projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
ac: use the raw tbuffer version for 16-bit SSBO loads
[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 enum chip_class chip_class, enum radeon_family family)
62 {
63 LLVMValueRef args[1];
64
65 ctx->context = LLVMContextCreate();
66
67 ctx->chip_class = chip_class;
68 ctx->family = family;
69 ctx->module = NULL;
70 ctx->builder = NULL;
71
72 ctx->voidt = LLVMVoidTypeInContext(ctx->context);
73 ctx->i1 = LLVMInt1TypeInContext(ctx->context);
74 ctx->i8 = LLVMInt8TypeInContext(ctx->context);
75 ctx->i16 = LLVMIntTypeInContext(ctx->context, 16);
76 ctx->i32 = LLVMIntTypeInContext(ctx->context, 32);
77 ctx->i64 = LLVMIntTypeInContext(ctx->context, 64);
78 ctx->intptr = ctx->i32;
79 ctx->f16 = LLVMHalfTypeInContext(ctx->context);
80 ctx->f32 = LLVMFloatTypeInContext(ctx->context);
81 ctx->f64 = LLVMDoubleTypeInContext(ctx->context);
82 ctx->v2i16 = LLVMVectorType(ctx->i16, 2);
83 ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
84 ctx->v3i32 = LLVMVectorType(ctx->i32, 3);
85 ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
86 ctx->v2f32 = LLVMVectorType(ctx->f32, 2);
87 ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
88 ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
89
90 ctx->i16_0 = LLVMConstInt(ctx->i16, 0, false);
91 ctx->i16_1 = LLVMConstInt(ctx->i16, 1, false);
92 ctx->i32_0 = LLVMConstInt(ctx->i32, 0, false);
93 ctx->i32_1 = LLVMConstInt(ctx->i32, 1, false);
94 ctx->i64_0 = LLVMConstInt(ctx->i64, 0, false);
95 ctx->i64_1 = LLVMConstInt(ctx->i64, 1, false);
96 ctx->f32_0 = LLVMConstReal(ctx->f32, 0.0);
97 ctx->f32_1 = LLVMConstReal(ctx->f32, 1.0);
98 ctx->f64_0 = LLVMConstReal(ctx->f64, 0.0);
99 ctx->f64_1 = LLVMConstReal(ctx->f64, 1.0);
100
101 ctx->i1false = LLVMConstInt(ctx->i1, 0, false);
102 ctx->i1true = LLVMConstInt(ctx->i1, 1, false);
103
104 ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context,
105 "range", 5);
106
107 ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context,
108 "invariant.load", 14);
109
110 ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->context, "fpmath", 6);
111
112 args[0] = LLVMConstReal(ctx->f32, 2.5);
113 ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->context, args, 1);
114
115 ctx->uniform_md_kind = LLVMGetMDKindIDInContext(ctx->context,
116 "amdgpu.uniform", 14);
117
118 ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
119 }
120
121 void
122 ac_llvm_context_dispose(struct ac_llvm_context *ctx)
123 {
124 free(ctx->flow);
125 ctx->flow = NULL;
126 ctx->flow_depth_max = 0;
127 }
128
129 int
130 ac_get_llvm_num_components(LLVMValueRef value)
131 {
132 LLVMTypeRef type = LLVMTypeOf(value);
133 unsigned num_components = LLVMGetTypeKind(type) == LLVMVectorTypeKind
134 ? LLVMGetVectorSize(type)
135 : 1;
136 return num_components;
137 }
138
139 LLVMValueRef
140 ac_llvm_extract_elem(struct ac_llvm_context *ac,
141 LLVMValueRef value,
142 int index)
143 {
144 if (LLVMGetTypeKind(LLVMTypeOf(value)) != LLVMVectorTypeKind) {
145 assert(index == 0);
146 return value;
147 }
148
149 return LLVMBuildExtractElement(ac->builder, value,
150 LLVMConstInt(ac->i32, index, false), "");
151 }
152
153 int
154 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type)
155 {
156 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
157 type = LLVMGetElementType(type);
158
159 if (LLVMGetTypeKind(type) == LLVMIntegerTypeKind)
160 return LLVMGetIntTypeWidth(type);
161
162 if (type == ctx->f16)
163 return 16;
164 if (type == ctx->f32)
165 return 32;
166 if (type == ctx->f64)
167 return 64;
168
169 unreachable("Unhandled type kind in get_elem_bits");
170 }
171
172 unsigned
173 ac_get_type_size(LLVMTypeRef type)
174 {
175 LLVMTypeKind kind = LLVMGetTypeKind(type);
176
177 switch (kind) {
178 case LLVMIntegerTypeKind:
179 return LLVMGetIntTypeWidth(type) / 8;
180 case LLVMHalfTypeKind:
181 return 2;
182 case LLVMFloatTypeKind:
183 return 4;
184 case LLVMDoubleTypeKind:
185 return 8;
186 case LLVMPointerTypeKind:
187 if (LLVMGetPointerAddressSpace(type) == AC_ADDR_SPACE_CONST_32BIT)
188 return 4;
189 return 8;
190 case LLVMVectorTypeKind:
191 return LLVMGetVectorSize(type) *
192 ac_get_type_size(LLVMGetElementType(type));
193 case LLVMArrayTypeKind:
194 return LLVMGetArrayLength(type) *
195 ac_get_type_size(LLVMGetElementType(type));
196 default:
197 assert(0);
198 return 0;
199 }
200 }
201
202 static LLVMTypeRef to_integer_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
203 {
204 if (t == ctx->f16 || t == ctx->i16)
205 return ctx->i16;
206 else if (t == ctx->f32 || t == ctx->i32)
207 return ctx->i32;
208 else if (t == ctx->f64 || t == ctx->i64)
209 return ctx->i64;
210 else
211 unreachable("Unhandled integer size");
212 }
213
214 LLVMTypeRef
215 ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
216 {
217 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
218 LLVMTypeRef elem_type = LLVMGetElementType(t);
219 return LLVMVectorType(to_integer_type_scalar(ctx, elem_type),
220 LLVMGetVectorSize(t));
221 }
222 if (LLVMGetTypeKind(t) == LLVMPointerTypeKind) {
223 switch (LLVMGetPointerAddressSpace(t)) {
224 case AC_ADDR_SPACE_GLOBAL:
225 return ctx->i64;
226 case AC_ADDR_SPACE_LDS:
227 return ctx->i32;
228 default:
229 unreachable("unhandled address space");
230 }
231 }
232 return to_integer_type_scalar(ctx, t);
233 }
234
235 LLVMValueRef
236 ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v)
237 {
238 LLVMTypeRef type = LLVMTypeOf(v);
239 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
240 return LLVMBuildPtrToInt(ctx->builder, v, ac_to_integer_type(ctx, type), "");
241 }
242 return LLVMBuildBitCast(ctx->builder, v, ac_to_integer_type(ctx, type), "");
243 }
244
245 LLVMValueRef
246 ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v)
247 {
248 LLVMTypeRef type = LLVMTypeOf(v);
249 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind)
250 return v;
251 return ac_to_integer(ctx, v);
252 }
253
254 static LLVMTypeRef to_float_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
255 {
256 if (t == ctx->i16 || t == ctx->f16)
257 return ctx->f16;
258 else if (t == ctx->i32 || t == ctx->f32)
259 return ctx->f32;
260 else if (t == ctx->i64 || t == ctx->f64)
261 return ctx->f64;
262 else
263 unreachable("Unhandled float size");
264 }
265
266 LLVMTypeRef
267 ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
268 {
269 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
270 LLVMTypeRef elem_type = LLVMGetElementType(t);
271 return LLVMVectorType(to_float_type_scalar(ctx, elem_type),
272 LLVMGetVectorSize(t));
273 }
274 return to_float_type_scalar(ctx, t);
275 }
276
277 LLVMValueRef
278 ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v)
279 {
280 LLVMTypeRef type = LLVMTypeOf(v);
281 return LLVMBuildBitCast(ctx->builder, v, ac_to_float_type(ctx, type), "");
282 }
283
284
285 LLVMValueRef
286 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
287 LLVMTypeRef return_type, LLVMValueRef *params,
288 unsigned param_count, unsigned attrib_mask)
289 {
290 LLVMValueRef function, call;
291 bool set_callsite_attrs = !(attrib_mask & AC_FUNC_ATTR_LEGACY);
292
293 function = LLVMGetNamedFunction(ctx->module, name);
294 if (!function) {
295 LLVMTypeRef param_types[32], function_type;
296 unsigned i;
297
298 assert(param_count <= 32);
299
300 for (i = 0; i < param_count; ++i) {
301 assert(params[i]);
302 param_types[i] = LLVMTypeOf(params[i]);
303 }
304 function_type =
305 LLVMFunctionType(return_type, param_types, param_count, 0);
306 function = LLVMAddFunction(ctx->module, name, function_type);
307
308 LLVMSetFunctionCallConv(function, LLVMCCallConv);
309 LLVMSetLinkage(function, LLVMExternalLinkage);
310
311 if (!set_callsite_attrs)
312 ac_add_func_attributes(ctx->context, function, attrib_mask);
313 }
314
315 call = LLVMBuildCall(ctx->builder, function, params, param_count, "");
316 if (set_callsite_attrs)
317 ac_add_func_attributes(ctx->context, call, attrib_mask);
318 return call;
319 }
320
321 /**
322 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
323 * intrinsic names).
324 */
325 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize)
326 {
327 LLVMTypeRef elem_type = type;
328
329 assert(bufsize >= 8);
330
331 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
332 int ret = snprintf(buf, bufsize, "v%u",
333 LLVMGetVectorSize(type));
334 if (ret < 0) {
335 char *type_name = LLVMPrintTypeToString(type);
336 fprintf(stderr, "Error building type name for: %s\n",
337 type_name);
338 return;
339 }
340 elem_type = LLVMGetElementType(type);
341 buf += ret;
342 bufsize -= ret;
343 }
344 switch (LLVMGetTypeKind(elem_type)) {
345 default: break;
346 case LLVMIntegerTypeKind:
347 snprintf(buf, bufsize, "i%d", LLVMGetIntTypeWidth(elem_type));
348 break;
349 case LLVMHalfTypeKind:
350 snprintf(buf, bufsize, "f16");
351 break;
352 case LLVMFloatTypeKind:
353 snprintf(buf, bufsize, "f32");
354 break;
355 case LLVMDoubleTypeKind:
356 snprintf(buf, bufsize, "f64");
357 break;
358 }
359 }
360
361 /**
362 * Helper function that builds an LLVM IR PHI node and immediately adds
363 * incoming edges.
364 */
365 LLVMValueRef
366 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
367 unsigned count_incoming, LLVMValueRef *values,
368 LLVMBasicBlockRef *blocks)
369 {
370 LLVMValueRef phi = LLVMBuildPhi(ctx->builder, type, "");
371 LLVMAddIncoming(phi, values, blocks, count_incoming);
372 return phi;
373 }
374
375 void ac_build_s_barrier(struct ac_llvm_context *ctx)
376 {
377 ac_build_intrinsic(ctx, "llvm.amdgcn.s.barrier", ctx->voidt, NULL,
378 0, AC_FUNC_ATTR_CONVERGENT);
379 }
380
381 /* Prevent optimizations (at least of memory accesses) across the current
382 * point in the program by emitting empty inline assembly that is marked as
383 * having side effects.
384 *
385 * Optionally, a value can be passed through the inline assembly to prevent
386 * LLVM from hoisting calls to ReadNone functions.
387 */
388 void
389 ac_build_optimization_barrier(struct ac_llvm_context *ctx,
390 LLVMValueRef *pvgpr)
391 {
392 static int counter = 0;
393
394 LLVMBuilderRef builder = ctx->builder;
395 char code[16];
396
397 snprintf(code, sizeof(code), "; %d", p_atomic_inc_return(&counter));
398
399 if (!pvgpr) {
400 LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
401 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "", true, false);
402 LLVMBuildCall(builder, inlineasm, NULL, 0, "");
403 } else {
404 LLVMTypeRef ftype = LLVMFunctionType(ctx->i32, &ctx->i32, 1, false);
405 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "=v,0", true, false);
406 LLVMValueRef vgpr = *pvgpr;
407 LLVMTypeRef vgpr_type = LLVMTypeOf(vgpr);
408 unsigned vgpr_size = ac_get_type_size(vgpr_type);
409 LLVMValueRef vgpr0;
410
411 assert(vgpr_size % 4 == 0);
412
413 vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
414 vgpr0 = LLVMBuildExtractElement(builder, vgpr, ctx->i32_0, "");
415 vgpr0 = LLVMBuildCall(builder, inlineasm, &vgpr0, 1, "");
416 vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
417 vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
418
419 *pvgpr = vgpr;
420 }
421 }
422
423 LLVMValueRef
424 ac_build_shader_clock(struct ac_llvm_context *ctx)
425 {
426 LLVMValueRef tmp = ac_build_intrinsic(ctx, "llvm.readcyclecounter",
427 ctx->i64, NULL, 0, 0);
428 return LLVMBuildBitCast(ctx->builder, tmp, ctx->v2i32, "");
429 }
430
431 LLVMValueRef
432 ac_build_ballot(struct ac_llvm_context *ctx,
433 LLVMValueRef value)
434 {
435 LLVMValueRef args[3] = {
436 value,
437 ctx->i32_0,
438 LLVMConstInt(ctx->i32, LLVMIntNE, 0)
439 };
440
441 /* We currently have no other way to prevent LLVM from lifting the icmp
442 * calls to a dominating basic block.
443 */
444 ac_build_optimization_barrier(ctx, &args[0]);
445
446 args[0] = ac_to_integer(ctx, args[0]);
447
448 return ac_build_intrinsic(ctx,
449 "llvm.amdgcn.icmp.i32",
450 ctx->i64, args, 3,
451 AC_FUNC_ATTR_NOUNWIND |
452 AC_FUNC_ATTR_READNONE |
453 AC_FUNC_ATTR_CONVERGENT);
454 }
455
456 LLVMValueRef
457 ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value)
458 {
459 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
460 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
461 return LLVMBuildICmp(ctx->builder, LLVMIntEQ, vote_set, active_set, "");
462 }
463
464 LLVMValueRef
465 ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value)
466 {
467 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
468 return LLVMBuildICmp(ctx->builder, LLVMIntNE, vote_set,
469 LLVMConstInt(ctx->i64, 0, 0), "");
470 }
471
472 LLVMValueRef
473 ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value)
474 {
475 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
476 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
477
478 LLVMValueRef all = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
479 vote_set, active_set, "");
480 LLVMValueRef none = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
481 vote_set,
482 LLVMConstInt(ctx->i64, 0, 0), "");
483 return LLVMBuildOr(ctx->builder, all, none, "");
484 }
485
486 LLVMValueRef
487 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
488 unsigned value_count, unsigned component)
489 {
490 LLVMValueRef vec = NULL;
491
492 if (value_count == 1) {
493 return values[component];
494 } else if (!value_count)
495 unreachable("value_count is 0");
496
497 for (unsigned i = component; i < value_count + component; i++) {
498 LLVMValueRef value = values[i];
499
500 if (i == component)
501 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
502 LLVMValueRef index = LLVMConstInt(ctx->i32, i - component, false);
503 vec = LLVMBuildInsertElement(ctx->builder, vec, value, index, "");
504 }
505 return vec;
506 }
507
508 LLVMValueRef
509 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
510 LLVMValueRef *values,
511 unsigned value_count,
512 unsigned value_stride,
513 bool load,
514 bool always_vector)
515 {
516 LLVMBuilderRef builder = ctx->builder;
517 LLVMValueRef vec = NULL;
518 unsigned i;
519
520 if (value_count == 1 && !always_vector) {
521 if (load)
522 return LLVMBuildLoad(builder, values[0], "");
523 return values[0];
524 } else if (!value_count)
525 unreachable("value_count is 0");
526
527 for (i = 0; i < value_count; i++) {
528 LLVMValueRef value = values[i * value_stride];
529 if (load)
530 value = LLVMBuildLoad(builder, value, "");
531
532 if (!i)
533 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
534 LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
535 vec = LLVMBuildInsertElement(builder, vec, value, index, "");
536 }
537 return vec;
538 }
539
540 LLVMValueRef
541 ac_build_gather_values(struct ac_llvm_context *ctx,
542 LLVMValueRef *values,
543 unsigned value_count)
544 {
545 return ac_build_gather_values_extended(ctx, values, value_count, 1, false, false);
546 }
547
548 /* Expand a scalar or vector to <dst_channels x type> by filling the remaining
549 * channels with undef. Extract at most src_channels components from the input.
550 */
551 static LLVMValueRef
552 ac_build_expand(struct ac_llvm_context *ctx,
553 LLVMValueRef value,
554 unsigned src_channels,
555 unsigned dst_channels)
556 {
557 LLVMTypeRef elemtype;
558 LLVMValueRef chan[dst_channels];
559
560 if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMVectorTypeKind) {
561 unsigned vec_size = LLVMGetVectorSize(LLVMTypeOf(value));
562
563 if (src_channels == dst_channels && vec_size == dst_channels)
564 return value;
565
566 src_channels = MIN2(src_channels, vec_size);
567
568 for (unsigned i = 0; i < src_channels; i++)
569 chan[i] = ac_llvm_extract_elem(ctx, value, i);
570
571 elemtype = LLVMGetElementType(LLVMTypeOf(value));
572 } else {
573 if (src_channels) {
574 assert(src_channels == 1);
575 chan[0] = value;
576 }
577 elemtype = LLVMTypeOf(value);
578 }
579
580 for (unsigned i = src_channels; i < dst_channels; i++)
581 chan[i] = LLVMGetUndef(elemtype);
582
583 return ac_build_gather_values(ctx, chan, dst_channels);
584 }
585
586 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
587 * with undef. Extract at most num_channels components from the input.
588 */
589 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
590 LLVMValueRef value,
591 unsigned num_channels)
592 {
593 return ac_build_expand(ctx, value, num_channels, 4);
594 }
595
596 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value)
597 {
598 unsigned type_size = ac_get_type_size(LLVMTypeOf(value));
599 const char *name;
600
601 if (type_size == 2)
602 name = "llvm.rint.f16";
603 else if (type_size == 4)
604 name = "llvm.rint.f32";
605 else
606 name = "llvm.rint.f64";
607
608 return ac_build_intrinsic(ctx, name, LLVMTypeOf(value), &value, 1,
609 AC_FUNC_ATTR_READNONE);
610 }
611
612 LLVMValueRef
613 ac_build_fdiv(struct ac_llvm_context *ctx,
614 LLVMValueRef num,
615 LLVMValueRef den)
616 {
617 /* If we do (num / den), LLVM >= 7.0 does:
618 * return num * v_rcp_f32(den * (fabs(den) > 0x1.0p+96f ? 0x1.0p-32f : 1.0f));
619 *
620 * If we do (num * (1 / den)), LLVM does:
621 * return num * v_rcp_f32(den);
622 */
623 LLVMValueRef one = LLVMConstReal(LLVMTypeOf(num), 1.0);
624 LLVMValueRef rcp = LLVMBuildFDiv(ctx->builder, one, den, "");
625 LLVMValueRef ret = LLVMBuildFMul(ctx->builder, num, rcp, "");
626
627 /* Use v_rcp_f32 instead of precise division. */
628 if (!LLVMIsConstant(ret))
629 LLVMSetMetadata(ret, ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
630 return ret;
631 }
632
633 /* See fast_idiv_by_const.h. */
634 /* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
635 LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
636 LLVMValueRef num,
637 LLVMValueRef multiplier,
638 LLVMValueRef pre_shift,
639 LLVMValueRef post_shift,
640 LLVMValueRef increment)
641 {
642 LLVMBuilderRef builder = ctx->builder;
643
644 num = LLVMBuildLShr(builder, num, pre_shift, "");
645 num = LLVMBuildMul(builder,
646 LLVMBuildZExt(builder, num, ctx->i64, ""),
647 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
648 num = LLVMBuildAdd(builder, num,
649 LLVMBuildZExt(builder, increment, ctx->i64, ""), "");
650 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
651 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
652 return LLVMBuildLShr(builder, num, post_shift, "");
653 }
654
655 /* See fast_idiv_by_const.h. */
656 /* If num != UINT_MAX, this more efficient version can be used. */
657 /* Set: increment = util_fast_udiv_info::increment; */
658 LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
659 LLVMValueRef num,
660 LLVMValueRef multiplier,
661 LLVMValueRef pre_shift,
662 LLVMValueRef post_shift,
663 LLVMValueRef increment)
664 {
665 LLVMBuilderRef builder = ctx->builder;
666
667 num = LLVMBuildLShr(builder, num, pre_shift, "");
668 num = LLVMBuildNUWAdd(builder, num, increment, "");
669 num = LLVMBuildMul(builder,
670 LLVMBuildZExt(builder, num, ctx->i64, ""),
671 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
672 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
673 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
674 return LLVMBuildLShr(builder, num, post_shift, "");
675 }
676
677 /* See fast_idiv_by_const.h. */
678 /* Both operands must fit in 31 bits and the divisor must not be 1. */
679 LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
680 LLVMValueRef num,
681 LLVMValueRef multiplier,
682 LLVMValueRef post_shift)
683 {
684 LLVMBuilderRef builder = ctx->builder;
685
686 num = LLVMBuildMul(builder,
687 LLVMBuildZExt(builder, num, ctx->i64, ""),
688 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
689 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
690 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
691 return LLVMBuildLShr(builder, num, post_shift, "");
692 }
693
694 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
695 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
696 * already multiplied by two. id is the cube face number.
697 */
698 struct cube_selection_coords {
699 LLVMValueRef stc[2];
700 LLVMValueRef ma;
701 LLVMValueRef id;
702 };
703
704 static void
705 build_cube_intrinsic(struct ac_llvm_context *ctx,
706 LLVMValueRef in[3],
707 struct cube_selection_coords *out)
708 {
709 LLVMTypeRef f32 = ctx->f32;
710
711 out->stc[1] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubetc",
712 f32, in, 3, AC_FUNC_ATTR_READNONE);
713 out->stc[0] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubesc",
714 f32, in, 3, AC_FUNC_ATTR_READNONE);
715 out->ma = ac_build_intrinsic(ctx, "llvm.amdgcn.cubema",
716 f32, in, 3, AC_FUNC_ATTR_READNONE);
717 out->id = ac_build_intrinsic(ctx, "llvm.amdgcn.cubeid",
718 f32, in, 3, AC_FUNC_ATTR_READNONE);
719 }
720
721 /**
722 * Build a manual selection sequence for cube face sc/tc coordinates and
723 * major axis vector (multiplied by 2 for consistency) for the given
724 * vec3 \p coords, for the face implied by \p selcoords.
725 *
726 * For the major axis, we always adjust the sign to be in the direction of
727 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
728 * the selcoords major axis.
729 */
730 static void build_cube_select(struct ac_llvm_context *ctx,
731 const struct cube_selection_coords *selcoords,
732 const LLVMValueRef *coords,
733 LLVMValueRef *out_st,
734 LLVMValueRef *out_ma)
735 {
736 LLVMBuilderRef builder = ctx->builder;
737 LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
738 LLVMValueRef is_ma_positive;
739 LLVMValueRef sgn_ma;
740 LLVMValueRef is_ma_z, is_not_ma_z;
741 LLVMValueRef is_ma_y;
742 LLVMValueRef is_ma_x;
743 LLVMValueRef sgn;
744 LLVMValueRef tmp;
745
746 is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
747 selcoords->ma, LLVMConstReal(f32, 0.0), "");
748 sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
749 LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
750
751 is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
752 is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
753 is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
754 LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
755 is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
756
757 /* Select sc */
758 tmp = LLVMBuildSelect(builder, is_ma_x, coords[2], coords[0], "");
759 sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
760 LLVMBuildSelect(builder, is_ma_z, sgn_ma,
761 LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
762 out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
763
764 /* Select tc */
765 tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
766 sgn = LLVMBuildSelect(builder, is_ma_y, sgn_ma,
767 LLVMConstReal(f32, -1.0), "");
768 out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
769
770 /* Select ma */
771 tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
772 LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
773 tmp = ac_build_intrinsic(ctx, "llvm.fabs.f32",
774 ctx->f32, &tmp, 1, AC_FUNC_ATTR_READNONE);
775 *out_ma = LLVMBuildFMul(builder, tmp, LLVMConstReal(f32, 2.0), "");
776 }
777
778 void
779 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
780 bool is_deriv, bool is_array, bool is_lod,
781 LLVMValueRef *coords_arg,
782 LLVMValueRef *derivs_arg)
783 {
784
785 LLVMBuilderRef builder = ctx->builder;
786 struct cube_selection_coords selcoords;
787 LLVMValueRef coords[3];
788 LLVMValueRef invma;
789
790 if (is_array && !is_lod) {
791 LLVMValueRef tmp = ac_build_round(ctx, coords_arg[3]);
792
793 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
794 *
795 * "For Array forms, the array layer used will be
796 *
797 * max(0, min(d−1, floor(layer+0.5)))
798 *
799 * where d is the depth of the texture array and layer
800 * comes from the component indicated in the tables below.
801 * Workaroudn for an issue where the layer is taken from a
802 * helper invocation which happens to fall on a different
803 * layer due to extrapolation."
804 *
805 * VI and earlier attempt to implement this in hardware by
806 * clamping the value of coords[2] = (8 * layer) + face.
807 * Unfortunately, this means that the we end up with the wrong
808 * face when clamping occurs.
809 *
810 * Clamp the layer earlier to work around the issue.
811 */
812 if (ctx->chip_class <= VI) {
813 LLVMValueRef ge0;
814 ge0 = LLVMBuildFCmp(builder, LLVMRealOGE, tmp, ctx->f32_0, "");
815 tmp = LLVMBuildSelect(builder, ge0, tmp, ctx->f32_0, "");
816 }
817
818 coords_arg[3] = tmp;
819 }
820
821 build_cube_intrinsic(ctx, coords_arg, &selcoords);
822
823 invma = ac_build_intrinsic(ctx, "llvm.fabs.f32",
824 ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
825 invma = ac_build_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
826
827 for (int i = 0; i < 2; ++i)
828 coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
829
830 coords[2] = selcoords.id;
831
832 if (is_deriv && derivs_arg) {
833 LLVMValueRef derivs[4];
834 int axis;
835
836 /* Convert cube derivatives to 2D derivatives. */
837 for (axis = 0; axis < 2; axis++) {
838 LLVMValueRef deriv_st[2];
839 LLVMValueRef deriv_ma;
840
841 /* Transform the derivative alongside the texture
842 * coordinate. Mathematically, the correct formula is
843 * as follows. Assume we're projecting onto the +Z face
844 * and denote by dx/dh the derivative of the (original)
845 * X texture coordinate with respect to horizontal
846 * window coordinates. The projection onto the +Z face
847 * plane is:
848 *
849 * f(x,z) = x/z
850 *
851 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
852 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
853 *
854 * This motivatives the implementation below.
855 *
856 * Whether this actually gives the expected results for
857 * apps that might feed in derivatives obtained via
858 * finite differences is anyone's guess. The OpenGL spec
859 * seems awfully quiet about how textureGrad for cube
860 * maps should be handled.
861 */
862 build_cube_select(ctx, &selcoords, &derivs_arg[axis * 3],
863 deriv_st, &deriv_ma);
864
865 deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
866
867 for (int i = 0; i < 2; ++i)
868 derivs[axis * 2 + i] =
869 LLVMBuildFSub(builder,
870 LLVMBuildFMul(builder, deriv_st[i], invma, ""),
871 LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
872 }
873
874 memcpy(derivs_arg, derivs, sizeof(derivs));
875 }
876
877 /* Shift the texture coordinate. This must be applied after the
878 * derivative calculation.
879 */
880 for (int i = 0; i < 2; ++i)
881 coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
882
883 if (is_array) {
884 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
885 /* coords_arg.w component - array_index for cube arrays */
886 coords[2] = ac_build_fmad(ctx, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), coords[2]);
887 }
888
889 memcpy(coords_arg, coords, sizeof(coords));
890 }
891
892
893 LLVMValueRef
894 ac_build_fs_interp(struct ac_llvm_context *ctx,
895 LLVMValueRef llvm_chan,
896 LLVMValueRef attr_number,
897 LLVMValueRef params,
898 LLVMValueRef i,
899 LLVMValueRef j)
900 {
901 LLVMValueRef args[5];
902 LLVMValueRef p1;
903
904 args[0] = i;
905 args[1] = llvm_chan;
906 args[2] = attr_number;
907 args[3] = params;
908
909 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1",
910 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
911
912 args[0] = p1;
913 args[1] = j;
914 args[2] = llvm_chan;
915 args[3] = attr_number;
916 args[4] = params;
917
918 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2",
919 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
920 }
921
922 LLVMValueRef
923 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
924 LLVMValueRef llvm_chan,
925 LLVMValueRef attr_number,
926 LLVMValueRef params,
927 LLVMValueRef i,
928 LLVMValueRef j)
929 {
930 LLVMValueRef args[6];
931 LLVMValueRef p1;
932
933 args[0] = i;
934 args[1] = llvm_chan;
935 args[2] = attr_number;
936 args[3] = ctx->i1false;
937 args[4] = params;
938
939 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1.f16",
940 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
941
942 args[0] = p1;
943 args[1] = j;
944 args[2] = llvm_chan;
945 args[3] = attr_number;
946 args[4] = ctx->i1false;
947 args[5] = params;
948
949 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2.f16",
950 ctx->f16, args, 6, AC_FUNC_ATTR_READNONE);
951 }
952
953 LLVMValueRef
954 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
955 LLVMValueRef parameter,
956 LLVMValueRef llvm_chan,
957 LLVMValueRef attr_number,
958 LLVMValueRef params)
959 {
960 LLVMValueRef args[4];
961
962 args[0] = parameter;
963 args[1] = llvm_chan;
964 args[2] = attr_number;
965 args[3] = params;
966
967 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.mov",
968 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
969 }
970
971 LLVMValueRef
972 ac_build_gep_ptr(struct ac_llvm_context *ctx,
973 LLVMValueRef base_ptr,
974 LLVMValueRef index)
975 {
976 return LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
977 }
978
979 LLVMValueRef
980 ac_build_gep0(struct ac_llvm_context *ctx,
981 LLVMValueRef base_ptr,
982 LLVMValueRef index)
983 {
984 LLVMValueRef indices[2] = {
985 ctx->i32_0,
986 index,
987 };
988 return LLVMBuildGEP(ctx->builder, base_ptr, indices, 2, "");
989 }
990
991 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
992 LLVMValueRef index)
993 {
994 return LLVMBuildPointerCast(ctx->builder,
995 ac_build_gep0(ctx, ptr, index),
996 LLVMTypeOf(ptr), "");
997 }
998
999 void
1000 ac_build_indexed_store(struct ac_llvm_context *ctx,
1001 LLVMValueRef base_ptr, LLVMValueRef index,
1002 LLVMValueRef value)
1003 {
1004 LLVMBuildStore(ctx->builder, value,
1005 ac_build_gep0(ctx, base_ptr, index));
1006 }
1007
1008 /**
1009 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
1010 * It's equivalent to doing a load from &base_ptr[index].
1011 *
1012 * \param base_ptr Where the array starts.
1013 * \param index The element index into the array.
1014 * \param uniform Whether the base_ptr and index can be assumed to be
1015 * dynamically uniform (i.e. load to an SGPR)
1016 * \param invariant Whether the load is invariant (no other opcodes affect it)
1017 * \param no_unsigned_wraparound
1018 * For all possible re-associations and re-distributions of an expression
1019 * "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
1020 * without inbounds in base_ptr), this parameter is true if "addr + offset"
1021 * does not result in an unsigned integer wraparound. This is used for
1022 * optimal code generation of 32-bit pointer arithmetic.
1023 *
1024 * For example, a 32-bit immediate offset that causes a 32-bit unsigned
1025 * integer wraparound can't be an imm offset in s_load_dword, because
1026 * the instruction performs "addr + offset" in 64 bits.
1027 *
1028 * Expected usage for bindless textures by chaining GEPs:
1029 * // possible unsigned wraparound, don't use InBounds:
1030 * ptr1 = LLVMBuildGEP(base_ptr, index);
1031 * image = load(ptr1); // becomes "s_load ptr1, 0"
1032 *
1033 * ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1034 * sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1035 */
1036 static LLVMValueRef
1037 ac_build_load_custom(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1038 LLVMValueRef index, bool uniform, bool invariant,
1039 bool no_unsigned_wraparound)
1040 {
1041 LLVMValueRef pointer, result;
1042 LLVMValueRef indices[2] = {ctx->i32_0, index};
1043
1044 if (no_unsigned_wraparound &&
1045 LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr)) == AC_ADDR_SPACE_CONST_32BIT)
1046 pointer = LLVMBuildInBoundsGEP(ctx->builder, base_ptr, indices, 2, "");
1047 else
1048 pointer = LLVMBuildGEP(ctx->builder, base_ptr, indices, 2, "");
1049
1050 if (uniform)
1051 LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
1052 result = LLVMBuildLoad(ctx->builder, pointer, "");
1053 if (invariant)
1054 LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
1055 return result;
1056 }
1057
1058 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1059 LLVMValueRef index)
1060 {
1061 return ac_build_load_custom(ctx, base_ptr, index, false, false, false);
1062 }
1063
1064 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
1065 LLVMValueRef base_ptr, LLVMValueRef index)
1066 {
1067 return ac_build_load_custom(ctx, base_ptr, index, false, true, false);
1068 }
1069
1070 /* This assumes that there is no unsigned integer wraparound during the address
1071 * computation, excluding all GEPs within base_ptr. */
1072 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
1073 LLVMValueRef base_ptr, LLVMValueRef index)
1074 {
1075 return ac_build_load_custom(ctx, base_ptr, index, true, true, true);
1076 }
1077
1078 /* See ac_build_load_custom() documentation. */
1079 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
1080 LLVMValueRef base_ptr, LLVMValueRef index)
1081 {
1082 return ac_build_load_custom(ctx, base_ptr, index, true, true, false);
1083 }
1084
1085 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1086 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1087 * or v4i32 (num_channels=3,4).
1088 */
1089 void
1090 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
1091 LLVMValueRef rsrc,
1092 LLVMValueRef vdata,
1093 unsigned num_channels,
1094 LLVMValueRef voffset,
1095 LLVMValueRef soffset,
1096 unsigned inst_offset,
1097 bool glc,
1098 bool slc,
1099 bool writeonly_memory,
1100 bool swizzle_enable_hint)
1101 {
1102 /* Split 3 channel stores, becase LLVM doesn't support 3-channel
1103 * intrinsics. */
1104 if (num_channels == 3) {
1105 LLVMValueRef v[3], v01;
1106
1107 for (int i = 0; i < 3; i++) {
1108 v[i] = LLVMBuildExtractElement(ctx->builder, vdata,
1109 LLVMConstInt(ctx->i32, i, 0), "");
1110 }
1111 v01 = ac_build_gather_values(ctx, v, 2);
1112
1113 ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset,
1114 soffset, inst_offset, glc, slc,
1115 writeonly_memory, swizzle_enable_hint);
1116 ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset,
1117 soffset, inst_offset + 8,
1118 glc, slc,
1119 writeonly_memory, swizzle_enable_hint);
1120 return;
1121 }
1122
1123 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1124 * (voffset is swizzled, but soffset isn't swizzled).
1125 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1126 */
1127 if (!swizzle_enable_hint) {
1128 LLVMValueRef offset = soffset;
1129
1130 static const char *types[] = {"f32", "v2f32", "v4f32"};
1131
1132 if (inst_offset)
1133 offset = LLVMBuildAdd(ctx->builder, offset,
1134 LLVMConstInt(ctx->i32, inst_offset, 0), "");
1135 if (voffset)
1136 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1137
1138 LLVMValueRef args[] = {
1139 ac_to_float(ctx, vdata),
1140 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
1141 ctx->i32_0,
1142 offset,
1143 LLVMConstInt(ctx->i1, glc, 0),
1144 LLVMConstInt(ctx->i1, slc, 0),
1145 };
1146
1147 char name[256];
1148 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.store.%s",
1149 types[CLAMP(num_channels, 1, 3) - 1]);
1150
1151 ac_build_intrinsic(ctx, name, ctx->voidt,
1152 args, ARRAY_SIZE(args),
1153 writeonly_memory ?
1154 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY :
1155 AC_FUNC_ATTR_WRITEONLY);
1156 return;
1157 }
1158
1159 static const unsigned dfmt[] = {
1160 V_008F0C_BUF_DATA_FORMAT_32,
1161 V_008F0C_BUF_DATA_FORMAT_32_32,
1162 V_008F0C_BUF_DATA_FORMAT_32_32_32,
1163 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1164 };
1165 static const char *types[] = {"i32", "v2i32", "v4i32"};
1166 LLVMValueRef args[] = {
1167 vdata,
1168 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
1169 ctx->i32_0,
1170 voffset ? voffset : ctx->i32_0,
1171 soffset,
1172 LLVMConstInt(ctx->i32, inst_offset, 0),
1173 LLVMConstInt(ctx->i32, dfmt[num_channels - 1], 0),
1174 LLVMConstInt(ctx->i32, V_008F0C_BUF_NUM_FORMAT_UINT, 0),
1175 LLVMConstInt(ctx->i1, glc, 0),
1176 LLVMConstInt(ctx->i1, slc, 0),
1177 };
1178 char name[256];
1179 snprintf(name, sizeof(name), "llvm.amdgcn.tbuffer.store.%s",
1180 types[CLAMP(num_channels, 1, 3) - 1]);
1181
1182 ac_build_intrinsic(ctx, name, ctx->voidt,
1183 args, ARRAY_SIZE(args),
1184 writeonly_memory ?
1185 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY :
1186 AC_FUNC_ATTR_WRITEONLY);
1187 }
1188
1189 static LLVMValueRef
1190 ac_build_buffer_load_common(struct ac_llvm_context *ctx,
1191 LLVMValueRef rsrc,
1192 LLVMValueRef vindex,
1193 LLVMValueRef voffset,
1194 unsigned num_channels,
1195 bool glc,
1196 bool slc,
1197 bool can_speculate,
1198 bool use_format)
1199 {
1200 LLVMValueRef args[] = {
1201 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
1202 vindex ? vindex : ctx->i32_0,
1203 voffset,
1204 LLVMConstInt(ctx->i1, glc, 0),
1205 LLVMConstInt(ctx->i1, slc, 0)
1206 };
1207 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1208
1209 LLVMTypeRef types[] = {ctx->f32, ctx->v2f32, ctx->v4f32};
1210 const char *type_names[] = {"f32", "v2f32", "v4f32"};
1211 char name[256];
1212
1213 if (use_format) {
1214 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.format.%s",
1215 type_names[func]);
1216 } else {
1217 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.%s",
1218 type_names[func]);
1219 }
1220
1221 return ac_build_intrinsic(ctx, name, types[func], args,
1222 ARRAY_SIZE(args),
1223 ac_get_load_intr_attribs(can_speculate));
1224 }
1225
1226 static LLVMValueRef
1227 ac_build_llvm8_buffer_load_common(struct ac_llvm_context *ctx,
1228 LLVMValueRef rsrc,
1229 LLVMValueRef vindex,
1230 LLVMValueRef voffset,
1231 LLVMValueRef soffset,
1232 unsigned num_channels,
1233 bool glc,
1234 bool slc,
1235 bool can_speculate,
1236 bool use_format,
1237 bool structurized)
1238 {
1239 LLVMValueRef args[5];
1240 int idx = 0;
1241 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1242 if (structurized)
1243 args[idx++] = vindex ? vindex : ctx->i32_0;
1244 args[idx++] = voffset ? voffset : ctx->i32_0;
1245 args[idx++] = soffset ? soffset : ctx->i32_0;
1246 args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0);
1247 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1248
1249 LLVMTypeRef types[] = {ctx->f32, ctx->v2f32, ctx->v4f32};
1250 const char *type_names[] = {"f32", "v2f32", "v4f32"};
1251 const char *indexing_kind = structurized ? "struct" : "raw";
1252 char name[256];
1253
1254 if (use_format) {
1255 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.format.%s",
1256 indexing_kind, type_names[func]);
1257 } else {
1258 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.%s",
1259 indexing_kind, type_names[func]);
1260 }
1261
1262 return ac_build_intrinsic(ctx, name, types[func], args,
1263 idx,
1264 ac_get_load_intr_attribs(can_speculate));
1265 }
1266
1267 LLVMValueRef
1268 ac_build_buffer_load(struct ac_llvm_context *ctx,
1269 LLVMValueRef rsrc,
1270 int num_channels,
1271 LLVMValueRef vindex,
1272 LLVMValueRef voffset,
1273 LLVMValueRef soffset,
1274 unsigned inst_offset,
1275 unsigned glc,
1276 unsigned slc,
1277 bool can_speculate,
1278 bool allow_smem)
1279 {
1280 LLVMValueRef offset = LLVMConstInt(ctx->i32, inst_offset, 0);
1281 if (voffset)
1282 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1283 if (soffset)
1284 offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
1285
1286 if (allow_smem && !slc &&
1287 (!glc || (HAVE_LLVM >= 0x0800 && ctx->chip_class >= VI))) {
1288 assert(vindex == NULL);
1289
1290 LLVMValueRef result[8];
1291
1292 for (int i = 0; i < num_channels; i++) {
1293 if (i) {
1294 offset = LLVMBuildAdd(ctx->builder, offset,
1295 LLVMConstInt(ctx->i32, 4, 0), "");
1296 }
1297 const char *intrname =
1298 HAVE_LLVM >= 0x0800 ? "llvm.amdgcn.s.buffer.load.f32"
1299 : "llvm.SI.load.const.v4i32";
1300 unsigned num_args = HAVE_LLVM >= 0x0800 ? 3 : 2;
1301 LLVMValueRef args[3] = {
1302 rsrc,
1303 offset,
1304 glc ? ctx->i32_1 : ctx->i32_0,
1305 };
1306 result[i] = ac_build_intrinsic(ctx, intrname,
1307 ctx->f32, args, num_args,
1308 AC_FUNC_ATTR_READNONE |
1309 (HAVE_LLVM < 0x0800 ? AC_FUNC_ATTR_LEGACY : 0));
1310 }
1311 if (num_channels == 1)
1312 return result[0];
1313
1314 if (num_channels == 3)
1315 result[num_channels++] = LLVMGetUndef(ctx->f32);
1316 return ac_build_gather_values(ctx, result, num_channels);
1317 }
1318
1319 return ac_build_buffer_load_common(ctx, rsrc, vindex, offset,
1320 num_channels, glc, slc,
1321 can_speculate, false);
1322 }
1323
1324 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
1325 LLVMValueRef rsrc,
1326 LLVMValueRef vindex,
1327 LLVMValueRef voffset,
1328 unsigned num_channels,
1329 bool glc,
1330 bool can_speculate)
1331 {
1332 if (HAVE_LLVM >= 0x800) {
1333 return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
1334 num_channels, glc, false,
1335 can_speculate, true, true);
1336 }
1337 return ac_build_buffer_load_common(ctx, rsrc, vindex, voffset,
1338 num_channels, glc, false,
1339 can_speculate, true);
1340 }
1341
1342 LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
1343 LLVMValueRef rsrc,
1344 LLVMValueRef vindex,
1345 LLVMValueRef voffset,
1346 unsigned num_channels,
1347 bool glc,
1348 bool can_speculate)
1349 {
1350 if (HAVE_LLVM >= 0x800) {
1351 return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
1352 num_channels, glc, false,
1353 can_speculate, true, true);
1354 }
1355
1356 LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->builder, rsrc, LLVMConstInt(ctx->i32, 2, 0), "");
1357 LLVMValueRef stride = LLVMBuildExtractElement(ctx->builder, rsrc, ctx->i32_1, "");
1358 stride = LLVMBuildLShr(ctx->builder, stride, LLVMConstInt(ctx->i32, 16, 0), "");
1359
1360 LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->builder,
1361 LLVMBuildICmp(ctx->builder, LLVMIntUGT, elem_count, stride, ""),
1362 elem_count, stride, "");
1363
1364 LLVMValueRef new_rsrc = LLVMBuildInsertElement(ctx->builder, rsrc, new_elem_count,
1365 LLVMConstInt(ctx->i32, 2, 0), "");
1366
1367 return ac_build_buffer_load_common(ctx, new_rsrc, vindex, voffset,
1368 num_channels, glc, false,
1369 can_speculate, true);
1370 }
1371
1372 static LLVMValueRef
1373 ac_build_llvm8_tbuffer_load(struct ac_llvm_context *ctx,
1374 LLVMValueRef rsrc,
1375 LLVMValueRef vindex,
1376 LLVMValueRef voffset,
1377 LLVMValueRef soffset,
1378 unsigned num_channels,
1379 unsigned dfmt,
1380 unsigned nfmt,
1381 bool glc,
1382 bool slc,
1383 bool can_speculate,
1384 bool structurized)
1385 {
1386 LLVMValueRef args[6];
1387 int idx = 0;
1388 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1389 if (structurized)
1390 args[idx++] = vindex ? vindex : ctx->i32_0;
1391 args[idx++] = voffset ? voffset : ctx->i32_0;
1392 args[idx++] = soffset ? soffset : ctx->i32_0;
1393 args[idx++] = LLVMConstInt(ctx->i32, dfmt | (nfmt << 4), 0);
1394 args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0);
1395 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1396
1397 LLVMTypeRef types[] = {ctx->i32, ctx->v2i32, ctx->v4i32};
1398 const char *type_names[] = {"i32", "v2i32", "v4i32"};
1399 const char *indexing_kind = structurized ? "struct" : "raw";
1400 char name[256];
1401
1402 snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.load.%s",
1403 indexing_kind, type_names[func]);
1404
1405 return ac_build_intrinsic(ctx, name, types[func], args,
1406 idx,
1407 ac_get_load_intr_attribs(can_speculate));
1408 }
1409
1410 static LLVMValueRef
1411 ac_build_tbuffer_load(struct ac_llvm_context *ctx,
1412 LLVMValueRef rsrc,
1413 LLVMValueRef vindex,
1414 LLVMValueRef voffset,
1415 LLVMValueRef soffset,
1416 LLVMValueRef immoffset,
1417 unsigned num_channels,
1418 unsigned dfmt,
1419 unsigned nfmt,
1420 bool glc,
1421 bool slc,
1422 bool can_speculate,
1423 bool structurized) /* only matters for LLVM 8+ */
1424 {
1425 if (HAVE_LLVM >= 0x800) {
1426 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1427
1428 return ac_build_llvm8_tbuffer_load(ctx, rsrc, vindex, voffset,
1429 soffset, num_channels,
1430 dfmt, nfmt, glc, slc,
1431 can_speculate, structurized);
1432 }
1433
1434 LLVMValueRef args[] = {
1435 rsrc,
1436 vindex ? vindex : ctx->i32_0,
1437 voffset,
1438 soffset,
1439 immoffset,
1440 LLVMConstInt(ctx->i32, dfmt, false),
1441 LLVMConstInt(ctx->i32, nfmt, false),
1442 LLVMConstInt(ctx->i32, glc, false),
1443 LLVMConstInt(ctx->i32, slc, false),
1444 };
1445 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1446 LLVMTypeRef types[] = {ctx->i32, ctx->v2i32, ctx->v4i32};
1447 const char *type_names[] = {"i32", "v2i32", "v4i32"};
1448 char name[256];
1449
1450 snprintf(name, sizeof(name), "llvm.amdgcn.tbuffer.load.%s",
1451 type_names[func]);
1452
1453 return ac_build_intrinsic(ctx, name, types[func], args, 9,
1454 ac_get_load_intr_attribs(can_speculate));
1455 }
1456
1457 LLVMValueRef
1458 ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
1459 LLVMValueRef rsrc,
1460 LLVMValueRef vindex,
1461 LLVMValueRef voffset,
1462 LLVMValueRef soffset,
1463 LLVMValueRef immoffset,
1464 unsigned num_channels,
1465 unsigned dfmt,
1466 unsigned nfmt,
1467 bool glc,
1468 bool slc,
1469 bool can_speculate)
1470 {
1471 return ac_build_tbuffer_load(ctx, rsrc, vindex, voffset, soffset,
1472 immoffset, num_channels, dfmt, nfmt, glc,
1473 slc, can_speculate, true);
1474 }
1475
1476 LLVMValueRef
1477 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
1478 LLVMValueRef rsrc,
1479 LLVMValueRef voffset,
1480 LLVMValueRef soffset,
1481 LLVMValueRef immoffset,
1482 unsigned num_channels,
1483 unsigned dfmt,
1484 unsigned nfmt,
1485 bool glc,
1486 bool slc,
1487 bool can_speculate)
1488 {
1489 return ac_build_tbuffer_load(ctx, rsrc, NULL, voffset, soffset,
1490 immoffset, num_channels, dfmt, nfmt, glc,
1491 slc, can_speculate, false);
1492 }
1493
1494 LLVMValueRef
1495 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
1496 LLVMValueRef rsrc,
1497 LLVMValueRef voffset,
1498 LLVMValueRef soffset,
1499 LLVMValueRef immoffset,
1500 bool glc)
1501 {
1502 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
1503 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1504 LLVMValueRef res;
1505
1506 res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
1507 immoffset, 1, dfmt, nfmt, glc, false,
1508 false);
1509
1510 return LLVMBuildTrunc(ctx->builder, res, ctx->i16, "");
1511 }
1512
1513 /**
1514 * Set range metadata on an instruction. This can only be used on load and
1515 * call instructions. If you know an instruction can only produce the values
1516 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1517 * \p lo is the minimum value inclusive.
1518 * \p hi is the maximum value exclusive.
1519 */
1520 static void set_range_metadata(struct ac_llvm_context *ctx,
1521 LLVMValueRef value, unsigned lo, unsigned hi)
1522 {
1523 LLVMValueRef range_md, md_args[2];
1524 LLVMTypeRef type = LLVMTypeOf(value);
1525 LLVMContextRef context = LLVMGetTypeContext(type);
1526
1527 md_args[0] = LLVMConstInt(type, lo, false);
1528 md_args[1] = LLVMConstInt(type, hi, false);
1529 range_md = LLVMMDNodeInContext(context, md_args, 2);
1530 LLVMSetMetadata(value, ctx->range_md_kind, range_md);
1531 }
1532
1533 LLVMValueRef
1534 ac_get_thread_id(struct ac_llvm_context *ctx)
1535 {
1536 LLVMValueRef tid;
1537
1538 LLVMValueRef tid_args[2];
1539 tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
1540 tid_args[1] = ctx->i32_0;
1541 tid_args[1] = ac_build_intrinsic(ctx,
1542 "llvm.amdgcn.mbcnt.lo", ctx->i32,
1543 tid_args, 2, AC_FUNC_ATTR_READNONE);
1544
1545 tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
1546 ctx->i32, tid_args,
1547 2, AC_FUNC_ATTR_READNONE);
1548 set_range_metadata(ctx, tid, 0, 64);
1549 return tid;
1550 }
1551
1552 /*
1553 * SI implements derivatives using the local data store (LDS)
1554 * All writes to the LDS happen in all executing threads at
1555 * the same time. TID is the Thread ID for the current
1556 * thread and is a value between 0 and 63, representing
1557 * the thread's position in the wavefront.
1558 *
1559 * For the pixel shader threads are grouped into quads of four pixels.
1560 * The TIDs of the pixels of a quad are:
1561 *
1562 * +------+------+
1563 * |4n + 0|4n + 1|
1564 * +------+------+
1565 * |4n + 2|4n + 3|
1566 * +------+------+
1567 *
1568 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
1569 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
1570 * the current pixel's column, and masking with 0xfffffffe yields the TID
1571 * of the left pixel of the current pixel's row.
1572 *
1573 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
1574 * adding 2 yields the TID of the pixel below the top pixel.
1575 */
1576 LLVMValueRef
1577 ac_build_ddxy(struct ac_llvm_context *ctx,
1578 uint32_t mask,
1579 int idx,
1580 LLVMValueRef val)
1581 {
1582 unsigned tl_lanes[4], trbl_lanes[4];
1583 LLVMValueRef tl, trbl;
1584 LLVMValueRef result;
1585
1586 for (unsigned i = 0; i < 4; ++i) {
1587 tl_lanes[i] = i & mask;
1588 trbl_lanes[i] = (i & mask) + idx;
1589 }
1590
1591 tl = ac_build_quad_swizzle(ctx, val,
1592 tl_lanes[0], tl_lanes[1],
1593 tl_lanes[2], tl_lanes[3]);
1594 trbl = ac_build_quad_swizzle(ctx, val,
1595 trbl_lanes[0], trbl_lanes[1],
1596 trbl_lanes[2], trbl_lanes[3]);
1597
1598 tl = LLVMBuildBitCast(ctx->builder, tl, ctx->f32, "");
1599 trbl = LLVMBuildBitCast(ctx->builder, trbl, ctx->f32, "");
1600 result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
1601
1602 result = ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.f32", ctx->f32,
1603 &result, 1, 0);
1604
1605 return result;
1606 }
1607
1608 void
1609 ac_build_sendmsg(struct ac_llvm_context *ctx,
1610 uint32_t msg,
1611 LLVMValueRef wave_id)
1612 {
1613 LLVMValueRef args[2];
1614 args[0] = LLVMConstInt(ctx->i32, msg, false);
1615 args[1] = wave_id;
1616 ac_build_intrinsic(ctx, "llvm.amdgcn.s.sendmsg", ctx->voidt, args, 2, 0);
1617 }
1618
1619 LLVMValueRef
1620 ac_build_imsb(struct ac_llvm_context *ctx,
1621 LLVMValueRef arg,
1622 LLVMTypeRef dst_type)
1623 {
1624 LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.amdgcn.sffbh.i32",
1625 dst_type, &arg, 1,
1626 AC_FUNC_ATTR_READNONE);
1627
1628 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
1629 * the index from LSB. Invert it by doing "31 - msb". */
1630 msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
1631 msb, "");
1632
1633 LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
1634 LLVMValueRef cond = LLVMBuildOr(ctx->builder,
1635 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
1636 arg, ctx->i32_0, ""),
1637 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
1638 arg, all_ones, ""), "");
1639
1640 return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
1641 }
1642
1643 LLVMValueRef
1644 ac_build_umsb(struct ac_llvm_context *ctx,
1645 LLVMValueRef arg,
1646 LLVMTypeRef dst_type)
1647 {
1648 const char *intrin_name;
1649 LLVMTypeRef type;
1650 LLVMValueRef highest_bit;
1651 LLVMValueRef zero;
1652 unsigned bitsize;
1653
1654 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(arg));
1655 switch (bitsize) {
1656 case 64:
1657 intrin_name = "llvm.ctlz.i64";
1658 type = ctx->i64;
1659 highest_bit = LLVMConstInt(ctx->i64, 63, false);
1660 zero = ctx->i64_0;
1661 break;
1662 case 32:
1663 intrin_name = "llvm.ctlz.i32";
1664 type = ctx->i32;
1665 highest_bit = LLVMConstInt(ctx->i32, 31, false);
1666 zero = ctx->i32_0;
1667 break;
1668 case 16:
1669 intrin_name = "llvm.ctlz.i16";
1670 type = ctx->i16;
1671 highest_bit = LLVMConstInt(ctx->i16, 15, false);
1672 zero = ctx->i16_0;
1673 break;
1674 default:
1675 unreachable(!"invalid bitsize");
1676 break;
1677 }
1678
1679 LLVMValueRef params[2] = {
1680 arg,
1681 ctx->i1true,
1682 };
1683
1684 LLVMValueRef msb = ac_build_intrinsic(ctx, intrin_name, type,
1685 params, 2,
1686 AC_FUNC_ATTR_READNONE);
1687
1688 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
1689 * the index from LSB. Invert it by doing "31 - msb". */
1690 msb = LLVMBuildSub(ctx->builder, highest_bit, msb, "");
1691 msb = LLVMBuildTruncOrBitCast(ctx->builder, msb, ctx->i32, "");
1692
1693 /* check for zero */
1694 return LLVMBuildSelect(ctx->builder,
1695 LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg, zero, ""),
1696 LLVMConstInt(ctx->i32, -1, true), msb, "");
1697 }
1698
1699 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
1700 LLVMValueRef b)
1701 {
1702 char name[64];
1703 snprintf(name, sizeof(name), "llvm.minnum.f%d", ac_get_elem_bits(ctx, LLVMTypeOf(a)));
1704 LLVMValueRef args[2] = {a, b};
1705 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
1706 AC_FUNC_ATTR_READNONE);
1707 }
1708
1709 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
1710 LLVMValueRef b)
1711 {
1712 char name[64];
1713 snprintf(name, sizeof(name), "llvm.maxnum.f%d", ac_get_elem_bits(ctx, LLVMTypeOf(a)));
1714 LLVMValueRef args[2] = {a, b};
1715 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
1716 AC_FUNC_ATTR_READNONE);
1717 }
1718
1719 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
1720 LLVMValueRef b)
1721 {
1722 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSLE, a, b, "");
1723 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
1724 }
1725
1726 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
1727 LLVMValueRef b)
1728 {
1729 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, a, b, "");
1730 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
1731 }
1732
1733 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a,
1734 LLVMValueRef b)
1735 {
1736 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntULE, a, b, "");
1737 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
1738 }
1739
1740 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value)
1741 {
1742 LLVMTypeRef t = LLVMTypeOf(value);
1743 return ac_build_fmin(ctx, ac_build_fmax(ctx, value, LLVMConstReal(t, 0.0)),
1744 LLVMConstReal(t, 1.0));
1745 }
1746
1747 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a)
1748 {
1749 LLVMValueRef args[9];
1750
1751 args[0] = LLVMConstInt(ctx->i32, a->target, 0);
1752 args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
1753
1754 if (a->compr) {
1755 LLVMTypeRef i16 = LLVMInt16TypeInContext(ctx->context);
1756 LLVMTypeRef v2i16 = LLVMVectorType(i16, 2);
1757
1758 args[2] = LLVMBuildBitCast(ctx->builder, a->out[0],
1759 v2i16, "");
1760 args[3] = LLVMBuildBitCast(ctx->builder, a->out[1],
1761 v2i16, "");
1762 args[4] = LLVMConstInt(ctx->i1, a->done, 0);
1763 args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
1764
1765 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16",
1766 ctx->voidt, args, 6, 0);
1767 } else {
1768 args[2] = a->out[0];
1769 args[3] = a->out[1];
1770 args[4] = a->out[2];
1771 args[5] = a->out[3];
1772 args[6] = LLVMConstInt(ctx->i1, a->done, 0);
1773 args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
1774
1775 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.f32",
1776 ctx->voidt, args, 8, 0);
1777 }
1778 }
1779
1780 void ac_build_export_null(struct ac_llvm_context *ctx)
1781 {
1782 struct ac_export_args args;
1783
1784 args.enabled_channels = 0x0; /* enabled channels */
1785 args.valid_mask = 1; /* whether the EXEC mask is valid */
1786 args.done = 1; /* DONE bit */
1787 args.target = V_008DFC_SQ_EXP_NULL;
1788 args.compr = 0; /* COMPR flag (0 = 32-bit export) */
1789 args.out[0] = LLVMGetUndef(ctx->f32); /* R */
1790 args.out[1] = LLVMGetUndef(ctx->f32); /* G */
1791 args.out[2] = LLVMGetUndef(ctx->f32); /* B */
1792 args.out[3] = LLVMGetUndef(ctx->f32); /* A */
1793
1794 ac_build_export(ctx, &args);
1795 }
1796
1797 static unsigned ac_num_coords(enum ac_image_dim dim)
1798 {
1799 switch (dim) {
1800 case ac_image_1d:
1801 return 1;
1802 case ac_image_2d:
1803 case ac_image_1darray:
1804 return 2;
1805 case ac_image_3d:
1806 case ac_image_cube:
1807 case ac_image_2darray:
1808 case ac_image_2dmsaa:
1809 return 3;
1810 case ac_image_2darraymsaa:
1811 return 4;
1812 default:
1813 unreachable("ac_num_coords: bad dim");
1814 }
1815 }
1816
1817 static unsigned ac_num_derivs(enum ac_image_dim dim)
1818 {
1819 switch (dim) {
1820 case ac_image_1d:
1821 case ac_image_1darray:
1822 return 2;
1823 case ac_image_2d:
1824 case ac_image_2darray:
1825 case ac_image_cube:
1826 return 4;
1827 case ac_image_3d:
1828 return 6;
1829 case ac_image_2dmsaa:
1830 case ac_image_2darraymsaa:
1831 default:
1832 unreachable("derivatives not supported");
1833 }
1834 }
1835
1836 static const char *get_atomic_name(enum ac_atomic_op op)
1837 {
1838 switch (op) {
1839 case ac_atomic_swap: return "swap";
1840 case ac_atomic_add: return "add";
1841 case ac_atomic_sub: return "sub";
1842 case ac_atomic_smin: return "smin";
1843 case ac_atomic_umin: return "umin";
1844 case ac_atomic_smax: return "smax";
1845 case ac_atomic_umax: return "umax";
1846 case ac_atomic_and: return "and";
1847 case ac_atomic_or: return "or";
1848 case ac_atomic_xor: return "xor";
1849 }
1850 unreachable("bad atomic op");
1851 }
1852
1853 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
1854 struct ac_image_args *a)
1855 {
1856 const char *overload[3] = { "", "", "" };
1857 unsigned num_overloads = 0;
1858 LLVMValueRef args[18];
1859 unsigned num_args = 0;
1860 enum ac_image_dim dim = a->dim;
1861
1862 assert(!a->lod || a->lod == ctx->i32_0 || a->lod == ctx->f32_0 ||
1863 !a->level_zero);
1864 assert((a->opcode != ac_image_get_resinfo && a->opcode != ac_image_load_mip &&
1865 a->opcode != ac_image_store_mip) ||
1866 a->lod);
1867 assert(a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
1868 (!a->compare && !a->offset));
1869 assert((a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
1870 a->opcode == ac_image_get_lod) ||
1871 !a->bias);
1872 assert((a->bias ? 1 : 0) +
1873 (a->lod ? 1 : 0) +
1874 (a->level_zero ? 1 : 0) +
1875 (a->derivs[0] ? 1 : 0) <= 1);
1876
1877 if (a->opcode == ac_image_get_lod) {
1878 switch (dim) {
1879 case ac_image_1darray:
1880 dim = ac_image_1d;
1881 break;
1882 case ac_image_2darray:
1883 case ac_image_cube:
1884 dim = ac_image_2d;
1885 break;
1886 default:
1887 break;
1888 }
1889 }
1890
1891 bool sample = a->opcode == ac_image_sample ||
1892 a->opcode == ac_image_gather4 ||
1893 a->opcode == ac_image_get_lod;
1894 bool atomic = a->opcode == ac_image_atomic ||
1895 a->opcode == ac_image_atomic_cmpswap;
1896 LLVMTypeRef coord_type = sample ? ctx->f32 : ctx->i32;
1897
1898 if (atomic || a->opcode == ac_image_store || a->opcode == ac_image_store_mip) {
1899 args[num_args++] = a->data[0];
1900 if (a->opcode == ac_image_atomic_cmpswap)
1901 args[num_args++] = a->data[1];
1902 }
1903
1904 if (!atomic)
1905 args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, false);
1906
1907 if (a->offset)
1908 args[num_args++] = ac_to_integer(ctx, a->offset);
1909 if (a->bias) {
1910 args[num_args++] = ac_to_float(ctx, a->bias);
1911 overload[num_overloads++] = ".f32";
1912 }
1913 if (a->compare)
1914 args[num_args++] = ac_to_float(ctx, a->compare);
1915 if (a->derivs[0]) {
1916 unsigned count = ac_num_derivs(dim);
1917 for (unsigned i = 0; i < count; ++i)
1918 args[num_args++] = ac_to_float(ctx, a->derivs[i]);
1919 overload[num_overloads++] = ".f32";
1920 }
1921 unsigned num_coords =
1922 a->opcode != ac_image_get_resinfo ? ac_num_coords(dim) : 0;
1923 for (unsigned i = 0; i < num_coords; ++i)
1924 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->coords[i], coord_type, "");
1925 if (a->lod)
1926 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->lod, coord_type, "");
1927 overload[num_overloads++] = sample ? ".f32" : ".i32";
1928
1929 args[num_args++] = a->resource;
1930 if (sample) {
1931 args[num_args++] = a->sampler;
1932 args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, false);
1933 }
1934
1935 args[num_args++] = ctx->i32_0; /* texfailctrl */
1936 args[num_args++] = LLVMConstInt(ctx->i32, a->cache_policy, false);
1937
1938 const char *name;
1939 const char *atomic_subop = "";
1940 switch (a->opcode) {
1941 case ac_image_sample: name = "sample"; break;
1942 case ac_image_gather4: name = "gather4"; break;
1943 case ac_image_load: name = "load"; break;
1944 case ac_image_load_mip: name = "load.mip"; break;
1945 case ac_image_store: name = "store"; break;
1946 case ac_image_store_mip: name = "store.mip"; break;
1947 case ac_image_atomic:
1948 name = "atomic.";
1949 atomic_subop = get_atomic_name(a->atomic);
1950 break;
1951 case ac_image_atomic_cmpswap:
1952 name = "atomic.";
1953 atomic_subop = "cmpswap";
1954 break;
1955 case ac_image_get_lod: name = "getlod"; break;
1956 case ac_image_get_resinfo: name = "getresinfo"; break;
1957 default: unreachable("invalid image opcode");
1958 }
1959
1960 const char *dimname;
1961 switch (dim) {
1962 case ac_image_1d: dimname = "1d"; break;
1963 case ac_image_2d: dimname = "2d"; break;
1964 case ac_image_3d: dimname = "3d"; break;
1965 case ac_image_cube: dimname = "cube"; break;
1966 case ac_image_1darray: dimname = "1darray"; break;
1967 case ac_image_2darray: dimname = "2darray"; break;
1968 case ac_image_2dmsaa: dimname = "2dmsaa"; break;
1969 case ac_image_2darraymsaa: dimname = "2darraymsaa"; break;
1970 default: unreachable("invalid dim");
1971 }
1972
1973 bool lod_suffix =
1974 a->lod && (a->opcode == ac_image_sample || a->opcode == ac_image_gather4);
1975 char intr_name[96];
1976 snprintf(intr_name, sizeof(intr_name),
1977 "llvm.amdgcn.image.%s%s" /* base name */
1978 "%s%s%s" /* sample/gather modifiers */
1979 ".%s.%s%s%s%s", /* dimension and type overloads */
1980 name, atomic_subop,
1981 a->compare ? ".c" : "",
1982 a->bias ? ".b" :
1983 lod_suffix ? ".l" :
1984 a->derivs[0] ? ".d" :
1985 a->level_zero ? ".lz" : "",
1986 a->offset ? ".o" : "",
1987 dimname,
1988 atomic ? "i32" : "v4f32",
1989 overload[0], overload[1], overload[2]);
1990
1991 LLVMTypeRef retty;
1992 if (atomic)
1993 retty = ctx->i32;
1994 else if (a->opcode == ac_image_store || a->opcode == ac_image_store_mip)
1995 retty = ctx->voidt;
1996 else
1997 retty = ctx->v4f32;
1998
1999 LLVMValueRef result =
2000 ac_build_intrinsic(ctx, intr_name, retty, args, num_args,
2001 a->attributes);
2002 if (!sample && retty == ctx->v4f32) {
2003 result = LLVMBuildBitCast(ctx->builder, result,
2004 ctx->v4i32, "");
2005 }
2006 return result;
2007 }
2008
2009 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
2010 LLVMValueRef args[2])
2011 {
2012 LLVMTypeRef v2f16 =
2013 LLVMVectorType(LLVMHalfTypeInContext(ctx->context), 2);
2014
2015 return ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz", v2f16,
2016 args, 2, AC_FUNC_ATTR_READNONE);
2017 }
2018
2019 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
2020 LLVMValueRef args[2])
2021 {
2022 LLVMValueRef res =
2023 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16",
2024 ctx->v2i16, args, 2,
2025 AC_FUNC_ATTR_READNONE);
2026 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2027 }
2028
2029 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
2030 LLVMValueRef args[2])
2031 {
2032 LLVMValueRef res =
2033 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16",
2034 ctx->v2i16, args, 2,
2035 AC_FUNC_ATTR_READNONE);
2036 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2037 }
2038
2039 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2040 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
2041 LLVMValueRef args[2], unsigned bits, bool hi)
2042 {
2043 assert(bits == 8 || bits == 10 || bits == 16);
2044
2045 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2046 bits == 8 ? 127 : bits == 10 ? 511 : 32767, 0);
2047 LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
2048 bits == 8 ? -128 : bits == 10 ? -512 : -32768, 0);
2049 LLVMValueRef max_alpha =
2050 bits != 10 ? max_rgb : ctx->i32_1;
2051 LLVMValueRef min_alpha =
2052 bits != 10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
2053
2054 /* Clamp. */
2055 if (bits != 16) {
2056 for (int i = 0; i < 2; i++) {
2057 bool alpha = hi && i == 1;
2058 args[i] = ac_build_imin(ctx, args[i],
2059 alpha ? max_alpha : max_rgb);
2060 args[i] = ac_build_imax(ctx, args[i],
2061 alpha ? min_alpha : min_rgb);
2062 }
2063 }
2064
2065 LLVMValueRef res =
2066 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16",
2067 ctx->v2i16, args, 2,
2068 AC_FUNC_ATTR_READNONE);
2069 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2070 }
2071
2072 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2073 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
2074 LLVMValueRef args[2], unsigned bits, bool hi)
2075 {
2076 assert(bits == 8 || bits == 10 || bits == 16);
2077
2078 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2079 bits == 8 ? 255 : bits == 10 ? 1023 : 65535, 0);
2080 LLVMValueRef max_alpha =
2081 bits != 10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
2082
2083 /* Clamp. */
2084 if (bits != 16) {
2085 for (int i = 0; i < 2; i++) {
2086 bool alpha = hi && i == 1;
2087 args[i] = ac_build_umin(ctx, args[i],
2088 alpha ? max_alpha : max_rgb);
2089 }
2090 }
2091
2092 LLVMValueRef res =
2093 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16",
2094 ctx->v2i16, args, 2,
2095 AC_FUNC_ATTR_READNONE);
2096 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2097 }
2098
2099 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1)
2100 {
2101 return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1,
2102 &i1, 1, AC_FUNC_ATTR_READNONE);
2103 }
2104
2105 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1)
2106 {
2107 ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt,
2108 &i1, 1, 0);
2109 }
2110
2111 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
2112 LLVMValueRef offset, LLVMValueRef width,
2113 bool is_signed)
2114 {
2115 LLVMValueRef args[] = {
2116 input,
2117 offset,
2118 width,
2119 };
2120
2121 return ac_build_intrinsic(ctx,
2122 is_signed ? "llvm.amdgcn.sbfe.i32" :
2123 "llvm.amdgcn.ubfe.i32",
2124 ctx->i32, args, 3,
2125 AC_FUNC_ATTR_READNONE);
2126 }
2127
2128 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2129 LLVMValueRef s1, LLVMValueRef s2)
2130 {
2131 return LLVMBuildAdd(ctx->builder,
2132 LLVMBuildMul(ctx->builder, s0, s1, ""), s2, "");
2133 }
2134
2135 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2136 LLVMValueRef s1, LLVMValueRef s2)
2137 {
2138 return LLVMBuildFAdd(ctx->builder,
2139 LLVMBuildFMul(ctx->builder, s0, s1, ""), s2, "");
2140 }
2141
2142 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16)
2143 {
2144 LLVMValueRef args[1] = {
2145 LLVMConstInt(ctx->i32, simm16, false),
2146 };
2147 ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt",
2148 ctx->voidt, args, 1, 0);
2149 }
2150
2151 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
2152 unsigned bitsize)
2153 {
2154 LLVMTypeRef type;
2155 char *intr;
2156
2157 if (bitsize == 32) {
2158 intr = "llvm.floor.f32";
2159 type = ctx->f32;
2160 } else {
2161 intr = "llvm.floor.f64";
2162 type = ctx->f64;
2163 }
2164
2165 LLVMValueRef params[] = {
2166 src0,
2167 };
2168 LLVMValueRef floor = ac_build_intrinsic(ctx, intr, type, params, 1,
2169 AC_FUNC_ATTR_READNONE);
2170 return LLVMBuildFSub(ctx->builder, src0, floor, "");
2171 }
2172
2173 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2174 unsigned bitsize)
2175 {
2176 LLVMTypeRef type = LLVMIntTypeInContext(ctx->context, bitsize);
2177 LLVMValueRef zero = LLVMConstInt(type, 0, false);
2178 LLVMValueRef one = LLVMConstInt(type, 1, false);
2179
2180 LLVMValueRef cmp, val;
2181 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, src0, zero, "");
2182 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2183 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGE, val, zero, "");
2184 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstInt(type, -1, true), "");
2185 return val;
2186 }
2187
2188 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2189 unsigned bitsize)
2190 {
2191 LLVMValueRef cmp, val, zero, one;
2192 LLVMTypeRef type;
2193
2194 if (bitsize == 32) {
2195 type = ctx->f32;
2196 zero = ctx->f32_0;
2197 one = ctx->f32_1;
2198 } else {
2199 type = ctx->f64;
2200 zero = ctx->f64_0;
2201 one = ctx->f64_1;
2202 }
2203
2204 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src0, zero, "");
2205 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2206 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGE, val, zero, "");
2207 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstReal(type, -1.0), "");
2208 return val;
2209 }
2210
2211 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0)
2212 {
2213 LLVMValueRef result;
2214 unsigned bitsize;
2215
2216 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2217
2218 switch (bitsize) {
2219 case 64:
2220 result = ac_build_intrinsic(ctx, "llvm.ctpop.i64", ctx->i64,
2221 (LLVMValueRef []) { src0 }, 1,
2222 AC_FUNC_ATTR_READNONE);
2223
2224 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
2225 break;
2226 case 32:
2227 result = ac_build_intrinsic(ctx, "llvm.ctpop.i32", ctx->i32,
2228 (LLVMValueRef []) { src0 }, 1,
2229 AC_FUNC_ATTR_READNONE);
2230 break;
2231 case 16:
2232 result = ac_build_intrinsic(ctx, "llvm.ctpop.i16", ctx->i16,
2233 (LLVMValueRef []) { src0 }, 1,
2234 AC_FUNC_ATTR_READNONE);
2235 break;
2236 default:
2237 unreachable(!"invalid bitsize");
2238 break;
2239 }
2240
2241 return result;
2242 }
2243
2244 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
2245 LLVMValueRef src0)
2246 {
2247 LLVMValueRef result;
2248 unsigned bitsize;
2249
2250 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2251
2252 switch (bitsize) {
2253 case 32:
2254 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i32", ctx->i32,
2255 (LLVMValueRef []) { src0 }, 1,
2256 AC_FUNC_ATTR_READNONE);
2257 break;
2258 case 16:
2259 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i16", ctx->i16,
2260 (LLVMValueRef []) { src0 }, 1,
2261 AC_FUNC_ATTR_READNONE);
2262 break;
2263 default:
2264 unreachable(!"invalid bitsize");
2265 break;
2266 }
2267
2268 return result;
2269 }
2270
2271 #define AC_EXP_TARGET 0
2272 #define AC_EXP_ENABLED_CHANNELS 1
2273 #define AC_EXP_OUT0 2
2274
2275 enum ac_ir_type {
2276 AC_IR_UNDEF,
2277 AC_IR_CONST,
2278 AC_IR_VALUE,
2279 };
2280
2281 struct ac_vs_exp_chan
2282 {
2283 LLVMValueRef value;
2284 float const_float;
2285 enum ac_ir_type type;
2286 };
2287
2288 struct ac_vs_exp_inst {
2289 unsigned offset;
2290 LLVMValueRef inst;
2291 struct ac_vs_exp_chan chan[4];
2292 };
2293
2294 struct ac_vs_exports {
2295 unsigned num;
2296 struct ac_vs_exp_inst exp[VARYING_SLOT_MAX];
2297 };
2298
2299 /* Return true if the PARAM export has been eliminated. */
2300 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset,
2301 uint32_t num_outputs,
2302 struct ac_vs_exp_inst *exp)
2303 {
2304 unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
2305 bool is_zero[4] = {}, is_one[4] = {};
2306
2307 for (i = 0; i < 4; i++) {
2308 /* It's a constant expression. Undef outputs are eliminated too. */
2309 if (exp->chan[i].type == AC_IR_UNDEF) {
2310 is_zero[i] = true;
2311 is_one[i] = true;
2312 } else if (exp->chan[i].type == AC_IR_CONST) {
2313 if (exp->chan[i].const_float == 0)
2314 is_zero[i] = true;
2315 else if (exp->chan[i].const_float == 1)
2316 is_one[i] = true;
2317 else
2318 return false; /* other constant */
2319 } else
2320 return false;
2321 }
2322
2323 /* Only certain combinations of 0 and 1 can be eliminated. */
2324 if (is_zero[0] && is_zero[1] && is_zero[2])
2325 default_val = is_zero[3] ? 0 : 1;
2326 else if (is_one[0] && is_one[1] && is_one[2])
2327 default_val = is_zero[3] ? 2 : 3;
2328 else
2329 return false;
2330
2331 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
2332 LLVMInstructionEraseFromParent(exp->inst);
2333
2334 /* Change OFFSET to DEFAULT_VAL. */
2335 for (i = 0; i < num_outputs; i++) {
2336 if (vs_output_param_offset[i] == exp->offset) {
2337 vs_output_param_offset[i] =
2338 AC_EXP_PARAM_DEFAULT_VAL_0000 + default_val;
2339 break;
2340 }
2341 }
2342 return true;
2343 }
2344
2345 static bool ac_eliminate_duplicated_output(struct ac_llvm_context *ctx,
2346 uint8_t *vs_output_param_offset,
2347 uint32_t num_outputs,
2348 struct ac_vs_exports *processed,
2349 struct ac_vs_exp_inst *exp)
2350 {
2351 unsigned p, copy_back_channels = 0;
2352
2353 /* See if the output is already in the list of processed outputs.
2354 * The LLVMValueRef comparison relies on SSA.
2355 */
2356 for (p = 0; p < processed->num; p++) {
2357 bool different = false;
2358
2359 for (unsigned j = 0; j < 4; j++) {
2360 struct ac_vs_exp_chan *c1 = &processed->exp[p].chan[j];
2361 struct ac_vs_exp_chan *c2 = &exp->chan[j];
2362
2363 /* Treat undef as a match. */
2364 if (c2->type == AC_IR_UNDEF)
2365 continue;
2366
2367 /* If c1 is undef but c2 isn't, we can copy c2 to c1
2368 * and consider the instruction duplicated.
2369 */
2370 if (c1->type == AC_IR_UNDEF) {
2371 copy_back_channels |= 1 << j;
2372 continue;
2373 }
2374
2375 /* Test whether the channels are not equal. */
2376 if (c1->type != c2->type ||
2377 (c1->type == AC_IR_CONST &&
2378 c1->const_float != c2->const_float) ||
2379 (c1->type == AC_IR_VALUE &&
2380 c1->value != c2->value)) {
2381 different = true;
2382 break;
2383 }
2384 }
2385 if (!different)
2386 break;
2387
2388 copy_back_channels = 0;
2389 }
2390 if (p == processed->num)
2391 return false;
2392
2393 /* If a match was found, but the matching export has undef where the new
2394 * one has a normal value, copy the normal value to the undef channel.
2395 */
2396 struct ac_vs_exp_inst *match = &processed->exp[p];
2397
2398 /* Get current enabled channels mask. */
2399 LLVMValueRef arg = LLVMGetOperand(match->inst, AC_EXP_ENABLED_CHANNELS);
2400 unsigned enabled_channels = LLVMConstIntGetZExtValue(arg);
2401
2402 while (copy_back_channels) {
2403 unsigned chan = u_bit_scan(&copy_back_channels);
2404
2405 assert(match->chan[chan].type == AC_IR_UNDEF);
2406 LLVMSetOperand(match->inst, AC_EXP_OUT0 + chan,
2407 exp->chan[chan].value);
2408 match->chan[chan] = exp->chan[chan];
2409
2410 /* Update number of enabled channels because the original mask
2411 * is not always 0xf.
2412 */
2413 enabled_channels |= (1 << chan);
2414 LLVMSetOperand(match->inst, AC_EXP_ENABLED_CHANNELS,
2415 LLVMConstInt(ctx->i32, enabled_channels, 0));
2416 }
2417
2418 /* The PARAM export is duplicated. Kill it. */
2419 LLVMInstructionEraseFromParent(exp->inst);
2420
2421 /* Change OFFSET to the matching export. */
2422 for (unsigned i = 0; i < num_outputs; i++) {
2423 if (vs_output_param_offset[i] == exp->offset) {
2424 vs_output_param_offset[i] = match->offset;
2425 break;
2426 }
2427 }
2428 return true;
2429 }
2430
2431 void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
2432 LLVMValueRef main_fn,
2433 uint8_t *vs_output_param_offset,
2434 uint32_t num_outputs,
2435 uint8_t *num_param_exports)
2436 {
2437 LLVMBasicBlockRef bb;
2438 bool removed_any = false;
2439 struct ac_vs_exports exports;
2440
2441 exports.num = 0;
2442
2443 /* Process all LLVM instructions. */
2444 bb = LLVMGetFirstBasicBlock(main_fn);
2445 while (bb) {
2446 LLVMValueRef inst = LLVMGetFirstInstruction(bb);
2447
2448 while (inst) {
2449 LLVMValueRef cur = inst;
2450 inst = LLVMGetNextInstruction(inst);
2451 struct ac_vs_exp_inst exp;
2452
2453 if (LLVMGetInstructionOpcode(cur) != LLVMCall)
2454 continue;
2455
2456 LLVMValueRef callee = ac_llvm_get_called_value(cur);
2457
2458 if (!ac_llvm_is_function(callee))
2459 continue;
2460
2461 const char *name = LLVMGetValueName(callee);
2462 unsigned num_args = LLVMCountParams(callee);
2463
2464 /* Check if this is an export instruction. */
2465 if ((num_args != 9 && num_args != 8) ||
2466 (strcmp(name, "llvm.SI.export") &&
2467 strcmp(name, "llvm.amdgcn.exp.f32")))
2468 continue;
2469
2470 LLVMValueRef arg = LLVMGetOperand(cur, AC_EXP_TARGET);
2471 unsigned target = LLVMConstIntGetZExtValue(arg);
2472
2473 if (target < V_008DFC_SQ_EXP_PARAM)
2474 continue;
2475
2476 target -= V_008DFC_SQ_EXP_PARAM;
2477
2478 /* Parse the instruction. */
2479 memset(&exp, 0, sizeof(exp));
2480 exp.offset = target;
2481 exp.inst = cur;
2482
2483 for (unsigned i = 0; i < 4; i++) {
2484 LLVMValueRef v = LLVMGetOperand(cur, AC_EXP_OUT0 + i);
2485
2486 exp.chan[i].value = v;
2487
2488 if (LLVMIsUndef(v)) {
2489 exp.chan[i].type = AC_IR_UNDEF;
2490 } else if (LLVMIsAConstantFP(v)) {
2491 LLVMBool loses_info;
2492 exp.chan[i].type = AC_IR_CONST;
2493 exp.chan[i].const_float =
2494 LLVMConstRealGetDouble(v, &loses_info);
2495 } else {
2496 exp.chan[i].type = AC_IR_VALUE;
2497 }
2498 }
2499
2500 /* Eliminate constant and duplicated PARAM exports. */
2501 if (ac_eliminate_const_output(vs_output_param_offset,
2502 num_outputs, &exp) ||
2503 ac_eliminate_duplicated_output(ctx,
2504 vs_output_param_offset,
2505 num_outputs, &exports,
2506 &exp)) {
2507 removed_any = true;
2508 } else {
2509 exports.exp[exports.num++] = exp;
2510 }
2511 }
2512 bb = LLVMGetNextBasicBlock(bb);
2513 }
2514
2515 /* Remove holes in export memory due to removed PARAM exports.
2516 * This is done by renumbering all PARAM exports.
2517 */
2518 if (removed_any) {
2519 uint8_t old_offset[VARYING_SLOT_MAX];
2520 unsigned out, i;
2521
2522 /* Make a copy of the offsets. We need the old version while
2523 * we are modifying some of them. */
2524 memcpy(old_offset, vs_output_param_offset,
2525 sizeof(old_offset));
2526
2527 for (i = 0; i < exports.num; i++) {
2528 unsigned offset = exports.exp[i].offset;
2529
2530 /* Update vs_output_param_offset. Multiple outputs can
2531 * have the same offset.
2532 */
2533 for (out = 0; out < num_outputs; out++) {
2534 if (old_offset[out] == offset)
2535 vs_output_param_offset[out] = i;
2536 }
2537
2538 /* Change the PARAM offset in the instruction. */
2539 LLVMSetOperand(exports.exp[i].inst, AC_EXP_TARGET,
2540 LLVMConstInt(ctx->i32,
2541 V_008DFC_SQ_EXP_PARAM + i, 0));
2542 }
2543 *num_param_exports = exports.num;
2544 }
2545 }
2546
2547 void ac_init_exec_full_mask(struct ac_llvm_context *ctx)
2548 {
2549 LLVMValueRef full_mask = LLVMConstInt(ctx->i64, ~0ull, 0);
2550 ac_build_intrinsic(ctx,
2551 "llvm.amdgcn.init.exec", ctx->voidt,
2552 &full_mask, 1, AC_FUNC_ATTR_CONVERGENT);
2553 }
2554
2555 void ac_declare_lds_as_pointer(struct ac_llvm_context *ctx)
2556 {
2557 unsigned lds_size = ctx->chip_class >= CIK ? 65536 : 32768;
2558 ctx->lds = LLVMBuildIntToPtr(ctx->builder, ctx->i32_0,
2559 LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), AC_ADDR_SPACE_LDS),
2560 "lds");
2561 }
2562
2563 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
2564 LLVMValueRef dw_addr)
2565 {
2566 return ac_build_load(ctx, ctx->lds, dw_addr);
2567 }
2568
2569 void ac_lds_store(struct ac_llvm_context *ctx,
2570 LLVMValueRef dw_addr,
2571 LLVMValueRef value)
2572 {
2573 value = ac_to_integer(ctx, value);
2574 ac_build_indexed_store(ctx, ctx->lds,
2575 dw_addr, value);
2576 }
2577
2578 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
2579 LLVMTypeRef dst_type,
2580 LLVMValueRef src0)
2581 {
2582 unsigned src0_bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2583 const char *intrin_name;
2584 LLVMTypeRef type;
2585 LLVMValueRef zero;
2586
2587 switch (src0_bitsize) {
2588 case 64:
2589 intrin_name = "llvm.cttz.i64";
2590 type = ctx->i64;
2591 zero = ctx->i64_0;
2592 break;
2593 case 32:
2594 intrin_name = "llvm.cttz.i32";
2595 type = ctx->i32;
2596 zero = ctx->i32_0;
2597 break;
2598 case 16:
2599 intrin_name = "llvm.cttz.i16";
2600 type = ctx->i16;
2601 zero = ctx->i16_0;
2602 break;
2603 default:
2604 unreachable(!"invalid bitsize");
2605 }
2606
2607 LLVMValueRef params[2] = {
2608 src0,
2609
2610 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
2611 * add special code to check for x=0. The reason is that
2612 * the LLVM behavior for x=0 is different from what we
2613 * need here. However, LLVM also assumes that ffs(x) is
2614 * in [0, 31], but GLSL expects that ffs(0) = -1, so
2615 * a conditional assignment to handle 0 is still required.
2616 *
2617 * The hardware already implements the correct behavior.
2618 */
2619 ctx->i1true,
2620 };
2621
2622 LLVMValueRef lsb = ac_build_intrinsic(ctx, intrin_name, type,
2623 params, 2,
2624 AC_FUNC_ATTR_READNONE);
2625
2626 if (src0_bitsize == 64) {
2627 lsb = LLVMBuildTrunc(ctx->builder, lsb, ctx->i32, "");
2628 }
2629
2630 /* TODO: We need an intrinsic to skip this conditional. */
2631 /* Check for zero: */
2632 return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder,
2633 LLVMIntEQ, src0,
2634 zero, ""),
2635 LLVMConstInt(ctx->i32, -1, 0), lsb, "");
2636 }
2637
2638 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type)
2639 {
2640 return LLVMPointerType(LLVMArrayType(elem_type, 0),
2641 AC_ADDR_SPACE_CONST);
2642 }
2643
2644 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type)
2645 {
2646 return LLVMPointerType(LLVMArrayType(elem_type, 0),
2647 AC_ADDR_SPACE_CONST_32BIT);
2648 }
2649
2650 static struct ac_llvm_flow *
2651 get_current_flow(struct ac_llvm_context *ctx)
2652 {
2653 if (ctx->flow_depth > 0)
2654 return &ctx->flow[ctx->flow_depth - 1];
2655 return NULL;
2656 }
2657
2658 static struct ac_llvm_flow *
2659 get_innermost_loop(struct ac_llvm_context *ctx)
2660 {
2661 for (unsigned i = ctx->flow_depth; i > 0; --i) {
2662 if (ctx->flow[i - 1].loop_entry_block)
2663 return &ctx->flow[i - 1];
2664 }
2665 return NULL;
2666 }
2667
2668 static struct ac_llvm_flow *
2669 push_flow(struct ac_llvm_context *ctx)
2670 {
2671 struct ac_llvm_flow *flow;
2672
2673 if (ctx->flow_depth >= ctx->flow_depth_max) {
2674 unsigned new_max = MAX2(ctx->flow_depth << 1,
2675 AC_LLVM_INITIAL_CF_DEPTH);
2676
2677 ctx->flow = realloc(ctx->flow, new_max * sizeof(*ctx->flow));
2678 ctx->flow_depth_max = new_max;
2679 }
2680
2681 flow = &ctx->flow[ctx->flow_depth];
2682 ctx->flow_depth++;
2683
2684 flow->next_block = NULL;
2685 flow->loop_entry_block = NULL;
2686 return flow;
2687 }
2688
2689 static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base,
2690 int label_id)
2691 {
2692 char buf[32];
2693 snprintf(buf, sizeof(buf), "%s%d", base, label_id);
2694 LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
2695 }
2696
2697 /* Append a basic block at the level of the parent flow.
2698 */
2699 static LLVMBasicBlockRef append_basic_block(struct ac_llvm_context *ctx,
2700 const char *name)
2701 {
2702 assert(ctx->flow_depth >= 1);
2703
2704 if (ctx->flow_depth >= 2) {
2705 struct ac_llvm_flow *flow = &ctx->flow[ctx->flow_depth - 2];
2706
2707 return LLVMInsertBasicBlockInContext(ctx->context,
2708 flow->next_block, name);
2709 }
2710
2711 LLVMValueRef main_fn =
2712 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->builder));
2713 return LLVMAppendBasicBlockInContext(ctx->context, main_fn, name);
2714 }
2715
2716 /* Emit a branch to the given default target for the current block if
2717 * applicable -- that is, if the current block does not already contain a
2718 * branch from a break or continue.
2719 */
2720 static void emit_default_branch(LLVMBuilderRef builder,
2721 LLVMBasicBlockRef target)
2722 {
2723 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
2724 LLVMBuildBr(builder, target);
2725 }
2726
2727 void ac_build_bgnloop(struct ac_llvm_context *ctx, int label_id)
2728 {
2729 struct ac_llvm_flow *flow = push_flow(ctx);
2730 flow->loop_entry_block = append_basic_block(ctx, "LOOP");
2731 flow->next_block = append_basic_block(ctx, "ENDLOOP");
2732 set_basicblock_name(flow->loop_entry_block, "loop", label_id);
2733 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
2734 LLVMPositionBuilderAtEnd(ctx->builder, flow->loop_entry_block);
2735 }
2736
2737 void ac_build_break(struct ac_llvm_context *ctx)
2738 {
2739 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
2740 LLVMBuildBr(ctx->builder, flow->next_block);
2741 }
2742
2743 void ac_build_continue(struct ac_llvm_context *ctx)
2744 {
2745 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
2746 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
2747 }
2748
2749 void ac_build_else(struct ac_llvm_context *ctx, int label_id)
2750 {
2751 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
2752 LLVMBasicBlockRef endif_block;
2753
2754 assert(!current_branch->loop_entry_block);
2755
2756 endif_block = append_basic_block(ctx, "ENDIF");
2757 emit_default_branch(ctx->builder, endif_block);
2758
2759 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
2760 set_basicblock_name(current_branch->next_block, "else", label_id);
2761
2762 current_branch->next_block = endif_block;
2763 }
2764
2765 void ac_build_endif(struct ac_llvm_context *ctx, int label_id)
2766 {
2767 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
2768
2769 assert(!current_branch->loop_entry_block);
2770
2771 emit_default_branch(ctx->builder, current_branch->next_block);
2772 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
2773 set_basicblock_name(current_branch->next_block, "endif", label_id);
2774
2775 ctx->flow_depth--;
2776 }
2777
2778 void ac_build_endloop(struct ac_llvm_context *ctx, int label_id)
2779 {
2780 struct ac_llvm_flow *current_loop = get_current_flow(ctx);
2781
2782 assert(current_loop->loop_entry_block);
2783
2784 emit_default_branch(ctx->builder, current_loop->loop_entry_block);
2785
2786 LLVMPositionBuilderAtEnd(ctx->builder, current_loop->next_block);
2787 set_basicblock_name(current_loop->next_block, "endloop", label_id);
2788 ctx->flow_depth--;
2789 }
2790
2791 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id)
2792 {
2793 struct ac_llvm_flow *flow = push_flow(ctx);
2794 LLVMBasicBlockRef if_block;
2795
2796 if_block = append_basic_block(ctx, "IF");
2797 flow->next_block = append_basic_block(ctx, "ELSE");
2798 set_basicblock_name(if_block, "if", label_id);
2799 LLVMBuildCondBr(ctx->builder, cond, if_block, flow->next_block);
2800 LLVMPositionBuilderAtEnd(ctx->builder, if_block);
2801 }
2802
2803 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
2804 int label_id)
2805 {
2806 LLVMValueRef cond = LLVMBuildFCmp(ctx->builder, LLVMRealUNE,
2807 value, ctx->f32_0, "");
2808 ac_build_ifcc(ctx, cond, label_id);
2809 }
2810
2811 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
2812 int label_id)
2813 {
2814 LLVMValueRef cond = LLVMBuildICmp(ctx->builder, LLVMIntNE,
2815 ac_to_integer(ctx, value),
2816 ctx->i32_0, "");
2817 ac_build_ifcc(ctx, cond, label_id);
2818 }
2819
2820 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
2821 const char *name)
2822 {
2823 LLVMBuilderRef builder = ac->builder;
2824 LLVMBasicBlockRef current_block = LLVMGetInsertBlock(builder);
2825 LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
2826 LLVMBasicBlockRef first_block = LLVMGetEntryBasicBlock(function);
2827 LLVMValueRef first_instr = LLVMGetFirstInstruction(first_block);
2828 LLVMBuilderRef first_builder = LLVMCreateBuilderInContext(ac->context);
2829 LLVMValueRef res;
2830
2831 if (first_instr) {
2832 LLVMPositionBuilderBefore(first_builder, first_instr);
2833 } else {
2834 LLVMPositionBuilderAtEnd(first_builder, first_block);
2835 }
2836
2837 res = LLVMBuildAlloca(first_builder, type, name);
2838 LLVMDisposeBuilder(first_builder);
2839 return res;
2840 }
2841
2842 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac,
2843 LLVMTypeRef type, const char *name)
2844 {
2845 LLVMValueRef ptr = ac_build_alloca_undef(ac, type, name);
2846 LLVMBuildStore(ac->builder, LLVMConstNull(type), ptr);
2847 return ptr;
2848 }
2849
2850 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
2851 LLVMTypeRef type)
2852 {
2853 int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
2854 return LLVMBuildBitCast(ctx->builder, ptr,
2855 LLVMPointerType(type, addr_space), "");
2856 }
2857
2858 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
2859 unsigned count)
2860 {
2861 unsigned num_components = ac_get_llvm_num_components(value);
2862 if (count == num_components)
2863 return value;
2864
2865 LLVMValueRef masks[MAX2(count, 2)];
2866 masks[0] = ctx->i32_0;
2867 masks[1] = ctx->i32_1;
2868 for (unsigned i = 2; i < count; i++)
2869 masks[i] = LLVMConstInt(ctx->i32, i, false);
2870
2871 if (count == 1)
2872 return LLVMBuildExtractElement(ctx->builder, value, masks[0],
2873 "");
2874
2875 LLVMValueRef swizzle = LLVMConstVector(masks, count);
2876 return LLVMBuildShuffleVector(ctx->builder, value, value, swizzle, "");
2877 }
2878
2879 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
2880 unsigned rshift, unsigned bitwidth)
2881 {
2882 LLVMValueRef value = param;
2883 if (rshift)
2884 value = LLVMBuildLShr(ctx->builder, value,
2885 LLVMConstInt(ctx->i32, rshift, false), "");
2886
2887 if (rshift + bitwidth < 32) {
2888 unsigned mask = (1 << bitwidth) - 1;
2889 value = LLVMBuildAnd(ctx->builder, value,
2890 LLVMConstInt(ctx->i32, mask, false), "");
2891 }
2892 return value;
2893 }
2894
2895 /* Adjust the sample index according to FMASK.
2896 *
2897 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
2898 * which is the identity mapping. Each nibble says which physical sample
2899 * should be fetched to get that sample.
2900 *
2901 * For example, 0x11111100 means there are only 2 samples stored and
2902 * the second sample covers 3/4 of the pixel. When reading samples 0
2903 * and 1, return physical sample 0 (determined by the first two 0s
2904 * in FMASK), otherwise return physical sample 1.
2905 *
2906 * The sample index should be adjusted as follows:
2907 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
2908 */
2909 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
2910 LLVMValueRef *addr, bool is_array_tex)
2911 {
2912 struct ac_image_args fmask_load = {};
2913 fmask_load.opcode = ac_image_load;
2914 fmask_load.resource = fmask;
2915 fmask_load.dmask = 0xf;
2916 fmask_load.dim = is_array_tex ? ac_image_2darray : ac_image_2d;
2917
2918 fmask_load.coords[0] = addr[0];
2919 fmask_load.coords[1] = addr[1];
2920 if (is_array_tex)
2921 fmask_load.coords[2] = addr[2];
2922
2923 LLVMValueRef fmask_value = ac_build_image_opcode(ac, &fmask_load);
2924 fmask_value = LLVMBuildExtractElement(ac->builder, fmask_value,
2925 ac->i32_0, "");
2926
2927 /* Apply the formula. */
2928 unsigned sample_chan = is_array_tex ? 3 : 2;
2929 LLVMValueRef final_sample;
2930 final_sample = LLVMBuildMul(ac->builder, addr[sample_chan],
2931 LLVMConstInt(ac->i32, 4, 0), "");
2932 final_sample = LLVMBuildLShr(ac->builder, fmask_value, final_sample, "");
2933 /* Mask the sample index by 0x7, because 0x8 means an unknown value
2934 * with EQAA, so those will map to 0. */
2935 final_sample = LLVMBuildAnd(ac->builder, final_sample,
2936 LLVMConstInt(ac->i32, 0x7, 0), "");
2937
2938 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
2939 * resource descriptor is 0 (invalid).
2940 */
2941 LLVMValueRef tmp;
2942 tmp = LLVMBuildBitCast(ac->builder, fmask, ac->v8i32, "");
2943 tmp = LLVMBuildExtractElement(ac->builder, tmp, ac->i32_1, "");
2944 tmp = LLVMBuildICmp(ac->builder, LLVMIntNE, tmp, ac->i32_0, "");
2945
2946 /* Replace the MSAA sample index. */
2947 addr[sample_chan] = LLVMBuildSelect(ac->builder, tmp, final_sample,
2948 addr[sample_chan], "");
2949 }
2950
2951 static LLVMValueRef
2952 _ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
2953 {
2954 ac_build_optimization_barrier(ctx, &src);
2955 return ac_build_intrinsic(ctx,
2956 lane == NULL ? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
2957 LLVMTypeOf(src), (LLVMValueRef []) {
2958 src, lane },
2959 lane == NULL ? 1 : 2,
2960 AC_FUNC_ATTR_READNONE |
2961 AC_FUNC_ATTR_CONVERGENT);
2962 }
2963
2964 /**
2965 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
2966 * @param ctx
2967 * @param src
2968 * @param lane - id of the lane or NULL for the first active lane
2969 * @return value of the lane
2970 */
2971 LLVMValueRef
2972 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
2973 {
2974 LLVMTypeRef src_type = LLVMTypeOf(src);
2975 src = ac_to_integer(ctx, src);
2976 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
2977 LLVMValueRef ret;
2978
2979 if (bits == 32) {
2980 ret = _ac_build_readlane(ctx, src, lane);
2981 } else {
2982 assert(bits % 32 == 0);
2983 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
2984 LLVMValueRef src_vector =
2985 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
2986 ret = LLVMGetUndef(vec_type);
2987 for (unsigned i = 0; i < bits / 32; i++) {
2988 src = LLVMBuildExtractElement(ctx->builder, src_vector,
2989 LLVMConstInt(ctx->i32, i, 0), "");
2990 LLVMValueRef ret_comp = _ac_build_readlane(ctx, src, lane);
2991 ret = LLVMBuildInsertElement(ctx->builder, ret, ret_comp,
2992 LLVMConstInt(ctx->i32, i, 0), "");
2993 }
2994 }
2995 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
2996 }
2997
2998 LLVMValueRef
2999 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane)
3000 {
3001 /* TODO: Use the actual instruction when LLVM adds an intrinsic for it.
3002 */
3003 LLVMValueRef pred = LLVMBuildICmp(ctx->builder, LLVMIntEQ, lane,
3004 ac_get_thread_id(ctx), "");
3005 return LLVMBuildSelect(ctx->builder, pred, value, src, "");
3006 }
3007
3008 LLVMValueRef
3009 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask)
3010 {
3011 LLVMValueRef mask_vec = LLVMBuildBitCast(ctx->builder, mask,
3012 LLVMVectorType(ctx->i32, 2),
3013 "");
3014 LLVMValueRef mask_lo = LLVMBuildExtractElement(ctx->builder, mask_vec,
3015 ctx->i32_0, "");
3016 LLVMValueRef mask_hi = LLVMBuildExtractElement(ctx->builder, mask_vec,
3017 ctx->i32_1, "");
3018 LLVMValueRef val =
3019 ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
3020 (LLVMValueRef []) { mask_lo, ctx->i32_0 },
3021 2, AC_FUNC_ATTR_READNONE);
3022 val = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi", ctx->i32,
3023 (LLVMValueRef []) { mask_hi, val },
3024 2, AC_FUNC_ATTR_READNONE);
3025 return val;
3026 }
3027
3028 enum dpp_ctrl {
3029 _dpp_quad_perm = 0x000,
3030 _dpp_row_sl = 0x100,
3031 _dpp_row_sr = 0x110,
3032 _dpp_row_rr = 0x120,
3033 dpp_wf_sl1 = 0x130,
3034 dpp_wf_rl1 = 0x134,
3035 dpp_wf_sr1 = 0x138,
3036 dpp_wf_rr1 = 0x13C,
3037 dpp_row_mirror = 0x140,
3038 dpp_row_half_mirror = 0x141,
3039 dpp_row_bcast15 = 0x142,
3040 dpp_row_bcast31 = 0x143
3041 };
3042
3043 static inline enum dpp_ctrl
3044 dpp_quad_perm(unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
3045 {
3046 assert(lane0 < 4 && lane1 < 4 && lane2 < 4 && lane3 < 4);
3047 return _dpp_quad_perm | lane0 | (lane1 << 2) | (lane2 << 4) | (lane3 << 6);
3048 }
3049
3050 static inline enum dpp_ctrl
3051 dpp_row_sl(unsigned amount)
3052 {
3053 assert(amount > 0 && amount < 16);
3054 return _dpp_row_sl | amount;
3055 }
3056
3057 static inline enum dpp_ctrl
3058 dpp_row_sr(unsigned amount)
3059 {
3060 assert(amount > 0 && amount < 16);
3061 return _dpp_row_sr | amount;
3062 }
3063
3064 static LLVMValueRef
3065 _ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3066 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3067 bool bound_ctrl)
3068 {
3069 return ac_build_intrinsic(ctx, "llvm.amdgcn.update.dpp.i32",
3070 LLVMTypeOf(old),
3071 (LLVMValueRef[]) {
3072 old, src,
3073 LLVMConstInt(ctx->i32, dpp_ctrl, 0),
3074 LLVMConstInt(ctx->i32, row_mask, 0),
3075 LLVMConstInt(ctx->i32, bank_mask, 0),
3076 LLVMConstInt(ctx->i1, bound_ctrl, 0) },
3077 6, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3078 }
3079
3080 static LLVMValueRef
3081 ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3082 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3083 bool bound_ctrl)
3084 {
3085 LLVMTypeRef src_type = LLVMTypeOf(src);
3086 src = ac_to_integer(ctx, src);
3087 old = ac_to_integer(ctx, old);
3088 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3089 LLVMValueRef ret;
3090 if (bits == 32) {
3091 ret = _ac_build_dpp(ctx, old, src, dpp_ctrl, row_mask,
3092 bank_mask, bound_ctrl);
3093 } else {
3094 assert(bits % 32 == 0);
3095 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3096 LLVMValueRef src_vector =
3097 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3098 LLVMValueRef old_vector =
3099 LLVMBuildBitCast(ctx->builder, old, vec_type, "");
3100 ret = LLVMGetUndef(vec_type);
3101 for (unsigned i = 0; i < bits / 32; i++) {
3102 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3103 LLVMConstInt(ctx->i32, i,
3104 0), "");
3105 old = LLVMBuildExtractElement(ctx->builder, old_vector,
3106 LLVMConstInt(ctx->i32, i,
3107 0), "");
3108 LLVMValueRef ret_comp = _ac_build_dpp(ctx, old, src,
3109 dpp_ctrl,
3110 row_mask,
3111 bank_mask,
3112 bound_ctrl);
3113 ret = LLVMBuildInsertElement(ctx->builder, ret,
3114 ret_comp,
3115 LLVMConstInt(ctx->i32, i,
3116 0), "");
3117 }
3118 }
3119 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3120 }
3121
3122 static inline unsigned
3123 ds_pattern_bitmode(unsigned and_mask, unsigned or_mask, unsigned xor_mask)
3124 {
3125 assert(and_mask < 32 && or_mask < 32 && xor_mask < 32);
3126 return and_mask | (or_mask << 5) | (xor_mask << 10);
3127 }
3128
3129 static LLVMValueRef
3130 _ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
3131 {
3132 return ac_build_intrinsic(ctx, "llvm.amdgcn.ds.swizzle",
3133 LLVMTypeOf(src), (LLVMValueRef []) {
3134 src, LLVMConstInt(ctx->i32, mask, 0) },
3135 2, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3136 }
3137
3138 LLVMValueRef
3139 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
3140 {
3141 LLVMTypeRef src_type = LLVMTypeOf(src);
3142 src = ac_to_integer(ctx, src);
3143 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3144 LLVMValueRef ret;
3145 if (bits == 32) {
3146 ret = _ac_build_ds_swizzle(ctx, src, mask);
3147 } else {
3148 assert(bits % 32 == 0);
3149 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3150 LLVMValueRef src_vector =
3151 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3152 ret = LLVMGetUndef(vec_type);
3153 for (unsigned i = 0; i < bits / 32; i++) {
3154 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3155 LLVMConstInt(ctx->i32, i,
3156 0), "");
3157 LLVMValueRef ret_comp = _ac_build_ds_swizzle(ctx, src,
3158 mask);
3159 ret = LLVMBuildInsertElement(ctx->builder, ret,
3160 ret_comp,
3161 LLVMConstInt(ctx->i32, i,
3162 0), "");
3163 }
3164 }
3165 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3166 }
3167
3168 static LLVMValueRef
3169 ac_build_wwm(struct ac_llvm_context *ctx, LLVMValueRef src)
3170 {
3171 char name[32], type[8];
3172 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
3173 snprintf(name, sizeof(name), "llvm.amdgcn.wwm.%s", type);
3174 return ac_build_intrinsic(ctx, name, LLVMTypeOf(src),
3175 (LLVMValueRef []) { src }, 1,
3176 AC_FUNC_ATTR_READNONE);
3177 }
3178
3179 static LLVMValueRef
3180 ac_build_set_inactive(struct ac_llvm_context *ctx, LLVMValueRef src,
3181 LLVMValueRef inactive)
3182 {
3183 char name[33], type[8];
3184 LLVMTypeRef src_type = LLVMTypeOf(src);
3185 src = ac_to_integer(ctx, src);
3186 inactive = ac_to_integer(ctx, inactive);
3187 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
3188 snprintf(name, sizeof(name), "llvm.amdgcn.set.inactive.%s", type);
3189 LLVMValueRef ret =
3190 ac_build_intrinsic(ctx, name,
3191 LLVMTypeOf(src), (LLVMValueRef []) {
3192 src, inactive }, 2,
3193 AC_FUNC_ATTR_READNONE |
3194 AC_FUNC_ATTR_CONVERGENT);
3195 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3196 }
3197
3198 static LLVMValueRef
3199 get_reduction_identity(struct ac_llvm_context *ctx, nir_op op, unsigned type_size)
3200 {
3201 if (type_size == 4) {
3202 switch (op) {
3203 case nir_op_iadd: return ctx->i32_0;
3204 case nir_op_fadd: return ctx->f32_0;
3205 case nir_op_imul: return ctx->i32_1;
3206 case nir_op_fmul: return ctx->f32_1;
3207 case nir_op_imin: return LLVMConstInt(ctx->i32, INT32_MAX, 0);
3208 case nir_op_umin: return LLVMConstInt(ctx->i32, UINT32_MAX, 0);
3209 case nir_op_fmin: return LLVMConstReal(ctx->f32, INFINITY);
3210 case nir_op_imax: return LLVMConstInt(ctx->i32, INT32_MIN, 0);
3211 case nir_op_umax: return ctx->i32_0;
3212 case nir_op_fmax: return LLVMConstReal(ctx->f32, -INFINITY);
3213 case nir_op_iand: return LLVMConstInt(ctx->i32, -1, 0);
3214 case nir_op_ior: return ctx->i32_0;
3215 case nir_op_ixor: return ctx->i32_0;
3216 default:
3217 unreachable("bad reduction intrinsic");
3218 }
3219 } else { /* type_size == 64bit */
3220 switch (op) {
3221 case nir_op_iadd: return ctx->i64_0;
3222 case nir_op_fadd: return ctx->f64_0;
3223 case nir_op_imul: return ctx->i64_1;
3224 case nir_op_fmul: return ctx->f64_1;
3225 case nir_op_imin: return LLVMConstInt(ctx->i64, INT64_MAX, 0);
3226 case nir_op_umin: return LLVMConstInt(ctx->i64, UINT64_MAX, 0);
3227 case nir_op_fmin: return LLVMConstReal(ctx->f64, INFINITY);
3228 case nir_op_imax: return LLVMConstInt(ctx->i64, INT64_MIN, 0);
3229 case nir_op_umax: return ctx->i64_0;
3230 case nir_op_fmax: return LLVMConstReal(ctx->f64, -INFINITY);
3231 case nir_op_iand: return LLVMConstInt(ctx->i64, -1, 0);
3232 case nir_op_ior: return ctx->i64_0;
3233 case nir_op_ixor: return ctx->i64_0;
3234 default:
3235 unreachable("bad reduction intrinsic");
3236 }
3237 }
3238 }
3239
3240 static LLVMValueRef
3241 ac_build_alu_op(struct ac_llvm_context *ctx, LLVMValueRef lhs, LLVMValueRef rhs, nir_op op)
3242 {
3243 bool _64bit = ac_get_type_size(LLVMTypeOf(lhs)) == 8;
3244 switch (op) {
3245 case nir_op_iadd: return LLVMBuildAdd(ctx->builder, lhs, rhs, "");
3246 case nir_op_fadd: return LLVMBuildFAdd(ctx->builder, lhs, rhs, "");
3247 case nir_op_imul: return LLVMBuildMul(ctx->builder, lhs, rhs, "");
3248 case nir_op_fmul: return LLVMBuildFMul(ctx->builder, lhs, rhs, "");
3249 case nir_op_imin: return LLVMBuildSelect(ctx->builder,
3250 LLVMBuildICmp(ctx->builder, LLVMIntSLT, lhs, rhs, ""),
3251 lhs, rhs, "");
3252 case nir_op_umin: return LLVMBuildSelect(ctx->builder,
3253 LLVMBuildICmp(ctx->builder, LLVMIntULT, lhs, rhs, ""),
3254 lhs, rhs, "");
3255 case nir_op_fmin: return ac_build_intrinsic(ctx,
3256 _64bit ? "llvm.minnum.f64" : "llvm.minnum.f32",
3257 _64bit ? ctx->f64 : ctx->f32,
3258 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
3259 case nir_op_imax: return LLVMBuildSelect(ctx->builder,
3260 LLVMBuildICmp(ctx->builder, LLVMIntSGT, lhs, rhs, ""),
3261 lhs, rhs, "");
3262 case nir_op_umax: return LLVMBuildSelect(ctx->builder,
3263 LLVMBuildICmp(ctx->builder, LLVMIntUGT, lhs, rhs, ""),
3264 lhs, rhs, "");
3265 case nir_op_fmax: return ac_build_intrinsic(ctx,
3266 _64bit ? "llvm.maxnum.f64" : "llvm.maxnum.f32",
3267 _64bit ? ctx->f64 : ctx->f32,
3268 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
3269 case nir_op_iand: return LLVMBuildAnd(ctx->builder, lhs, rhs, "");
3270 case nir_op_ior: return LLVMBuildOr(ctx->builder, lhs, rhs, "");
3271 case nir_op_ixor: return LLVMBuildXor(ctx->builder, lhs, rhs, "");
3272 default:
3273 unreachable("bad reduction intrinsic");
3274 }
3275 }
3276
3277 /**
3278 * \param maxprefix specifies that the result only needs to be correct for a
3279 * prefix of this many threads
3280 *
3281 * TODO: add inclusive and excluse scan functions for SI chip class.
3282 */
3283 static LLVMValueRef
3284 ac_build_scan(struct ac_llvm_context *ctx, nir_op op, LLVMValueRef src, LLVMValueRef identity,
3285 unsigned maxprefix)
3286 {
3287 LLVMValueRef result, tmp;
3288 result = src;
3289 if (maxprefix <= 1)
3290 return result;
3291 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
3292 result = ac_build_alu_op(ctx, result, tmp, op);
3293 if (maxprefix <= 2)
3294 return result;
3295 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(2), 0xf, 0xf, false);
3296 result = ac_build_alu_op(ctx, result, tmp, op);
3297 if (maxprefix <= 3)
3298 return result;
3299 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(3), 0xf, 0xf, false);
3300 result = ac_build_alu_op(ctx, result, tmp, op);
3301 if (maxprefix <= 4)
3302 return result;
3303 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(4), 0xf, 0xe, false);
3304 result = ac_build_alu_op(ctx, result, tmp, op);
3305 if (maxprefix <= 8)
3306 return result;
3307 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(8), 0xf, 0xc, false);
3308 result = ac_build_alu_op(ctx, result, tmp, op);
3309 if (maxprefix <= 16)
3310 return result;
3311 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
3312 result = ac_build_alu_op(ctx, result, tmp, op);
3313 if (maxprefix <= 32)
3314 return result;
3315 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
3316 result = ac_build_alu_op(ctx, result, tmp, op);
3317 return result;
3318 }
3319
3320 LLVMValueRef
3321 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
3322 {
3323 LLVMValueRef result;
3324
3325 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
3326 LLVMBuilderRef builder = ctx->builder;
3327 src = LLVMBuildZExt(builder, src, ctx->i32, "");
3328 result = ac_build_ballot(ctx, src);
3329 result = ac_build_mbcnt(ctx, result);
3330 result = LLVMBuildAdd(builder, result, src, "");
3331 return result;
3332 }
3333
3334 ac_build_optimization_barrier(ctx, &src);
3335
3336 LLVMValueRef identity =
3337 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
3338 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
3339 LLVMTypeOf(identity), "");
3340 result = ac_build_scan(ctx, op, result, identity, 64);
3341
3342 return ac_build_wwm(ctx, result);
3343 }
3344
3345 LLVMValueRef
3346 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
3347 {
3348 LLVMValueRef result;
3349
3350 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
3351 LLVMBuilderRef builder = ctx->builder;
3352 src = LLVMBuildZExt(builder, src, ctx->i32, "");
3353 result = ac_build_ballot(ctx, src);
3354 result = ac_build_mbcnt(ctx, result);
3355 return result;
3356 }
3357
3358 ac_build_optimization_barrier(ctx, &src);
3359
3360 LLVMValueRef identity =
3361 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
3362 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
3363 LLVMTypeOf(identity), "");
3364 result = ac_build_dpp(ctx, identity, result, dpp_wf_sr1, 0xf, 0xf, false);
3365 result = ac_build_scan(ctx, op, result, identity, 64);
3366
3367 return ac_build_wwm(ctx, result);
3368 }
3369
3370 LLVMValueRef
3371 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size)
3372 {
3373 if (cluster_size == 1) return src;
3374 ac_build_optimization_barrier(ctx, &src);
3375 LLVMValueRef result, swap;
3376 LLVMValueRef identity = get_reduction_identity(ctx, op,
3377 ac_get_type_size(LLVMTypeOf(src)));
3378 result = LLVMBuildBitCast(ctx->builder,
3379 ac_build_set_inactive(ctx, src, identity),
3380 LLVMTypeOf(identity), "");
3381 swap = ac_build_quad_swizzle(ctx, result, 1, 0, 3, 2);
3382 result = ac_build_alu_op(ctx, result, swap, op);
3383 if (cluster_size == 2) return ac_build_wwm(ctx, result);
3384
3385 swap = ac_build_quad_swizzle(ctx, result, 2, 3, 0, 1);
3386 result = ac_build_alu_op(ctx, result, swap, op);
3387 if (cluster_size == 4) return ac_build_wwm(ctx, result);
3388
3389 if (ctx->chip_class >= VI)
3390 swap = ac_build_dpp(ctx, identity, result, dpp_row_half_mirror, 0xf, 0xf, false);
3391 else
3392 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x04));
3393 result = ac_build_alu_op(ctx, result, swap, op);
3394 if (cluster_size == 8) return ac_build_wwm(ctx, result);
3395
3396 if (ctx->chip_class >= VI)
3397 swap = ac_build_dpp(ctx, identity, result, dpp_row_mirror, 0xf, 0xf, false);
3398 else
3399 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x08));
3400 result = ac_build_alu_op(ctx, result, swap, op);
3401 if (cluster_size == 16) return ac_build_wwm(ctx, result);
3402
3403 if (ctx->chip_class >= VI && cluster_size != 32)
3404 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
3405 else
3406 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x10));
3407 result = ac_build_alu_op(ctx, result, swap, op);
3408 if (cluster_size == 32) return ac_build_wwm(ctx, result);
3409
3410 if (ctx->chip_class >= VI) {
3411 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
3412 result = ac_build_alu_op(ctx, result, swap, op);
3413 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 63, 0));
3414 return ac_build_wwm(ctx, result);
3415 } else {
3416 swap = ac_build_readlane(ctx, result, ctx->i32_0);
3417 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 32, 0));
3418 result = ac_build_alu_op(ctx, result, swap, op);
3419 return ac_build_wwm(ctx, result);
3420 }
3421 }
3422
3423 /**
3424 * "Top half" of a scan that reduces per-wave values across an entire
3425 * workgroup.
3426 *
3427 * The source value must be present in the highest lane of the wave, and the
3428 * highest lane must be live.
3429 */
3430 void
3431 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3432 {
3433 if (ws->maxwaves <= 1)
3434 return;
3435
3436 const LLVMValueRef i32_63 = LLVMConstInt(ctx->i32, 63, false);
3437 LLVMBuilderRef builder = ctx->builder;
3438 LLVMValueRef tid = ac_get_thread_id(ctx);
3439 LLVMValueRef tmp;
3440
3441 tmp = LLVMBuildICmp(builder, LLVMIntEQ, tid, i32_63, "");
3442 ac_build_ifcc(ctx, tmp, 1000);
3443 LLVMBuildStore(builder, ws->src, LLVMBuildGEP(builder, ws->scratch, &ws->waveidx, 1, ""));
3444 ac_build_endif(ctx, 1000);
3445 }
3446
3447 /**
3448 * "Bottom half" of a scan that reduces per-wave values across an entire
3449 * workgroup.
3450 *
3451 * The caller must place a barrier between the top and bottom halves.
3452 */
3453 void
3454 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3455 {
3456 const LLVMTypeRef type = LLVMTypeOf(ws->src);
3457 const LLVMValueRef identity =
3458 get_reduction_identity(ctx, ws->op, ac_get_type_size(type));
3459
3460 if (ws->maxwaves <= 1) {
3461 ws->result_reduce = ws->src;
3462 ws->result_inclusive = ws->src;
3463 ws->result_exclusive = identity;
3464 return;
3465 }
3466 assert(ws->maxwaves <= 32);
3467
3468 LLVMBuilderRef builder = ctx->builder;
3469 LLVMValueRef tid = ac_get_thread_id(ctx);
3470 LLVMBasicBlockRef bbs[2];
3471 LLVMValueRef phivalues_scan[2];
3472 LLVMValueRef tmp, tmp2;
3473
3474 bbs[0] = LLVMGetInsertBlock(builder);
3475 phivalues_scan[0] = LLVMGetUndef(type);
3476
3477 if (ws->enable_reduce)
3478 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->numwaves, "");
3479 else if (ws->enable_inclusive)
3480 tmp = LLVMBuildICmp(builder, LLVMIntULE, tid, ws->waveidx, "");
3481 else
3482 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->waveidx, "");
3483 ac_build_ifcc(ctx, tmp, 1001);
3484 {
3485 tmp = LLVMBuildLoad(builder, LLVMBuildGEP(builder, ws->scratch, &tid, 1, ""), "");
3486
3487 ac_build_optimization_barrier(ctx, &tmp);
3488
3489 bbs[1] = LLVMGetInsertBlock(builder);
3490 phivalues_scan[1] = ac_build_scan(ctx, ws->op, tmp, identity, ws->maxwaves);
3491 }
3492 ac_build_endif(ctx, 1001);
3493
3494 const LLVMValueRef scan = ac_build_phi(ctx, type, 2, phivalues_scan, bbs);
3495
3496 if (ws->enable_reduce) {
3497 tmp = LLVMBuildSub(builder, ws->numwaves, ctx->i32_1, "");
3498 ws->result_reduce = ac_build_readlane(ctx, scan, tmp);
3499 }
3500 if (ws->enable_inclusive)
3501 ws->result_inclusive = ac_build_readlane(ctx, scan, ws->waveidx);
3502 if (ws->enable_exclusive) {
3503 tmp = LLVMBuildSub(builder, ws->waveidx, ctx->i32_1, "");
3504 tmp = ac_build_readlane(ctx, scan, tmp);
3505 tmp2 = LLVMBuildICmp(builder, LLVMIntEQ, ws->waveidx, ctx->i32_0, "");
3506 ws->result_exclusive = LLVMBuildSelect(builder, tmp2, identity, tmp, "");
3507 }
3508 }
3509
3510 /**
3511 * Inclusive scan of a per-wave value across an entire workgroup.
3512 *
3513 * This implies an s_barrier instruction.
3514 *
3515 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
3516 * of the workgroup are live. (This requirement cannot easily be relaxed in a
3517 * useful manner because of the barrier in the algorithm.)
3518 */
3519 void
3520 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3521 {
3522 ac_build_wg_wavescan_top(ctx, ws);
3523 ac_build_s_barrier(ctx);
3524 ac_build_wg_wavescan_bottom(ctx, ws);
3525 }
3526
3527 /**
3528 * "Top half" of a scan that reduces per-thread values across an entire
3529 * workgroup.
3530 *
3531 * All lanes must be active when this code runs.
3532 */
3533 void
3534 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3535 {
3536 if (ws->enable_exclusive) {
3537 ws->extra = ac_build_exclusive_scan(ctx, ws->src, ws->op);
3538 if (LLVMTypeOf(ws->src) == ctx->i1 && ws->op == nir_op_iadd)
3539 ws->src = LLVMBuildZExt(ctx->builder, ws->src, ctx->i32, "");
3540 ws->src = ac_build_alu_op(ctx, ws->extra, ws->src, ws->op);
3541 } else {
3542 ws->src = ac_build_inclusive_scan(ctx, ws->src, ws->op);
3543 }
3544
3545 bool enable_inclusive = ws->enable_inclusive;
3546 bool enable_exclusive = ws->enable_exclusive;
3547 ws->enable_inclusive = false;
3548 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
3549 ac_build_wg_wavescan_top(ctx, ws);
3550 ws->enable_inclusive = enable_inclusive;
3551 ws->enable_exclusive = enable_exclusive;
3552 }
3553
3554 /**
3555 * "Bottom half" of a scan that reduces per-thread values across an entire
3556 * workgroup.
3557 *
3558 * The caller must place a barrier between the top and bottom halves.
3559 */
3560 void
3561 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3562 {
3563 bool enable_inclusive = ws->enable_inclusive;
3564 bool enable_exclusive = ws->enable_exclusive;
3565 ws->enable_inclusive = false;
3566 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
3567 ac_build_wg_wavescan_bottom(ctx, ws);
3568 ws->enable_inclusive = enable_inclusive;
3569 ws->enable_exclusive = enable_exclusive;
3570
3571 /* ws->result_reduce is already the correct value */
3572 if (ws->enable_inclusive)
3573 ws->result_inclusive = ac_build_alu_op(ctx, ws->result_exclusive, ws->src, ws->op);
3574 if (ws->enable_exclusive)
3575 ws->result_exclusive = ac_build_alu_op(ctx, ws->result_exclusive, ws->extra, ws->op);
3576 }
3577
3578 /**
3579 * A scan that reduces per-thread values across an entire workgroup.
3580 *
3581 * The caller must ensure that all lanes are active when this code runs
3582 * (WWM is insufficient!), because there is an implied barrier.
3583 */
3584 void
3585 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3586 {
3587 ac_build_wg_scan_top(ctx, ws);
3588 ac_build_s_barrier(ctx);
3589 ac_build_wg_scan_bottom(ctx, ws);
3590 }
3591
3592 LLVMValueRef
3593 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
3594 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
3595 {
3596 unsigned mask = dpp_quad_perm(lane0, lane1, lane2, lane3);
3597 if (ctx->chip_class >= VI) {
3598 return ac_build_dpp(ctx, src, src, mask, 0xf, 0xf, false);
3599 } else {
3600 return ac_build_ds_swizzle(ctx, src, (1 << 15) | mask);
3601 }
3602 }
3603
3604 LLVMValueRef
3605 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index)
3606 {
3607 index = LLVMBuildMul(ctx->builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
3608 return ac_build_intrinsic(ctx,
3609 "llvm.amdgcn.ds.bpermute", ctx->i32,
3610 (LLVMValueRef []) {index, src}, 2,
3611 AC_FUNC_ATTR_READNONE |
3612 AC_FUNC_ATTR_CONVERGENT);
3613 }