLVMTypeRef src_type = LLVMTypeOf(src);
4093 src = ac_to_integer(ctx, src);
4094 inactive = ac_to_integer(ctx, inactive);
4095 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
4096 snprintf(name, sizeof(name), "llvm.amdgcn.set.inactive.%s", type);
4097 LLVMValueRef ret =
4098 ac_build_intrinsic(ctx, name,
4099 LLVMTypeOf(src), (LLVMValueRef []) {
4100 src, inactive }, 2,
4101 AC_FUNC_ATTR_READNONE |
4102 AC_FUNC_ATTR_CONVERGENT);
4103 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
4104 }
4105
4106 static LLVMValueRef
4107 get_reduction_identity(struct ac_llvm_context *ctx, nir_op op, unsigned type_size)
4108 {
4109 if (type_size == 4) {
4110 switch (op) {
4111 case nir_op_iadd: return ctx->i32_0;
4112 case nir_op_fadd: return ctx->f32_0;
4113 case nir_op_imul: return ctx->i32_1;
4114 case nir_op_fmul: return ctx->f32_1;
4115 case nir_op_imin: return LLVMConstInt(ctx->i32, INT32_MAX, 0);
4116 case nir_op_umin: return LLVMConstInt(ctx->i32, UINT32_MAX, 0);
4117 case nir_op_fmin: return LLVMConstReal(ctx->f32, INFINITY);
4118 case nir_op_imax: return LLVMConstInt(ctx->i32, INT32_MIN, 0);
4119 case nir_op_umax: return ctx->i32_0;
4120 case nir_op_fmax: return LLVMConstReal(ctx->f32, -INFINITY);
4121 case nir_op_iand: return LLVMConstInt(ctx->i32, -1, 0);
4122 case nir_op_ior: return ctx->i32_0;
4123 case nir_op_ixor: return ctx->i32_0;
4124 default:
4125 unreachable("bad reduction intrinsic");
4126 }
4127 } else { /* type_size == 64bit */
4128 switch (op) {
4129 case nir_op_iadd: return ctx->i64_0;
4130 case nir_op_fadd: return ctx->f64_0;
4131 case nir_op_imul: return ctx->i64_1;
4132 case nir_op_fmul: return ctx->f64_1;
4133 case nir_op_imin: return LLVMConstInt(ctx->i64, INT64_MAX, 0);
4134 case nir_op_umin: return LLVMConstInt(ctx->i64, UINT64_MAX, 0);
4135 case nir_op_fmin: return LLVMConstReal(ctx->f64, INFINITY);
4136 case nir_op_imax: return LLVMConstInt(ctx->i64, INT64_MIN, 0);
4137 case nir_op_umax: return ctx->i64_0;
4138 case nir_op_fmax: return LLVMConstReal(ctx->f64, -INFINITY);
4139 case nir_op_iand: return LLVMConstInt(ctx->i64, -1, 0);
4140 case nir_op_ior: return ctx->i64_0;
4141 case nir_op_ixor: return ctx->i64_0;
4142 default:
4143 unreachable("bad reduction intrinsic");
4144 }
4145 }
4146 }
4147
4148 static LLVMValueRef
4149 ac_build_alu_op(struct ac_llvm_context *ctx, LLVMValueRef lhs, LLVMValueRef rhs, nir_op op)
4150 {
4151 bool _64bit = ac_get_type_size(LLVMTypeOf(lhs)) == 8;
4152 switch (op) {
4153 case nir_op_iadd: return LLVMBuildAdd(ctx->builder, lhs, rhs, "");
4154 case nir_op_fadd: return LLVMBuildFAdd(ctx->builder, lhs, rhs, "");
4155 case nir_op_imul: return LLVMBuildMul(ctx->builder, lhs, rhs, "");
4156 case nir_op_fmul: return LLVMBuildFMul(ctx->builder, lhs, rhs, "");
4157 case nir_op_imin: return LLVMBuildSelect(ctx->builder,
4158 LLVMBuildICmp(ctx->builder, LLVMIntSLT, lhs, rhs, ""),
4159 lhs, rhs, "");
4160 case nir_op_umin: return LLVMBuildSelect(ctx->builder,
4161 LLVMBuildICmp(ctx->builder, LLVMIntULT, lhs, rhs, ""),
4162 lhs, rhs, "");
4163 case nir_op_fmin: return ac_build_intrinsic(ctx,
4164 _64bit ? "llvm.minnum.f64" : "llvm.minnum.f32",
4165 _64bit ? ctx->f64 : ctx->f32,
4166 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
4167 case nir_op_imax: return LLVMBuildSelect(ctx->builder,
4168 LLVMBuildICmp(ctx->builder, LLVMIntSGT, lhs, rhs, ""),
4169 lhs, rhs, "");
4170 case nir_op_umax: return LLVMBuildSelect(ctx->builder,
4171 LLVMBuildICmp(ctx->builder, LLVMIntUGT, lhs, rhs, ""),
4172 lhs, rhs, "");
4173 case nir_op_fmax: return ac_build_intrinsic(ctx,
4174 _64bit ? "llvm.maxnum.f64" : "llvm.maxnum.f32",
4175 _64bit ? ctx->f64 : ctx->f32,
4176 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
4177 case nir_op_iand: return LLVMBuildAnd(ctx->builder, lhs, rhs, "");
4178 case nir_op_ior: return LLVMBuildOr(ctx->builder, lhs, rhs, "");
4179 case nir_op_ixor: return LLVMBuildXor(ctx->builder, lhs, rhs, "");
4180 default:
4181 unreachable("bad reduction intrinsic");
4182 }
4183 }
4184
4185 /**
4186 * \param maxprefix specifies that the result only needs to be correct for a
4187 * prefix of this many threads
4188 *
4189 * TODO: add inclusive and excluse scan functions for GFX6.
4190 */
4191 static LLVMValueRef
4192 ac_build_scan(struct ac_llvm_context *ctx, nir_op op, LLVMValueRef src, LLVMValueRef identity,
4193 unsigned maxprefix, bool inclusive)
4194 {
4195 LLVMValueRef result, tmp;
4196
4197 if (ctx->chip_class >= GFX10) {
4198 result = inclusive ? src : identity;
4199 } else {
4200 if (inclusive)
4201 result = src;
4202 else
4203 result = ac_build_dpp(ctx, identity, src, dpp_wf_sr1, 0xf, 0xf, false);
4204 }
4205 if (maxprefix <= 1)
4206 return result;
4207 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
4208 result = ac_build_alu_op(ctx, result, tmp, op);
4209 if (maxprefix <= 2)
4210 return result;
4211 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(2), 0xf, 0xf, false);
4212 result = ac_build_alu_op(ctx, result, tmp, op);
4213 if (maxprefix <= 3)
4214 return result;
4215 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(3), 0xf, 0xf, false);
4216 result = ac_build_alu_op(ctx, result, tmp, op);
4217 if (maxprefix <= 4)
4218 return result;
4219 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(4), 0xf, 0xe, false);
4220 result = ac_build_alu_op(ctx, result, tmp, op);
4221 if (maxprefix <= 8)
4222 return result;
4223 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(8), 0xf, 0xc, false);
4224 result = ac_build_alu_op(ctx, result, tmp, op);
4225 if (maxprefix <= 16)
4226 return result;
4227
4228 if (ctx->chip_class >= GFX10) {
4229 /* dpp_row_bcast{15,31} are not supported on gfx10. */
4230 LLVMBuilderRef builder = ctx->builder;
4231 LLVMValueRef tid = ac_get_thread_id(ctx);
4232 LLVMValueRef cc;
4233 /* TODO-GFX10: Can we get better code-gen by putting this into
4234 * a branch so that LLVM generates EXEC mask manipulations? */
4235 if (inclusive)
4236 tmp = result;
4237 else
4238 tmp = ac_build_alu_op(ctx, result, src, op);
4239 tmp = ac_build_permlane16(ctx, tmp, ~(uint64_t)0, true, false);
4240 tmp = ac_build_alu_op(ctx, result, tmp, op);
4241 cc = LLVMBuildAnd(builder, tid, LLVMConstInt(ctx->i32, 16, false), "");
4242 cc = LLVMBuildICmp(builder, LLVMIntNE, cc, ctx->i32_0, "");
4243 result = LLVMBuildSelect(builder, cc, tmp, result, "");
4244 if (maxprefix <= 32)
4245 return result;
4246
4247 if (inclusive)
4248 tmp = result;
4249 else
4250 tmp = ac_build_alu_op(ctx, result, src, op);
4251 tmp = ac_build_readlane(ctx, tmp, LLVMConstInt(ctx->i32, 31, false));
4252 tmp = ac_build_alu_op(ctx, result, tmp, op);
4253 cc = LLVMBuildICmp(builder, LLVMIntUGE, tid,
4254 LLVMConstInt(ctx->i32, 32, false), "");
4255 result = LLVMBuildSelect(builder, cc, tmp, result, "");
4256 return result;
4257 }
4258
4259 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
4260 result = ac_build_alu_op(ctx, result, tmp, op);
4261 if (maxprefix <= 32)
4262 return result;
4263 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
4264 result = ac_build_alu_op(ctx, result, tmp, op);
4265 return result;
4266 }
4267
4268 LLVMValueRef
4269 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
4270 {
4271 LLVMValueRef result;
4272
4273 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
4274 LLVMBuilderRef builder = ctx->builder;
4275 src = LLVMBuildZExt(builder, src, ctx->i32, "");
4276 result = ac_build_ballot(ctx, src);
4277 result = ac_build_mbcnt(ctx, result);
4278 result = LLVMBuildAdd(builder, result, src, "");
4279 return result;
4280 }
4281
4282 ac_build_optimization_barrier(ctx, &src);
4283
4284 LLVMValueRef identity =
4285 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
4286 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
4287 LLVMTypeOf(identity), "");
4288 result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, true);
4289
4290 return ac_build_wwm(ctx, result);
4291 }
4292
4293 LLVMValueRef
4294 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
4295 {
4296 LLVMValueRef result;
4297
4298 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
4299 LLVMBuilderRef builder = ctx->builder;
4300 src = LLVMBuildZExt(builder, src, ctx->i32, "");
4301 result = ac_build_ballot(ctx, src);
4302 result = ac_build_mbcnt(ctx, result);
4303 return result;
4304 }
4305
4306 ac_build_optimization_barrier(ctx, &src);
4307
4308 LLVMValueRef identity =
4309 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
4310 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
4311 LLVMTypeOf(identity), "");
4312 result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, false);
4313
4314 return ac_build_wwm(ctx, result);
4315 }
4316
4317 LLVMValueRef
4318 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size)
4319 {
4320 if (cluster_size == 1) return src;
4321 ac_build_optimization_barrier(ctx, &src);
4322 LLVMValueRef result, swap;
4323 LLVMValueRef identity = get_reduction_identity(ctx, op,
4324 ac_get_type_size(LLVMTypeOf(src)));
4325 result = LLVMBuildBitCast(ctx->builder,
4326 ac_build_set_inactive(ctx, src, identity),
4327 LLVMTypeOf(identity), "");
4328 swap = ac_build_quad_swizzle(ctx, result, 1, 0, 3, 2);
4329 result = ac_build_alu_op(ctx, result, swap, op);
4330 if (cluster_size == 2) return ac_build_wwm(ctx, result);
4331
4332 swap = ac_build_quad_swizzle(ctx, result, 2, 3, 0, 1);
4333 result = ac_build_alu_op(ctx, result, swap, op);
4334 if (cluster_size == 4) return ac_build_wwm(ctx, result);
4335
4336 if (ctx->chip_class >= GFX8)
4337 swap = ac_build_dpp(ctx, identity, result, dpp_row_half_mirror, 0xf, 0xf, false);
4338 else
4339 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x04));
4340 result = ac_build_alu_op(ctx, result, swap, op);
4341 if (cluster_size == 8) return ac_build_wwm(ctx, result);
4342
4343 if (ctx->chip_class >= GFX8)
4344 swap = ac_build_dpp(ctx, identity, result, dpp_row_mirror, 0xf, 0xf, false);
4345 else
4346 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x08));
4347 result = ac_build_alu_op(ctx, result, swap, op);
4348 if (cluster_size == 16) return ac_build_wwm(ctx, result);
4349
4350 if (ctx->chip_class >= GFX10)
4351 swap = ac_build_permlane16(ctx, result, 0, true, false);
4352 else if (ctx->chip_class >= GFX8 && cluster_size != 32)
4353 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
4354 else
4355 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x10));
4356 result = ac_build_alu_op(ctx, result, swap, op);
4357 if (cluster_size == 32) return ac_build_wwm(ctx, result);
4358
4359 if (ctx->chip_class >= GFX8) {
4360 if (ctx->chip_class >= GFX10)
4361 swap = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, false));
4362 else
4363 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
4364 result = ac_build_alu_op(ctx, result, swap, op);
4365 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 63, 0));
4366 return ac_build_wwm(ctx, result);
4367 } else {
4368 swap = ac_build_readlane(ctx, result, ctx->i32_0);
4369 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 32, 0));
4370 result = ac_build_alu_op(ctx, result, swap, op);
4371 return ac_build_wwm(ctx, result);
4372 }
4373 }
4374
4375 /**
4376 * "Top half" of a scan that reduces per-wave values across an entire
4377 * workgroup.
4378 *
4379 * The source value must be present in the highest lane of the wave, and the
4380 * highest lane must be live.
4381 */
4382 void
4383 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4384 {
4385 if (ws->maxwaves <= 1)
4386 return;
4387
4388 const LLVMValueRef last_lane = LLVMConstInt(ctx->i32, ctx->wave_size - 1, false);
4389 LLVMBuilderRef builder = ctx->builder;
4390 LLVMValueRef tid = ac_get_thread_id(ctx);
4391 LLVMValueRef tmp;
4392
4393 tmp = LLVMBuildICmp(builder, LLVMIntEQ, tid, last_lane, "");
4394 ac_build_ifcc(ctx, tmp, 1000);
4395 LLVMBuildStore(builder, ws->src, LLVMBuildGEP(builder, ws->scratch, &ws->waveidx, 1, ""));
4396 ac_build_endif(ctx, 1000);
4397 }
4398
4399 /**
4400 * "Bottom half" of a scan that reduces per-wave values across an entire
4401 * workgroup.
4402 *
4403 * The caller must place a barrier between the top and bottom halves.
4404 */
4405 void
4406 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4407 {
4408 const LLVMTypeRef type = LLVMTypeOf(ws->src);
4409 const LLVMValueRef identity =
4410 get_reduction_identity(ctx, ws->op, ac_get_type_size(type));
4411
4412 if (ws->maxwaves <= 1) {
4413 ws->result_reduce = ws->src;
4414 ws->result_inclusive = ws->src;
4415 ws->result_exclusive = identity;
4416 return;
4417 }
4418 assert(ws->maxwaves <= 32);
4419
4420 LLVMBuilderRef builder = ctx->builder;
4421 LLVMValueRef tid = ac_get_thread_id(ctx);
4422 LLVMBasicBlockRef bbs[2];
4423 LLVMValueRef phivalues_scan[2];
4424 LLVMValueRef tmp, tmp2;
4425
4426 bbs[0] = LLVMGetInsertBlock(builder);
4427 phivalues_scan[0] = LLVMGetUndef(type);
4428
4429 if (ws->enable_reduce)
4430 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->numwaves, "");
4431 else if (ws->enable_inclusive)
4432 tmp = LLVMBuildICmp(builder, LLVMIntULE, tid, ws->waveidx, "");
4433 else
4434 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->waveidx, "");
4435 ac_build_ifcc(ctx, tmp, 1001);
4436 {
4437 tmp = LLVMBuildLoad(builder, LLVMBuildGEP(builder, ws->scratch, &tid, 1, ""), "");
4438
4439 ac_build_optimization_barrier(ctx, &tmp);
4440
4441 bbs[1] = LLVMGetInsertBlock(builder);
4442 phivalues_scan[1] = ac_build_scan(ctx, ws->op, tmp, identity, ws->maxwaves, true);
4443 }
4444 ac_build_endif(ctx, 1001);
4445
4446 const LLVMValueRef scan = ac_build_phi(ctx, type, 2, phivalues_scan, bbs);
4447
4448 if (ws->enable_reduce) {
4449 tmp = LLVMBuildSub(builder, ws->numwaves, ctx->i32_1, "");
4450 ws->result_reduce = ac_build_readlane(ctx, scan, tmp);
4451 }
4452 if (ws->enable_inclusive)
4453 ws->result_inclusive = ac_build_readlane(ctx, scan, ws->waveidx);
4454 if (ws->enable_exclusive) {
4455 tmp = LLVMBuildSub(builder, ws->waveidx, ctx->i32_1, "");
4456 tmp = ac_build_readlane(ctx, scan, tmp);
4457 tmp2 = LLVMBuildICmp(builder, LLVMIntEQ, ws->waveidx, ctx->i32_0, "");
4458 ws->result_exclusive = LLVMBuildSelect(builder, tmp2, identity, tmp, "");
4459 }
4460 }
4461
4462 /**
4463 * Inclusive scan of a per-wave value across an entire workgroup.
4464 *
4465 * This implies an s_barrier instruction.
4466 *
4467 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
4468 * of the workgroup are live. (This requirement cannot easily be relaxed in a
4469 * useful manner because of the barrier in the algorithm.)
4470 */
4471 void
4472 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4473 {
4474 ac_build_wg_wavescan_top(ctx, ws);
4475 ac_build_s_barrier(ctx);
4476 ac_build_wg_wavescan_bottom(ctx, ws);
4477 }
4478
4479 /**
4480 * "Top half" of a scan that reduces per-thread values across an entire
4481 * workgroup.
4482 *
4483 * All lanes must be active when this code runs.
4484 */
4485 void
4486 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4487 {
4488 if (ws->enable_exclusive) {
4489 ws->extra = ac_build_exclusive_scan(ctx, ws->src, ws->op);
4490 if (LLVMTypeOf(ws->src) == ctx->i1 && ws->op == nir_op_iadd)
4491 ws->src = LLVMBuildZExt(ctx->builder, ws->src, ctx->i32, "");
4492 ws->src = ac_build_alu_op(ctx, ws->extra, ws->src, ws->op);
4493 } else {
4494 ws->src = ac_build_inclusive_scan(ctx, ws->src, ws->op);
4495 }
4496
4497 bool enable_inclusive = ws->enable_inclusive;
4498 bool enable_exclusive = ws->enable_exclusive;
4499 ws->enable_inclusive = false;
4500 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
4501 ac_build_wg_wavescan_top(ctx, ws);
4502 ws->enable_inclusive = enable_inclusive;
4503 ws->enable_exclusive = enable_exclusive;
4504 }
4505
4506 /**
4507 * "Bottom half" of a scan that reduces per-thread values across an entire
4508 * workgroup.
4509 *
4510 * The caller must place a barrier between the top and bottom halves.
4511 */
4512 void
4513 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4514 {
4515 bool enable_inclusive = ws->enable_inclusive;
4516 bool enable_exclusive = ws->enable_exclusive;
4517 ws->enable_inclusive = false;
4518 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
4519 ac_build_wg_wavescan_bottom(ctx, ws);
4520 ws->enable_inclusive = enable_inclusive;
4521 ws->enable_exclusive = enable_exclusive;
4522
4523 /* ws->result_reduce is already the correct value */
4524 if (ws->enable_inclusive)
4525 ws->result_inclusive = ac_build_alu_op(ctx, ws->result_inclusive, ws->src, ws->op);
4526 if (ws->enable_exclusive)
4527 ws->result_exclusive = ac_build_alu_op(ctx, ws->result_exclusive, ws->extra, ws->op);
4528 }
4529
4530 /**
4531 * A scan that reduces per-thread values across an entire workgroup.
4532 *
4533 * The caller must ensure that all lanes are active when this code runs
4534 * (WWM is insufficient!), because there is an implied barrier.
4535 */
4536 void
4537 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4538 {
4539 ac_build_wg_scan_top(ctx, ws);
4540 ac_build_s_barrier(ctx);
4541 ac_build_wg_scan_bottom(ctx, ws);
4542 }
4543
4544 LLVMValueRef
4545 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
4546 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
4547 {
4548 unsigned mask = dpp_quad_perm(lane0, lane1, lane2, lane3);
4549 if (ctx->chip_class >= GFX8) {
4550 return ac_build_dpp(ctx, src, src, mask, 0xf, 0xf, false);
4551 } else {
4552 return ac_build_ds_swizzle(ctx, src, (1 << 15) | mask);
4553 }
4554 }
4555
4556 LLVMValueRef
4557 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index)
4558 {
4559 index = LLVMBuildMul(ctx->builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
4560 return ac_build_intrinsic(ctx,
4561 "llvm.amdgcn.ds.bpermute", ctx->i32,
4562 (LLVMValueRef []) {index, src}, 2,
4563 AC_FUNC_ATTR_READNONE |
4564 AC_FUNC_ATTR_CONVERGENT);
4565 }
4566
4567 LLVMValueRef
4568 ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
4569 unsigned bitsize)
4570 {
4571 LLVMTypeRef type;
4572 char *intr;
4573
4574 if (bitsize == 16) {
4575 intr = "llvm.amdgcn.frexp.exp.i16.f16";
4576 type = ctx->i16;
4577 } else if (bitsize == 32) {
4578 intr = "llvm.amdgcn.frexp.exp.i32.f32";
4579 type = ctx->i32;
4580 } else {
4581 intr = "llvm.amdgcn.frexp.exp.i32.f64";
4582 type = ctx->i32;
4583 }
4584
4585 LLVMValueRef params[] = {
4586 src0,
4587 };
4588 return ac_build_intrinsic(ctx, intr, type, params, 1,
4589 AC_FUNC_ATTR_READNONE);
4590 }
4591 LLVMValueRef
4592 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
4593 unsigned bitsize)
4594 {
4595 LLVMTypeRef type;
4596 char *intr;
4597
4598 if (bitsize == 16) {
4599 intr = "llvm.amdgcn.frexp.mant.f16";
4600 type = ctx->f16;
4601 } else if (bitsize == 32) {
4602 intr = "llvm.amdgcn.frexp.mant.f32";
4603 type = ctx->f32;
4604 } else {
4605 intr = "llvm.amdgcn.frexp.mant.f64";
4606 type = ctx->f64;
4607 }
4608
4609 LLVMValueRef params[] = {
4610 src0,
4611 };
4612 return ac_build_intrinsic(ctx, intr, type, params, 1,
4613 AC_FUNC_ATTR_READNONE);
4614 }
4615
4616 /*
4617 * this takes an I,J coordinate pair,
4618 * and works out the X and Y derivatives.
4619 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
4620 */
4621 LLVMValueRef
4622 ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij)
4623 {
4624 LLVMValueRef result[4], a;
4625 unsigned i;
4626
4627 for (i = 0; i < 2; i++) {
4628 a = LLVMBuildExtractElement(ctx->builder, interp_ij,
4629 LLVMConstInt(ctx->i32, i, false), "");
4630 result[i] = ac_build_ddxy(ctx, AC_TID_MASK_TOP_LEFT, 1, a);
4631 result[2+i] = ac_build_ddxy(ctx, AC_TID_MASK_TOP_LEFT, 2, a);
4632 }
4633 return ac_build_gather_values(ctx, result, 4);
4634 }
4635
4636 LLVMValueRef
4637 ac_build_load_helper_invocation(struct ac_llvm_context *ctx)
4638 {
4639 LLVMValueRef result = ac_build_intrinsic(ctx, "llvm.amdgcn.ps.live",
4640 ctx->i1, NULL, 0,
4641 AC_FUNC_ATTR_READNONE);
4642 result = LLVMBuildNot(ctx->builder, result, "");
4643 return LLVMBuildSExt(ctx->builder, result, ctx->i32, "");
4644 }
4645
4646 LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
4647 LLVMValueRef *args, unsigned num_args)
4648 {
4649 LLVMValueRef ret = LLVMBuildCall(ctx->builder, func, args, num_args, "");
4650 LLVMSetInstructionCallConv(ret, LLVMGetFunctionCallConv(func));
4651 return ret;
4652 }