projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
radv: Fix float16 interpolation set up.
[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 LLVMValueRef
1373 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
1374 LLVMValueRef rsrc,
1375 LLVMValueRef vindex,
1376 LLVMValueRef voffset,
1377 LLVMValueRef soffset,
1378 LLVMValueRef immoffset,
1379 LLVMValueRef glc)
1380 {
1381 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
1382 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1383 LLVMValueRef res;
1384
1385 if (HAVE_LLVM >= 0x0800) {
1386 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1387
1388 res = ac_build_llvm8_tbuffer_load(ctx, rsrc, vindex, voffset,
1389 soffset, 1, dfmt, nfmt, glc,
1390 false, true, true);
1391 } else {
1392 const char *name = "llvm.amdgcn.tbuffer.load.i32";
1393 LLVMTypeRef type = ctx->i32;
1394 LLVMValueRef params[] = {
1395 rsrc,
1396 vindex,
1397 voffset,
1398 soffset,
1399 immoffset,
1400 LLVMConstInt(ctx->i32, dfmt, false),
1401 LLVMConstInt(ctx->i32, nfmt, false),
1402 glc,
1403 ctx->i1false,
1404 };
1405 res = ac_build_intrinsic(ctx, name, type, params, 9, 0);
1406 }
1407
1408 return LLVMBuildTrunc(ctx->builder, res, ctx->i16, "");
1409 }
1410
1411 LLVMValueRef
1412 ac_build_llvm8_tbuffer_load(struct ac_llvm_context *ctx,
1413 LLVMValueRef rsrc,
1414 LLVMValueRef vindex,
1415 LLVMValueRef voffset,
1416 LLVMValueRef soffset,
1417 unsigned num_channels,
1418 unsigned dfmt,
1419 unsigned nfmt,
1420 bool glc,
1421 bool slc,
1422 bool can_speculate,
1423 bool structurized)
1424 {
1425 LLVMValueRef args[6];
1426 int idx = 0;
1427 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1428 if (structurized)
1429 args[idx++] = vindex ? vindex : ctx->i32_0;
1430 args[idx++] = voffset ? voffset : ctx->i32_0;
1431 args[idx++] = soffset ? soffset : ctx->i32_0;
1432 args[idx++] = LLVMConstInt(ctx->i32, dfmt | (nfmt << 4), 0);
1433 args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0);
1434 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1435
1436 LLVMTypeRef types[] = {ctx->i32, ctx->v2i32, ctx->v4i32};
1437 const char *type_names[] = {"i32", "v2i32", "v4i32"};
1438 const char *indexing_kind = structurized ? "struct" : "raw";
1439 char name[256];
1440
1441 snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.load.%s",
1442 indexing_kind, type_names[func]);
1443
1444 return ac_build_intrinsic(ctx, name, types[func], args,
1445 idx,
1446 ac_get_load_intr_attribs(can_speculate));
1447 }
1448
1449 /**
1450 * Set range metadata on an instruction. This can only be used on load and
1451 * call instructions. If you know an instruction can only produce the values
1452 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1453 * \p lo is the minimum value inclusive.
1454 * \p hi is the maximum value exclusive.
1455 */
1456 static void set_range_metadata(struct ac_llvm_context *ctx,
1457 LLVMValueRef value, unsigned lo, unsigned hi)
1458 {
1459 LLVMValueRef range_md, md_args[2];
1460 LLVMTypeRef type = LLVMTypeOf(value);
1461 LLVMContextRef context = LLVMGetTypeContext(type);
1462
1463 md_args[0] = LLVMConstInt(type, lo, false);
1464 md_args[1] = LLVMConstInt(type, hi, false);
1465 range_md = LLVMMDNodeInContext(context, md_args, 2);
1466 LLVMSetMetadata(value, ctx->range_md_kind, range_md);
1467 }
1468
1469 LLVMValueRef
1470 ac_get_thread_id(struct ac_llvm_context *ctx)
1471 {
1472 LLVMValueRef tid;
1473
1474 LLVMValueRef tid_args[2];
1475 tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
1476 tid_args[1] = ctx->i32_0;
1477 tid_args[1] = ac_build_intrinsic(ctx,
1478 "llvm.amdgcn.mbcnt.lo", ctx->i32,
1479 tid_args, 2, AC_FUNC_ATTR_READNONE);
1480
1481 tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
1482 ctx->i32, tid_args,
1483 2, AC_FUNC_ATTR_READNONE);
1484 set_range_metadata(ctx, tid, 0, 64);
1485 return tid;
1486 }
1487
1488 /*
1489 * SI implements derivatives using the local data store (LDS)
1490 * All writes to the LDS happen in all executing threads at
1491 * the same time. TID is the Thread ID for the current
1492 * thread and is a value between 0 and 63, representing
1493 * the thread's position in the wavefront.
1494 *
1495 * For the pixel shader threads are grouped into quads of four pixels.
1496 * The TIDs of the pixels of a quad are:
1497 *
1498 * +------+------+
1499 * |4n + 0|4n + 1|
1500 * +------+------+
1501 * |4n + 2|4n + 3|
1502 * +------+------+
1503 *
1504 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
1505 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
1506 * the current pixel's column, and masking with 0xfffffffe yields the TID
1507 * of the left pixel of the current pixel's row.
1508 *
1509 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
1510 * adding 2 yields the TID of the pixel below the top pixel.
1511 */
1512 LLVMValueRef
1513 ac_build_ddxy(struct ac_llvm_context *ctx,
1514 uint32_t mask,
1515 int idx,
1516 LLVMValueRef val)
1517 {
1518 unsigned tl_lanes[4], trbl_lanes[4];
1519 LLVMValueRef tl, trbl;
1520 LLVMValueRef result;
1521
1522 for (unsigned i = 0; i < 4; ++i) {
1523 tl_lanes[i] = i & mask;
1524 trbl_lanes[i] = (i & mask) + idx;
1525 }
1526
1527 tl = ac_build_quad_swizzle(ctx, val,
1528 tl_lanes[0], tl_lanes[1],
1529 tl_lanes[2], tl_lanes[3]);
1530 trbl = ac_build_quad_swizzle(ctx, val,
1531 trbl_lanes[0], trbl_lanes[1],
1532 trbl_lanes[2], trbl_lanes[3]);
1533
1534 tl = LLVMBuildBitCast(ctx->builder, tl, ctx->f32, "");
1535 trbl = LLVMBuildBitCast(ctx->builder, trbl, ctx->f32, "");
1536 result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
1537
1538 result = ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.f32", ctx->f32,
1539 &result, 1, 0);
1540
1541 return result;
1542 }
1543
1544 void
1545 ac_build_sendmsg(struct ac_llvm_context *ctx,
1546 uint32_t msg,
1547 LLVMValueRef wave_id)
1548 {
1549 LLVMValueRef args[2];
1550 args[0] = LLVMConstInt(ctx->i32, msg, false);
1551 args[1] = wave_id;
1552 ac_build_intrinsic(ctx, "llvm.amdgcn.s.sendmsg", ctx->voidt, args, 2, 0);
1553 }
1554
1555 LLVMValueRef
1556 ac_build_imsb(struct ac_llvm_context *ctx,
1557 LLVMValueRef arg,
1558 LLVMTypeRef dst_type)
1559 {
1560 LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.amdgcn.sffbh.i32",
1561 dst_type, &arg, 1,
1562 AC_FUNC_ATTR_READNONE);
1563
1564 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
1565 * the index from LSB. Invert it by doing "31 - msb". */
1566 msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
1567 msb, "");
1568
1569 LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
1570 LLVMValueRef cond = LLVMBuildOr(ctx->builder,
1571 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
1572 arg, ctx->i32_0, ""),
1573 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
1574 arg, all_ones, ""), "");
1575
1576 return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
1577 }
1578
1579 LLVMValueRef
1580 ac_build_umsb(struct ac_llvm_context *ctx,
1581 LLVMValueRef arg,
1582 LLVMTypeRef dst_type)
1583 {
1584 const char *intrin_name;
1585 LLVMTypeRef type;
1586 LLVMValueRef highest_bit;
1587 LLVMValueRef zero;
1588 unsigned bitsize;
1589
1590 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(arg));
1591 switch (bitsize) {
1592 case 64:
1593 intrin_name = "llvm.ctlz.i64";
1594 type = ctx->i64;
1595 highest_bit = LLVMConstInt(ctx->i64, 63, false);
1596 zero = ctx->i64_0;
1597 break;
1598 case 32:
1599 intrin_name = "llvm.ctlz.i32";
1600 type = ctx->i32;
1601 highest_bit = LLVMConstInt(ctx->i32, 31, false);
1602 zero = ctx->i32_0;
1603 break;
1604 case 16:
1605 intrin_name = "llvm.ctlz.i16";
1606 type = ctx->i16;
1607 highest_bit = LLVMConstInt(ctx->i16, 15, false);
1608 zero = ctx->i16_0;
1609 break;
1610 default:
1611 unreachable(!"invalid bitsize");
1612 break;
1613 }
1614
1615 LLVMValueRef params[2] = {
1616 arg,
1617 ctx->i1true,
1618 };
1619
1620 LLVMValueRef msb = ac_build_intrinsic(ctx, intrin_name, type,
1621 params, 2,
1622 AC_FUNC_ATTR_READNONE);
1623
1624 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
1625 * the index from LSB. Invert it by doing "31 - msb". */
1626 msb = LLVMBuildSub(ctx->builder, highest_bit, msb, "");
1627 msb = LLVMBuildTruncOrBitCast(ctx->builder, msb, ctx->i32, "");
1628
1629 /* check for zero */
1630 return LLVMBuildSelect(ctx->builder,
1631 LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg, zero, ""),
1632 LLVMConstInt(ctx->i32, -1, true), msb, "");
1633 }
1634
1635 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
1636 LLVMValueRef b)
1637 {
1638 char name[64];
1639 snprintf(name, sizeof(name), "llvm.minnum.f%d", ac_get_elem_bits(ctx, LLVMTypeOf(a)));
1640 LLVMValueRef args[2] = {a, b};
1641 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
1642 AC_FUNC_ATTR_READNONE);
1643 }
1644
1645 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
1646 LLVMValueRef b)
1647 {
1648 char name[64];
1649 snprintf(name, sizeof(name), "llvm.maxnum.f%d", ac_get_elem_bits(ctx, LLVMTypeOf(a)));
1650 LLVMValueRef args[2] = {a, b};
1651 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
1652 AC_FUNC_ATTR_READNONE);
1653 }
1654
1655 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
1656 LLVMValueRef b)
1657 {
1658 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSLE, a, b, "");
1659 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
1660 }
1661
1662 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
1663 LLVMValueRef b)
1664 {
1665 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, a, b, "");
1666 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
1667 }
1668
1669 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a,
1670 LLVMValueRef b)
1671 {
1672 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntULE, a, b, "");
1673 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
1674 }
1675
1676 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value)
1677 {
1678 LLVMTypeRef t = LLVMTypeOf(value);
1679 return ac_build_fmin(ctx, ac_build_fmax(ctx, value, LLVMConstReal(t, 0.0)),
1680 LLVMConstReal(t, 1.0));
1681 }
1682
1683 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a)
1684 {
1685 LLVMValueRef args[9];
1686
1687 args[0] = LLVMConstInt(ctx->i32, a->target, 0);
1688 args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
1689
1690 if (a->compr) {
1691 LLVMTypeRef i16 = LLVMInt16TypeInContext(ctx->context);
1692 LLVMTypeRef v2i16 = LLVMVectorType(i16, 2);
1693
1694 args[2] = LLVMBuildBitCast(ctx->builder, a->out[0],
1695 v2i16, "");
1696 args[3] = LLVMBuildBitCast(ctx->builder, a->out[1],
1697 v2i16, "");
1698 args[4] = LLVMConstInt(ctx->i1, a->done, 0);
1699 args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
1700
1701 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16",
1702 ctx->voidt, args, 6, 0);
1703 } else {
1704 args[2] = a->out[0];
1705 args[3] = a->out[1];
1706 args[4] = a->out[2];
1707 args[5] = a->out[3];
1708 args[6] = LLVMConstInt(ctx->i1, a->done, 0);
1709 args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
1710
1711 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.f32",
1712 ctx->voidt, args, 8, 0);
1713 }
1714 }
1715
1716 void ac_build_export_null(struct ac_llvm_context *ctx)
1717 {
1718 struct ac_export_args args;
1719
1720 args.enabled_channels = 0x0; /* enabled channels */
1721 args.valid_mask = 1; /* whether the EXEC mask is valid */
1722 args.done = 1; /* DONE bit */
1723 args.target = V_008DFC_SQ_EXP_NULL;
1724 args.compr = 0; /* COMPR flag (0 = 32-bit export) */
1725 args.out[0] = LLVMGetUndef(ctx->f32); /* R */
1726 args.out[1] = LLVMGetUndef(ctx->f32); /* G */
1727 args.out[2] = LLVMGetUndef(ctx->f32); /* B */
1728 args.out[3] = LLVMGetUndef(ctx->f32); /* A */
1729
1730 ac_build_export(ctx, &args);
1731 }
1732
1733 static unsigned ac_num_coords(enum ac_image_dim dim)
1734 {
1735 switch (dim) {
1736 case ac_image_1d:
1737 return 1;
1738 case ac_image_2d:
1739 case ac_image_1darray:
1740 return 2;
1741 case ac_image_3d:
1742 case ac_image_cube:
1743 case ac_image_2darray:
1744 case ac_image_2dmsaa:
1745 return 3;
1746 case ac_image_2darraymsaa:
1747 return 4;
1748 default:
1749 unreachable("ac_num_coords: bad dim");
1750 }
1751 }
1752
1753 static unsigned ac_num_derivs(enum ac_image_dim dim)
1754 {
1755 switch (dim) {
1756 case ac_image_1d:
1757 case ac_image_1darray:
1758 return 2;
1759 case ac_image_2d:
1760 case ac_image_2darray:
1761 case ac_image_cube:
1762 return 4;
1763 case ac_image_3d:
1764 return 6;
1765 case ac_image_2dmsaa:
1766 case ac_image_2darraymsaa:
1767 default:
1768 unreachable("derivatives not supported");
1769 }
1770 }
1771
1772 static const char *get_atomic_name(enum ac_atomic_op op)
1773 {
1774 switch (op) {
1775 case ac_atomic_swap: return "swap";
1776 case ac_atomic_add: return "add";
1777 case ac_atomic_sub: return "sub";
1778 case ac_atomic_smin: return "smin";
1779 case ac_atomic_umin: return "umin";
1780 case ac_atomic_smax: return "smax";
1781 case ac_atomic_umax: return "umax";
1782 case ac_atomic_and: return "and";
1783 case ac_atomic_or: return "or";
1784 case ac_atomic_xor: return "xor";
1785 }
1786 unreachable("bad atomic op");
1787 }
1788
1789 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
1790 struct ac_image_args *a)
1791 {
1792 const char *overload[3] = { "", "", "" };
1793 unsigned num_overloads = 0;
1794 LLVMValueRef args[18];
1795 unsigned num_args = 0;
1796 enum ac_image_dim dim = a->dim;
1797
1798 assert(!a->lod || a->lod == ctx->i32_0 || a->lod == ctx->f32_0 ||
1799 !a->level_zero);
1800 assert((a->opcode != ac_image_get_resinfo && a->opcode != ac_image_load_mip &&
1801 a->opcode != ac_image_store_mip) ||
1802 a->lod);
1803 assert(a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
1804 (!a->compare && !a->offset));
1805 assert((a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
1806 a->opcode == ac_image_get_lod) ||
1807 !a->bias);
1808 assert((a->bias ? 1 : 0) +
1809 (a->lod ? 1 : 0) +
1810 (a->level_zero ? 1 : 0) +
1811 (a->derivs[0] ? 1 : 0) <= 1);
1812
1813 if (a->opcode == ac_image_get_lod) {
1814 switch (dim) {
1815 case ac_image_1darray:
1816 dim = ac_image_1d;
1817 break;
1818 case ac_image_2darray:
1819 case ac_image_cube:
1820 dim = ac_image_2d;
1821 break;
1822 default:
1823 break;
1824 }
1825 }
1826
1827 bool sample = a->opcode == ac_image_sample ||
1828 a->opcode == ac_image_gather4 ||
1829 a->opcode == ac_image_get_lod;
1830 bool atomic = a->opcode == ac_image_atomic ||
1831 a->opcode == ac_image_atomic_cmpswap;
1832 LLVMTypeRef coord_type = sample ? ctx->f32 : ctx->i32;
1833
1834 if (atomic || a->opcode == ac_image_store || a->opcode == ac_image_store_mip) {
1835 args[num_args++] = a->data[0];
1836 if (a->opcode == ac_image_atomic_cmpswap)
1837 args[num_args++] = a->data[1];
1838 }
1839
1840 if (!atomic)
1841 args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, false);
1842
1843 if (a->offset)
1844 args[num_args++] = ac_to_integer(ctx, a->offset);
1845 if (a->bias) {
1846 args[num_args++] = ac_to_float(ctx, a->bias);
1847 overload[num_overloads++] = ".f32";
1848 }
1849 if (a->compare)
1850 args[num_args++] = ac_to_float(ctx, a->compare);
1851 if (a->derivs[0]) {
1852 unsigned count = ac_num_derivs(dim);
1853 for (unsigned i = 0; i < count; ++i)
1854 args[num_args++] = ac_to_float(ctx, a->derivs[i]);
1855 overload[num_overloads++] = ".f32";
1856 }
1857 unsigned num_coords =
1858 a->opcode != ac_image_get_resinfo ? ac_num_coords(dim) : 0;
1859 for (unsigned i = 0; i < num_coords; ++i)
1860 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->coords[i], coord_type, "");
1861 if (a->lod)
1862 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->lod, coord_type, "");
1863 overload[num_overloads++] = sample ? ".f32" : ".i32";
1864
1865 args[num_args++] = a->resource;
1866 if (sample) {
1867 args[num_args++] = a->sampler;
1868 args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, false);
1869 }
1870
1871 args[num_args++] = ctx->i32_0; /* texfailctrl */
1872 args[num_args++] = LLVMConstInt(ctx->i32, a->cache_policy, false);
1873
1874 const char *name;
1875 const char *atomic_subop = "";
1876 switch (a->opcode) {
1877 case ac_image_sample: name = "sample"; break;
1878 case ac_image_gather4: name = "gather4"; break;
1879 case ac_image_load: name = "load"; break;
1880 case ac_image_load_mip: name = "load.mip"; break;
1881 case ac_image_store: name = "store"; break;
1882 case ac_image_store_mip: name = "store.mip"; break;
1883 case ac_image_atomic:
1884 name = "atomic.";
1885 atomic_subop = get_atomic_name(a->atomic);
1886 break;
1887 case ac_image_atomic_cmpswap:
1888 name = "atomic.";
1889 atomic_subop = "cmpswap";
1890 break;
1891 case ac_image_get_lod: name = "getlod"; break;
1892 case ac_image_get_resinfo: name = "getresinfo"; break;
1893 default: unreachable("invalid image opcode");
1894 }
1895
1896 const char *dimname;
1897 switch (dim) {
1898 case ac_image_1d: dimname = "1d"; break;
1899 case ac_image_2d: dimname = "2d"; break;
1900 case ac_image_3d: dimname = "3d"; break;
1901 case ac_image_cube: dimname = "cube"; break;
1902 case ac_image_1darray: dimname = "1darray"; break;
1903 case ac_image_2darray: dimname = "2darray"; break;
1904 case ac_image_2dmsaa: dimname = "2dmsaa"; break;
1905 case ac_image_2darraymsaa: dimname = "2darraymsaa"; break;
1906 default: unreachable("invalid dim");
1907 }
1908
1909 bool lod_suffix =
1910 a->lod && (a->opcode == ac_image_sample || a->opcode == ac_image_gather4);
1911 char intr_name[96];
1912 snprintf(intr_name, sizeof(intr_name),
1913 "llvm.amdgcn.image.%s%s" /* base name */
1914 "%s%s%s" /* sample/gather modifiers */
1915 ".%s.%s%s%s%s", /* dimension and type overloads */
1916 name, atomic_subop,
1917 a->compare ? ".c" : "",
1918 a->bias ? ".b" :
1919 lod_suffix ? ".l" :
1920 a->derivs[0] ? ".d" :
1921 a->level_zero ? ".lz" : "",
1922 a->offset ? ".o" : "",
1923 dimname,
1924 atomic ? "i32" : "v4f32",
1925 overload[0], overload[1], overload[2]);
1926
1927 LLVMTypeRef retty;
1928 if (atomic)
1929 retty = ctx->i32;
1930 else if (a->opcode == ac_image_store || a->opcode == ac_image_store_mip)
1931 retty = ctx->voidt;
1932 else
1933 retty = ctx->v4f32;
1934
1935 LLVMValueRef result =
1936 ac_build_intrinsic(ctx, intr_name, retty, args, num_args,
1937 a->attributes);
1938 if (!sample && retty == ctx->v4f32) {
1939 result = LLVMBuildBitCast(ctx->builder, result,
1940 ctx->v4i32, "");
1941 }
1942 return result;
1943 }
1944
1945 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
1946 LLVMValueRef args[2])
1947 {
1948 LLVMTypeRef v2f16 =
1949 LLVMVectorType(LLVMHalfTypeInContext(ctx->context), 2);
1950
1951 return ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz", v2f16,
1952 args, 2, AC_FUNC_ATTR_READNONE);
1953 }
1954
1955 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
1956 LLVMValueRef args[2])
1957 {
1958 LLVMValueRef res =
1959 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16",
1960 ctx->v2i16, args, 2,
1961 AC_FUNC_ATTR_READNONE);
1962 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1963 }
1964
1965 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
1966 LLVMValueRef args[2])
1967 {
1968 LLVMValueRef res =
1969 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16",
1970 ctx->v2i16, args, 2,
1971 AC_FUNC_ATTR_READNONE);
1972 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1973 }
1974
1975 /* The 8-bit and 10-bit clamping is for HW workarounds. */
1976 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
1977 LLVMValueRef args[2], unsigned bits, bool hi)
1978 {
1979 assert(bits == 8 || bits == 10 || bits == 16);
1980
1981 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
1982 bits == 8 ? 127 : bits == 10 ? 511 : 32767, 0);
1983 LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
1984 bits == 8 ? -128 : bits == 10 ? -512 : -32768, 0);
1985 LLVMValueRef max_alpha =
1986 bits != 10 ? max_rgb : ctx->i32_1;
1987 LLVMValueRef min_alpha =
1988 bits != 10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
1989
1990 /* Clamp. */
1991 if (bits != 16) {
1992 for (int i = 0; i < 2; i++) {
1993 bool alpha = hi && i == 1;
1994 args[i] = ac_build_imin(ctx, args[i],
1995 alpha ? max_alpha : max_rgb);
1996 args[i] = ac_build_imax(ctx, args[i],
1997 alpha ? min_alpha : min_rgb);
1998 }
1999 }
2000
2001 LLVMValueRef res =
2002 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16",
2003 ctx->v2i16, args, 2,
2004 AC_FUNC_ATTR_READNONE);
2005 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2006 }
2007
2008 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2009 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
2010 LLVMValueRef args[2], unsigned bits, bool hi)
2011 {
2012 assert(bits == 8 || bits == 10 || bits == 16);
2013
2014 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2015 bits == 8 ? 255 : bits == 10 ? 1023 : 65535, 0);
2016 LLVMValueRef max_alpha =
2017 bits != 10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
2018
2019 /* Clamp. */
2020 if (bits != 16) {
2021 for (int i = 0; i < 2; i++) {
2022 bool alpha = hi && i == 1;
2023 args[i] = ac_build_umin(ctx, args[i],
2024 alpha ? max_alpha : max_rgb);
2025 }
2026 }
2027
2028 LLVMValueRef res =
2029 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16",
2030 ctx->v2i16, args, 2,
2031 AC_FUNC_ATTR_READNONE);
2032 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2033 }
2034
2035 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1)
2036 {
2037 return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1,
2038 &i1, 1, AC_FUNC_ATTR_READNONE);
2039 }
2040
2041 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1)
2042 {
2043 ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt,
2044 &i1, 1, 0);
2045 }
2046
2047 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
2048 LLVMValueRef offset, LLVMValueRef width,
2049 bool is_signed)
2050 {
2051 LLVMValueRef args[] = {
2052 input,
2053 offset,
2054 width,
2055 };
2056
2057 return ac_build_intrinsic(ctx,
2058 is_signed ? "llvm.amdgcn.sbfe.i32" :
2059 "llvm.amdgcn.ubfe.i32",
2060 ctx->i32, args, 3,
2061 AC_FUNC_ATTR_READNONE);
2062 }
2063
2064 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2065 LLVMValueRef s1, LLVMValueRef s2)
2066 {
2067 return LLVMBuildAdd(ctx->builder,
2068 LLVMBuildMul(ctx->builder, s0, s1, ""), s2, "");
2069 }
2070
2071 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2072 LLVMValueRef s1, LLVMValueRef s2)
2073 {
2074 return LLVMBuildFAdd(ctx->builder,
2075 LLVMBuildFMul(ctx->builder, s0, s1, ""), s2, "");
2076 }
2077
2078 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16)
2079 {
2080 LLVMValueRef args[1] = {
2081 LLVMConstInt(ctx->i32, simm16, false),
2082 };
2083 ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt",
2084 ctx->voidt, args, 1, 0);
2085 }
2086
2087 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
2088 unsigned bitsize)
2089 {
2090 LLVMTypeRef type;
2091 char *intr;
2092
2093 if (bitsize == 32) {
2094 intr = "llvm.floor.f32";
2095 type = ctx->f32;
2096 } else {
2097 intr = "llvm.floor.f64";
2098 type = ctx->f64;
2099 }
2100
2101 LLVMValueRef params[] = {
2102 src0,
2103 };
2104 LLVMValueRef floor = ac_build_intrinsic(ctx, intr, type, params, 1,
2105 AC_FUNC_ATTR_READNONE);
2106 return LLVMBuildFSub(ctx->builder, src0, floor, "");
2107 }
2108
2109 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2110 unsigned bitsize)
2111 {
2112 LLVMTypeRef type = LLVMIntTypeInContext(ctx->context, bitsize);
2113 LLVMValueRef zero = LLVMConstInt(type, 0, false);
2114 LLVMValueRef one = LLVMConstInt(type, 1, false);
2115
2116 LLVMValueRef cmp, val;
2117 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, src0, zero, "");
2118 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2119 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGE, val, zero, "");
2120 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstInt(type, -1, true), "");
2121 return val;
2122 }
2123
2124 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2125 unsigned bitsize)
2126 {
2127 LLVMValueRef cmp, val, zero, one;
2128 LLVMTypeRef type;
2129
2130 if (bitsize == 32) {
2131 type = ctx->f32;
2132 zero = ctx->f32_0;
2133 one = ctx->f32_1;
2134 } else {
2135 type = ctx->f64;
2136 zero = ctx->f64_0;
2137 one = ctx->f64_1;
2138 }
2139
2140 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src0, zero, "");
2141 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2142 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGE, val, zero, "");
2143 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstReal(type, -1.0), "");
2144 return val;
2145 }
2146
2147 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0)
2148 {
2149 LLVMValueRef result;
2150 unsigned bitsize;
2151
2152 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2153
2154 switch (bitsize) {
2155 case 64:
2156 result = ac_build_intrinsic(ctx, "llvm.ctpop.i64", ctx->i64,
2157 (LLVMValueRef []) { src0 }, 1,
2158 AC_FUNC_ATTR_READNONE);
2159
2160 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
2161 break;
2162 case 32:
2163 result = ac_build_intrinsic(ctx, "llvm.ctpop.i32", ctx->i32,
2164 (LLVMValueRef []) { src0 }, 1,
2165 AC_FUNC_ATTR_READNONE);
2166 break;
2167 case 16:
2168 result = ac_build_intrinsic(ctx, "llvm.ctpop.i16", ctx->i16,
2169 (LLVMValueRef []) { src0 }, 1,
2170 AC_FUNC_ATTR_READNONE);
2171 break;
2172 default:
2173 unreachable(!"invalid bitsize");
2174 break;
2175 }
2176
2177 return result;
2178 }
2179
2180 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
2181 LLVMValueRef src0)
2182 {
2183 LLVMValueRef result;
2184 unsigned bitsize;
2185
2186 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2187
2188 switch (bitsize) {
2189 case 32:
2190 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i32", ctx->i32,
2191 (LLVMValueRef []) { src0 }, 1,
2192 AC_FUNC_ATTR_READNONE);
2193 break;
2194 case 16:
2195 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i16", ctx->i16,
2196 (LLVMValueRef []) { src0 }, 1,
2197 AC_FUNC_ATTR_READNONE);
2198 break;
2199 default:
2200 unreachable(!"invalid bitsize");
2201 break;
2202 }
2203
2204 return result;
2205 }
2206
2207 #define AC_EXP_TARGET 0
2208 #define AC_EXP_ENABLED_CHANNELS 1
2209 #define AC_EXP_OUT0 2
2210
2211 enum ac_ir_type {
2212 AC_IR_UNDEF,
2213 AC_IR_CONST,
2214 AC_IR_VALUE,
2215 };
2216
2217 struct ac_vs_exp_chan
2218 {
2219 LLVMValueRef value;
2220 float const_float;
2221 enum ac_ir_type type;
2222 };
2223
2224 struct ac_vs_exp_inst {
2225 unsigned offset;
2226 LLVMValueRef inst;
2227 struct ac_vs_exp_chan chan[4];
2228 };
2229
2230 struct ac_vs_exports {
2231 unsigned num;
2232 struct ac_vs_exp_inst exp[VARYING_SLOT_MAX];
2233 };
2234
2235 /* Return true if the PARAM export has been eliminated. */
2236 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset,
2237 uint32_t num_outputs,
2238 struct ac_vs_exp_inst *exp)
2239 {
2240 unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
2241 bool is_zero[4] = {}, is_one[4] = {};
2242
2243 for (i = 0; i < 4; i++) {
2244 /* It's a constant expression. Undef outputs are eliminated too. */
2245 if (exp->chan[i].type == AC_IR_UNDEF) {
2246 is_zero[i] = true;
2247 is_one[i] = true;
2248 } else if (exp->chan[i].type == AC_IR_CONST) {
2249 if (exp->chan[i].const_float == 0)
2250 is_zero[i] = true;
2251 else if (exp->chan[i].const_float == 1)
2252 is_one[i] = true;
2253 else
2254 return false; /* other constant */
2255 } else
2256 return false;
2257 }
2258
2259 /* Only certain combinations of 0 and 1 can be eliminated. */
2260 if (is_zero[0] && is_zero[1] && is_zero[2])
2261 default_val = is_zero[3] ? 0 : 1;
2262 else if (is_one[0] && is_one[1] && is_one[2])
2263 default_val = is_zero[3] ? 2 : 3;
2264 else
2265 return false;
2266
2267 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
2268 LLVMInstructionEraseFromParent(exp->inst);
2269
2270 /* Change OFFSET to DEFAULT_VAL. */
2271 for (i = 0; i < num_outputs; i++) {
2272 if (vs_output_param_offset[i] == exp->offset) {
2273 vs_output_param_offset[i] =
2274 AC_EXP_PARAM_DEFAULT_VAL_0000 + default_val;
2275 break;
2276 }
2277 }
2278 return true;
2279 }
2280
2281 static bool ac_eliminate_duplicated_output(struct ac_llvm_context *ctx,
2282 uint8_t *vs_output_param_offset,
2283 uint32_t num_outputs,
2284 struct ac_vs_exports *processed,
2285 struct ac_vs_exp_inst *exp)
2286 {
2287 unsigned p, copy_back_channels = 0;
2288
2289 /* See if the output is already in the list of processed outputs.
2290 * The LLVMValueRef comparison relies on SSA.
2291 */
2292 for (p = 0; p < processed->num; p++) {
2293 bool different = false;
2294
2295 for (unsigned j = 0; j < 4; j++) {
2296 struct ac_vs_exp_chan *c1 = &processed->exp[p].chan[j];
2297 struct ac_vs_exp_chan *c2 = &exp->chan[j];
2298
2299 /* Treat undef as a match. */
2300 if (c2->type == AC_IR_UNDEF)
2301 continue;
2302
2303 /* If c1 is undef but c2 isn't, we can copy c2 to c1
2304 * and consider the instruction duplicated.
2305 */
2306 if (c1->type == AC_IR_UNDEF) {
2307 copy_back_channels |= 1 << j;
2308 continue;
2309 }
2310
2311 /* Test whether the channels are not equal. */
2312 if (c1->type != c2->type ||
2313 (c1->type == AC_IR_CONST &&
2314 c1->const_float != c2->const_float) ||
2315 (c1->type == AC_IR_VALUE &&
2316 c1->value != c2->value)) {
2317 different = true;
2318 break;
2319 }
2320 }
2321 if (!different)
2322 break;
2323
2324 copy_back_channels = 0;
2325 }
2326 if (p == processed->num)
2327 return false;
2328
2329 /* If a match was found, but the matching export has undef where the new
2330 * one has a normal value, copy the normal value to the undef channel.
2331 */
2332 struct ac_vs_exp_inst *match = &processed->exp[p];
2333
2334 /* Get current enabled channels mask. */
2335 LLVMValueRef arg = LLVMGetOperand(match->inst, AC_EXP_ENABLED_CHANNELS);
2336 unsigned enabled_channels = LLVMConstIntGetZExtValue(arg);
2337
2338 while (copy_back_channels) {
2339 unsigned chan = u_bit_scan(&copy_back_channels);
2340
2341 assert(match->chan[chan].type == AC_IR_UNDEF);
2342 LLVMSetOperand(match->inst, AC_EXP_OUT0 + chan,
2343 exp->chan[chan].value);
2344 match->chan[chan] = exp->chan[chan];
2345
2346 /* Update number of enabled channels because the original mask
2347 * is not always 0xf.
2348 */
2349 enabled_channels |= (1 << chan);
2350 LLVMSetOperand(match->inst, AC_EXP_ENABLED_CHANNELS,
2351 LLVMConstInt(ctx->i32, enabled_channels, 0));
2352 }
2353
2354 /* The PARAM export is duplicated. Kill it. */
2355 LLVMInstructionEraseFromParent(exp->inst);
2356
2357 /* Change OFFSET to the matching export. */
2358 for (unsigned i = 0; i < num_outputs; i++) {
2359 if (vs_output_param_offset[i] == exp->offset) {
2360 vs_output_param_offset[i] = match->offset;
2361 break;
2362 }
2363 }
2364 return true;
2365 }
2366
2367 void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
2368 LLVMValueRef main_fn,
2369 uint8_t *vs_output_param_offset,
2370 uint32_t num_outputs,
2371 uint8_t *num_param_exports)
2372 {
2373 LLVMBasicBlockRef bb;
2374 bool removed_any = false;
2375 struct ac_vs_exports exports;
2376
2377 exports.num = 0;
2378
2379 /* Process all LLVM instructions. */
2380 bb = LLVMGetFirstBasicBlock(main_fn);
2381 while (bb) {
2382 LLVMValueRef inst = LLVMGetFirstInstruction(bb);
2383
2384 while (inst) {
2385 LLVMValueRef cur = inst;
2386 inst = LLVMGetNextInstruction(inst);
2387 struct ac_vs_exp_inst exp;
2388
2389 if (LLVMGetInstructionOpcode(cur) != LLVMCall)
2390 continue;
2391
2392 LLVMValueRef callee = ac_llvm_get_called_value(cur);
2393
2394 if (!ac_llvm_is_function(callee))
2395 continue;
2396
2397 const char *name = LLVMGetValueName(callee);
2398 unsigned num_args = LLVMCountParams(callee);
2399
2400 /* Check if this is an export instruction. */
2401 if ((num_args != 9 && num_args != 8) ||
2402 (strcmp(name, "llvm.SI.export") &&
2403 strcmp(name, "llvm.amdgcn.exp.f32")))
2404 continue;
2405
2406 LLVMValueRef arg = LLVMGetOperand(cur, AC_EXP_TARGET);
2407 unsigned target = LLVMConstIntGetZExtValue(arg);
2408
2409 if (target < V_008DFC_SQ_EXP_PARAM)
2410 continue;
2411
2412 target -= V_008DFC_SQ_EXP_PARAM;
2413
2414 /* Parse the instruction. */
2415 memset(&exp, 0, sizeof(exp));
2416 exp.offset = target;
2417 exp.inst = cur;
2418
2419 for (unsigned i = 0; i < 4; i++) {
2420 LLVMValueRef v = LLVMGetOperand(cur, AC_EXP_OUT0 + i);
2421
2422 exp.chan[i].value = v;
2423
2424 if (LLVMIsUndef(v)) {
2425 exp.chan[i].type = AC_IR_UNDEF;
2426 } else if (LLVMIsAConstantFP(v)) {
2427 LLVMBool loses_info;
2428 exp.chan[i].type = AC_IR_CONST;
2429 exp.chan[i].const_float =
2430 LLVMConstRealGetDouble(v, &loses_info);
2431 } else {
2432 exp.chan[i].type = AC_IR_VALUE;
2433 }
2434 }
2435
2436 /* Eliminate constant and duplicated PARAM exports. */
2437 if (ac_eliminate_const_output(vs_output_param_offset,
2438 num_outputs, &exp) ||
2439 ac_eliminate_duplicated_output(ctx,
2440 vs_output_param_offset,
2441 num_outputs, &exports,
2442 &exp)) {
2443 removed_any = true;
2444 } else {
2445 exports.exp[exports.num++] = exp;
2446 }
2447 }
2448 bb = LLVMGetNextBasicBlock(bb);
2449 }
2450
2451 /* Remove holes in export memory due to removed PARAM exports.
2452 * This is done by renumbering all PARAM exports.
2453 */
2454 if (removed_any) {
2455 uint8_t old_offset[VARYING_SLOT_MAX];
2456 unsigned out, i;
2457
2458 /* Make a copy of the offsets. We need the old version while
2459 * we are modifying some of them. */
2460 memcpy(old_offset, vs_output_param_offset,
2461 sizeof(old_offset));
2462
2463 for (i = 0; i < exports.num; i++) {
2464 unsigned offset = exports.exp[i].offset;
2465
2466 /* Update vs_output_param_offset. Multiple outputs can
2467 * have the same offset.
2468 */
2469 for (out = 0; out < num_outputs; out++) {
2470 if (old_offset[out] == offset)
2471 vs_output_param_offset[out] = i;
2472 }
2473
2474 /* Change the PARAM offset in the instruction. */
2475 LLVMSetOperand(exports.exp[i].inst, AC_EXP_TARGET,
2476 LLVMConstInt(ctx->i32,
2477 V_008DFC_SQ_EXP_PARAM + i, 0));
2478 }
2479 *num_param_exports = exports.num;
2480 }
2481 }
2482
2483 void ac_init_exec_full_mask(struct ac_llvm_context *ctx)
2484 {
2485 LLVMValueRef full_mask = LLVMConstInt(ctx->i64, ~0ull, 0);
2486 ac_build_intrinsic(ctx,
2487 "llvm.amdgcn.init.exec", ctx->voidt,
2488 &full_mask, 1, AC_FUNC_ATTR_CONVERGENT);
2489 }
2490
2491 void ac_declare_lds_as_pointer(struct ac_llvm_context *ctx)
2492 {
2493 unsigned lds_size = ctx->chip_class >= CIK ? 65536 : 32768;
2494 ctx->lds = LLVMBuildIntToPtr(ctx->builder, ctx->i32_0,
2495 LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), AC_ADDR_SPACE_LDS),
2496 "lds");
2497 }
2498
2499 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
2500 LLVMValueRef dw_addr)
2501 {
2502 return ac_build_load(ctx, ctx->lds, dw_addr);
2503 }
2504
2505 void ac_lds_store(struct ac_llvm_context *ctx,
2506 LLVMValueRef dw_addr,
2507 LLVMValueRef value)
2508 {
2509 value = ac_to_integer(ctx, value);
2510 ac_build_indexed_store(ctx, ctx->lds,
2511 dw_addr, value);
2512 }
2513
2514 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
2515 LLVMTypeRef dst_type,
2516 LLVMValueRef src0)
2517 {
2518 unsigned src0_bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2519 const char *intrin_name;
2520 LLVMTypeRef type;
2521 LLVMValueRef zero;
2522
2523 switch (src0_bitsize) {
2524 case 64:
2525 intrin_name = "llvm.cttz.i64";
2526 type = ctx->i64;
2527 zero = ctx->i64_0;
2528 break;
2529 case 32:
2530 intrin_name = "llvm.cttz.i32";
2531 type = ctx->i32;
2532 zero = ctx->i32_0;
2533 break;
2534 case 16:
2535 intrin_name = "llvm.cttz.i16";
2536 type = ctx->i16;
2537 zero = ctx->i16_0;
2538 break;
2539 default:
2540 unreachable(!"invalid bitsize");
2541 }
2542
2543 LLVMValueRef params[2] = {
2544 src0,
2545
2546 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
2547 * add special code to check for x=0. The reason is that
2548 * the LLVM behavior for x=0 is different from what we
2549 * need here. However, LLVM also assumes that ffs(x) is
2550 * in [0, 31], but GLSL expects that ffs(0) = -1, so
2551 * a conditional assignment to handle 0 is still required.
2552 *
2553 * The hardware already implements the correct behavior.
2554 */
2555 ctx->i1true,
2556 };
2557
2558 LLVMValueRef lsb = ac_build_intrinsic(ctx, intrin_name, type,
2559 params, 2,
2560 AC_FUNC_ATTR_READNONE);
2561
2562 if (src0_bitsize == 64) {
2563 lsb = LLVMBuildTrunc(ctx->builder, lsb, ctx->i32, "");
2564 }
2565
2566 /* TODO: We need an intrinsic to skip this conditional. */
2567 /* Check for zero: */
2568 return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder,
2569 LLVMIntEQ, src0,
2570 zero, ""),
2571 LLVMConstInt(ctx->i32, -1, 0), lsb, "");
2572 }
2573
2574 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type)
2575 {
2576 return LLVMPointerType(LLVMArrayType(elem_type, 0),
2577 AC_ADDR_SPACE_CONST);
2578 }
2579
2580 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type)
2581 {
2582 return LLVMPointerType(LLVMArrayType(elem_type, 0),
2583 AC_ADDR_SPACE_CONST_32BIT);
2584 }
2585
2586 static struct ac_llvm_flow *
2587 get_current_flow(struct ac_llvm_context *ctx)
2588 {
2589 if (ctx->flow_depth > 0)
2590 return &ctx->flow[ctx->flow_depth - 1];
2591 return NULL;
2592 }
2593
2594 static struct ac_llvm_flow *
2595 get_innermost_loop(struct ac_llvm_context *ctx)
2596 {
2597 for (unsigned i = ctx->flow_depth; i > 0; --i) {
2598 if (ctx->flow[i - 1].loop_entry_block)
2599 return &ctx->flow[i - 1];
2600 }
2601 return NULL;
2602 }
2603
2604 static struct ac_llvm_flow *
2605 push_flow(struct ac_llvm_context *ctx)
2606 {
2607 struct ac_llvm_flow *flow;
2608
2609 if (ctx->flow_depth >= ctx->flow_depth_max) {
2610 unsigned new_max = MAX2(ctx->flow_depth << 1,
2611 AC_LLVM_INITIAL_CF_DEPTH);
2612
2613 ctx->flow = realloc(ctx->flow, new_max * sizeof(*ctx->flow));
2614 ctx->flow_depth_max = new_max;
2615 }
2616
2617 flow = &ctx->flow[ctx->flow_depth];
2618 ctx->flow_depth++;
2619
2620 flow->next_block = NULL;
2621 flow->loop_entry_block = NULL;
2622 return flow;
2623 }
2624
2625 static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base,
2626 int label_id)
2627 {
2628 char buf[32];
2629 snprintf(buf, sizeof(buf), "%s%d", base, label_id);
2630 LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
2631 }
2632
2633 /* Append a basic block at the level of the parent flow.
2634 */
2635 static LLVMBasicBlockRef append_basic_block(struct ac_llvm_context *ctx,
2636 const char *name)
2637 {
2638 assert(ctx->flow_depth >= 1);
2639
2640 if (ctx->flow_depth >= 2) {
2641 struct ac_llvm_flow *flow = &ctx->flow[ctx->flow_depth - 2];
2642
2643 return LLVMInsertBasicBlockInContext(ctx->context,
2644 flow->next_block, name);
2645 }
2646
2647 LLVMValueRef main_fn =
2648 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->builder));
2649 return LLVMAppendBasicBlockInContext(ctx->context, main_fn, name);
2650 }
2651
2652 /* Emit a branch to the given default target for the current block if
2653 * applicable -- that is, if the current block does not already contain a
2654 * branch from a break or continue.
2655 */
2656 static void emit_default_branch(LLVMBuilderRef builder,
2657 LLVMBasicBlockRef target)
2658 {
2659 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
2660 LLVMBuildBr(builder, target);
2661 }
2662
2663 void ac_build_bgnloop(struct ac_llvm_context *ctx, int label_id)
2664 {
2665 struct ac_llvm_flow *flow = push_flow(ctx);
2666 flow->loop_entry_block = append_basic_block(ctx, "LOOP");
2667 flow->next_block = append_basic_block(ctx, "ENDLOOP");
2668 set_basicblock_name(flow->loop_entry_block, "loop", label_id);
2669 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
2670 LLVMPositionBuilderAtEnd(ctx->builder, flow->loop_entry_block);
2671 }
2672
2673 void ac_build_break(struct ac_llvm_context *ctx)
2674 {
2675 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
2676 LLVMBuildBr(ctx->builder, flow->next_block);
2677 }
2678
2679 void ac_build_continue(struct ac_llvm_context *ctx)
2680 {
2681 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
2682 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
2683 }
2684
2685 void ac_build_else(struct ac_llvm_context *ctx, int label_id)
2686 {
2687 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
2688 LLVMBasicBlockRef endif_block;
2689
2690 assert(!current_branch->loop_entry_block);
2691
2692 endif_block = append_basic_block(ctx, "ENDIF");
2693 emit_default_branch(ctx->builder, endif_block);
2694
2695 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
2696 set_basicblock_name(current_branch->next_block, "else", label_id);
2697
2698 current_branch->next_block = endif_block;
2699 }
2700
2701 void ac_build_endif(struct ac_llvm_context *ctx, int label_id)
2702 {
2703 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
2704
2705 assert(!current_branch->loop_entry_block);
2706
2707 emit_default_branch(ctx->builder, current_branch->next_block);
2708 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
2709 set_basicblock_name(current_branch->next_block, "endif", label_id);
2710
2711 ctx->flow_depth--;
2712 }
2713
2714 void ac_build_endloop(struct ac_llvm_context *ctx, int label_id)
2715 {
2716 struct ac_llvm_flow *current_loop = get_current_flow(ctx);
2717
2718 assert(current_loop->loop_entry_block);
2719
2720 emit_default_branch(ctx->builder, current_loop->loop_entry_block);
2721
2722 LLVMPositionBuilderAtEnd(ctx->builder, current_loop->next_block);
2723 set_basicblock_name(current_loop->next_block, "endloop", label_id);
2724 ctx->flow_depth--;
2725 }
2726
2727 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id)
2728 {
2729 struct ac_llvm_flow *flow = push_flow(ctx);
2730 LLVMBasicBlockRef if_block;
2731
2732 if_block = append_basic_block(ctx, "IF");
2733 flow->next_block = append_basic_block(ctx, "ELSE");
2734 set_basicblock_name(if_block, "if", label_id);
2735 LLVMBuildCondBr(ctx->builder, cond, if_block, flow->next_block);
2736 LLVMPositionBuilderAtEnd(ctx->builder, if_block);
2737 }
2738
2739 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
2740 int label_id)
2741 {
2742 LLVMValueRef cond = LLVMBuildFCmp(ctx->builder, LLVMRealUNE,
2743 value, ctx->f32_0, "");
2744 ac_build_ifcc(ctx, cond, label_id);
2745 }
2746
2747 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
2748 int label_id)
2749 {
2750 LLVMValueRef cond = LLVMBuildICmp(ctx->builder, LLVMIntNE,
2751 ac_to_integer(ctx, value),
2752 ctx->i32_0, "");
2753 ac_build_ifcc(ctx, cond, label_id);
2754 }
2755
2756 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
2757 const char *name)
2758 {
2759 LLVMBuilderRef builder = ac->builder;
2760 LLVMBasicBlockRef current_block = LLVMGetInsertBlock(builder);
2761 LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
2762 LLVMBasicBlockRef first_block = LLVMGetEntryBasicBlock(function);
2763 LLVMValueRef first_instr = LLVMGetFirstInstruction(first_block);
2764 LLVMBuilderRef first_builder = LLVMCreateBuilderInContext(ac->context);
2765 LLVMValueRef res;
2766
2767 if (first_instr) {
2768 LLVMPositionBuilderBefore(first_builder, first_instr);
2769 } else {
2770 LLVMPositionBuilderAtEnd(first_builder, first_block);
2771 }
2772
2773 res = LLVMBuildAlloca(first_builder, type, name);
2774 LLVMDisposeBuilder(first_builder);
2775 return res;
2776 }
2777
2778 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac,
2779 LLVMTypeRef type, const char *name)
2780 {
2781 LLVMValueRef ptr = ac_build_alloca_undef(ac, type, name);
2782 LLVMBuildStore(ac->builder, LLVMConstNull(type), ptr);
2783 return ptr;
2784 }
2785
2786 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
2787 LLVMTypeRef type)
2788 {
2789 int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
2790 return LLVMBuildBitCast(ctx->builder, ptr,
2791 LLVMPointerType(type, addr_space), "");
2792 }
2793
2794 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
2795 unsigned count)
2796 {
2797 unsigned num_components = ac_get_llvm_num_components(value);
2798 if (count == num_components)
2799 return value;
2800
2801 LLVMValueRef masks[MAX2(count, 2)];
2802 masks[0] = ctx->i32_0;
2803 masks[1] = ctx->i32_1;
2804 for (unsigned i = 2; i < count; i++)
2805 masks[i] = LLVMConstInt(ctx->i32, i, false);
2806
2807 if (count == 1)
2808 return LLVMBuildExtractElement(ctx->builder, value, masks[0],
2809 "");
2810
2811 LLVMValueRef swizzle = LLVMConstVector(masks, count);
2812 return LLVMBuildShuffleVector(ctx->builder, value, value, swizzle, "");
2813 }
2814
2815 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
2816 unsigned rshift, unsigned bitwidth)
2817 {
2818 LLVMValueRef value = param;
2819 if (rshift)
2820 value = LLVMBuildLShr(ctx->builder, value,
2821 LLVMConstInt(ctx->i32, rshift, false), "");
2822
2823 if (rshift + bitwidth < 32) {
2824 unsigned mask = (1 << bitwidth) - 1;
2825 value = LLVMBuildAnd(ctx->builder, value,
2826 LLVMConstInt(ctx->i32, mask, false), "");
2827 }
2828 return value;
2829 }
2830
2831 /* Adjust the sample index according to FMASK.
2832 *
2833 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
2834 * which is the identity mapping. Each nibble says which physical sample
2835 * should be fetched to get that sample.
2836 *
2837 * For example, 0x11111100 means there are only 2 samples stored and
2838 * the second sample covers 3/4 of the pixel. When reading samples 0
2839 * and 1, return physical sample 0 (determined by the first two 0s
2840 * in FMASK), otherwise return physical sample 1.
2841 *
2842 * The sample index should be adjusted as follows:
2843 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
2844 */
2845 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
2846 LLVMValueRef *addr, bool is_array_tex)
2847 {
2848 struct ac_image_args fmask_load = {};
2849 fmask_load.opcode = ac_image_load;
2850 fmask_load.resource = fmask;
2851 fmask_load.dmask = 0xf;
2852 fmask_load.dim = is_array_tex ? ac_image_2darray : ac_image_2d;
2853
2854 fmask_load.coords[0] = addr[0];
2855 fmask_load.coords[1] = addr[1];
2856 if (is_array_tex)
2857 fmask_load.coords[2] = addr[2];
2858
2859 LLVMValueRef fmask_value = ac_build_image_opcode(ac, &fmask_load);
2860 fmask_value = LLVMBuildExtractElement(ac->builder, fmask_value,
2861 ac->i32_0, "");
2862
2863 /* Apply the formula. */
2864 unsigned sample_chan = is_array_tex ? 3 : 2;
2865 LLVMValueRef final_sample;
2866 final_sample = LLVMBuildMul(ac->builder, addr[sample_chan],
2867 LLVMConstInt(ac->i32, 4, 0), "");
2868 final_sample = LLVMBuildLShr(ac->builder, fmask_value, final_sample, "");
2869 /* Mask the sample index by 0x7, because 0x8 means an unknown value
2870 * with EQAA, so those will map to 0. */
2871 final_sample = LLVMBuildAnd(ac->builder, final_sample,
2872 LLVMConstInt(ac->i32, 0x7, 0), "");
2873
2874 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
2875 * resource descriptor is 0 (invalid).
2876 */
2877 LLVMValueRef tmp;
2878 tmp = LLVMBuildBitCast(ac->builder, fmask, ac->v8i32, "");
2879 tmp = LLVMBuildExtractElement(ac->builder, tmp, ac->i32_1, "");
2880 tmp = LLVMBuildICmp(ac->builder, LLVMIntNE, tmp, ac->i32_0, "");
2881
2882 /* Replace the MSAA sample index. */
2883 addr[sample_chan] = LLVMBuildSelect(ac->builder, tmp, final_sample,
2884 addr[sample_chan], "");
2885 }
2886
2887 static LLVMValueRef
2888 _ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
2889 {
2890 ac_build_optimization_barrier(ctx, &src);
2891 return ac_build_intrinsic(ctx,
2892 lane == NULL ? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
2893 LLVMTypeOf(src), (LLVMValueRef []) {
2894 src, lane },
2895 lane == NULL ? 1 : 2,
2896 AC_FUNC_ATTR_READNONE |
2897 AC_FUNC_ATTR_CONVERGENT);
2898 }
2899
2900 /**
2901 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
2902 * @param ctx
2903 * @param src
2904 * @param lane - id of the lane or NULL for the first active lane
2905 * @return value of the lane
2906 */
2907 LLVMValueRef
2908 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
2909 {
2910 LLVMTypeRef src_type = LLVMTypeOf(src);
2911 src = ac_to_integer(ctx, src);
2912 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
2913 LLVMValueRef ret;
2914
2915 if (bits == 32) {
2916 ret = _ac_build_readlane(ctx, src, lane);
2917 } else {
2918 assert(bits % 32 == 0);
2919 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
2920 LLVMValueRef src_vector =
2921 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
2922 ret = LLVMGetUndef(vec_type);
2923 for (unsigned i = 0; i < bits / 32; i++) {
2924 src = LLVMBuildExtractElement(ctx->builder, src_vector,
2925 LLVMConstInt(ctx->i32, i, 0), "");
2926 LLVMValueRef ret_comp = _ac_build_readlane(ctx, src, lane);
2927 ret = LLVMBuildInsertElement(ctx->builder, ret, ret_comp,
2928 LLVMConstInt(ctx->i32, i, 0), "");
2929 }
2930 }
2931 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
2932 }
2933
2934 LLVMValueRef
2935 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane)
2936 {
2937 /* TODO: Use the actual instruction when LLVM adds an intrinsic for it.
2938 */
2939 LLVMValueRef pred = LLVMBuildICmp(ctx->builder, LLVMIntEQ, lane,
2940 ac_get_thread_id(ctx), "");
2941 return LLVMBuildSelect(ctx->builder, pred, value, src, "");
2942 }
2943
2944 LLVMValueRef
2945 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask)
2946 {
2947 LLVMValueRef mask_vec = LLVMBuildBitCast(ctx->builder, mask,
2948 LLVMVectorType(ctx->i32, 2),
2949 "");
2950 LLVMValueRef mask_lo = LLVMBuildExtractElement(ctx->builder, mask_vec,
2951 ctx->i32_0, "");
2952 LLVMValueRef mask_hi = LLVMBuildExtractElement(ctx->builder, mask_vec,
2953 ctx->i32_1, "");
2954 LLVMValueRef val =
2955 ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
2956 (LLVMValueRef []) { mask_lo, ctx->i32_0 },
2957 2, AC_FUNC_ATTR_READNONE);
2958 val = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi", ctx->i32,
2959 (LLVMValueRef []) { mask_hi, val },
2960 2, AC_FUNC_ATTR_READNONE);
2961 return val;
2962 }
2963
2964 enum dpp_ctrl {
2965 _dpp_quad_perm = 0x000,
2966 _dpp_row_sl = 0x100,
2967 _dpp_row_sr = 0x110,
2968 _dpp_row_rr = 0x120,
2969 dpp_wf_sl1 = 0x130,
2970 dpp_wf_rl1 = 0x134,
2971 dpp_wf_sr1 = 0x138,
2972 dpp_wf_rr1 = 0x13C,
2973 dpp_row_mirror = 0x140,
2974 dpp_row_half_mirror = 0x141,
2975 dpp_row_bcast15 = 0x142,
2976 dpp_row_bcast31 = 0x143
2977 };
2978
2979 static inline enum dpp_ctrl
2980 dpp_quad_perm(unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
2981 {
2982 assert(lane0 < 4 && lane1 < 4 && lane2 < 4 && lane3 < 4);
2983 return _dpp_quad_perm | lane0 | (lane1 << 2) | (lane2 << 4) | (lane3 << 6);
2984 }
2985
2986 static inline enum dpp_ctrl
2987 dpp_row_sl(unsigned amount)
2988 {
2989 assert(amount > 0 && amount < 16);
2990 return _dpp_row_sl | amount;
2991 }
2992
2993 static inline enum dpp_ctrl
2994 dpp_row_sr(unsigned amount)
2995 {
2996 assert(amount > 0 && amount < 16);
2997 return _dpp_row_sr | amount;
2998 }
2999
3000 static LLVMValueRef
3001 _ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3002 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3003 bool bound_ctrl)
3004 {
3005 return ac_build_intrinsic(ctx, "llvm.amdgcn.update.dpp.i32",
3006 LLVMTypeOf(old),
3007 (LLVMValueRef[]) {
3008 old, src,
3009 LLVMConstInt(ctx->i32, dpp_ctrl, 0),
3010 LLVMConstInt(ctx->i32, row_mask, 0),
3011 LLVMConstInt(ctx->i32, bank_mask, 0),
3012 LLVMConstInt(ctx->i1, bound_ctrl, 0) },
3013 6, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3014 }
3015
3016 static LLVMValueRef
3017 ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3018 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3019 bool bound_ctrl)
3020 {
3021 LLVMTypeRef src_type = LLVMTypeOf(src);
3022 src = ac_to_integer(ctx, src);
3023 old = ac_to_integer(ctx, old);
3024 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3025 LLVMValueRef ret;
3026 if (bits == 32) {
3027 ret = _ac_build_dpp(ctx, old, src, dpp_ctrl, row_mask,
3028 bank_mask, bound_ctrl);
3029 } else {
3030 assert(bits % 32 == 0);
3031 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3032 LLVMValueRef src_vector =
3033 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3034 LLVMValueRef old_vector =
3035 LLVMBuildBitCast(ctx->builder, old, vec_type, "");
3036 ret = LLVMGetUndef(vec_type);
3037 for (unsigned i = 0; i < bits / 32; i++) {
3038 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3039 LLVMConstInt(ctx->i32, i,
3040 0), "");
3041 old = LLVMBuildExtractElement(ctx->builder, old_vector,
3042 LLVMConstInt(ctx->i32, i,
3043 0), "");
3044 LLVMValueRef ret_comp = _ac_build_dpp(ctx, old, src,
3045 dpp_ctrl,
3046 row_mask,
3047 bank_mask,
3048 bound_ctrl);
3049 ret = LLVMBuildInsertElement(ctx->builder, ret,
3050 ret_comp,
3051 LLVMConstInt(ctx->i32, i,
3052 0), "");
3053 }
3054 }
3055 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3056 }
3057
3058 static inline unsigned
3059 ds_pattern_bitmode(unsigned and_mask, unsigned or_mask, unsigned xor_mask)
3060 {
3061 assert(and_mask < 32 && or_mask < 32 && xor_mask < 32);
3062 return and_mask | (or_mask << 5) | (xor_mask << 10);
3063 }
3064
3065 static LLVMValueRef
3066 _ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
3067 {
3068 return ac_build_intrinsic(ctx, "llvm.amdgcn.ds.swizzle",
3069 LLVMTypeOf(src), (LLVMValueRef []) {
3070 src, LLVMConstInt(ctx->i32, mask, 0) },
3071 2, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3072 }
3073
3074 LLVMValueRef
3075 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
3076 {
3077 LLVMTypeRef src_type = LLVMTypeOf(src);
3078 src = ac_to_integer(ctx, src);
3079 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3080 LLVMValueRef ret;
3081 if (bits == 32) {
3082 ret = _ac_build_ds_swizzle(ctx, src, mask);
3083 } else {
3084 assert(bits % 32 == 0);
3085 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3086 LLVMValueRef src_vector =
3087 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3088 ret = LLVMGetUndef(vec_type);
3089 for (unsigned i = 0; i < bits / 32; i++) {
3090 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3091 LLVMConstInt(ctx->i32, i,
3092 0), "");
3093 LLVMValueRef ret_comp = _ac_build_ds_swizzle(ctx, src,
3094 mask);
3095 ret = LLVMBuildInsertElement(ctx->builder, ret,
3096 ret_comp,
3097 LLVMConstInt(ctx->i32, i,
3098 0), "");
3099 }
3100 }
3101 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3102 }
3103
3104 static LLVMValueRef
3105 ac_build_wwm(struct ac_llvm_context *ctx, LLVMValueRef src)
3106 {
3107 char name[32], type[8];
3108 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
3109 snprintf(name, sizeof(name), "llvm.amdgcn.wwm.%s", type);
3110 return ac_build_intrinsic(ctx, name, LLVMTypeOf(src),
3111 (LLVMValueRef []) { src }, 1,
3112 AC_FUNC_ATTR_READNONE);
3113 }
3114
3115 static LLVMValueRef
3116 ac_build_set_inactive(struct ac_llvm_context *ctx, LLVMValueRef src,
3117 LLVMValueRef inactive)
3118 {
3119 char name[33], type[8];
3120 LLVMTypeRef src_type = LLVMTypeOf(src);
3121 src = ac_to_integer(ctx, src);
3122 inactive = ac_to_integer(ctx, inactive);
3123 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
3124 snprintf(name, sizeof(name), "llvm.amdgcn.set.inactive.%s", type);
3125 LLVMValueRef ret =
3126 ac_build_intrinsic(ctx, name,
3127 LLVMTypeOf(src), (LLVMValueRef []) {
3128 src, inactive }, 2,
3129 AC_FUNC_ATTR_READNONE |
3130 AC_FUNC_ATTR_CONVERGENT);
3131 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3132 }
3133
3134 static LLVMValueRef
3135 get_reduction_identity(struct ac_llvm_context *ctx, nir_op op, unsigned type_size)
3136 {
3137 if (type_size == 4) {
3138 switch (op) {
3139 case nir_op_iadd: return ctx->i32_0;
3140 case nir_op_fadd: return ctx->f32_0;
3141 case nir_op_imul: return ctx->i32_1;
3142 case nir_op_fmul: return ctx->f32_1;
3143 case nir_op_imin: return LLVMConstInt(ctx->i32, INT32_MAX, 0);
3144 case nir_op_umin: return LLVMConstInt(ctx->i32, UINT32_MAX, 0);
3145 case nir_op_fmin: return LLVMConstReal(ctx->f32, INFINITY);
3146 case nir_op_imax: return LLVMConstInt(ctx->i32, INT32_MIN, 0);
3147 case nir_op_umax: return ctx->i32_0;
3148 case nir_op_fmax: return LLVMConstReal(ctx->f32, -INFINITY);
3149 case nir_op_iand: return LLVMConstInt(ctx->i32, -1, 0);
3150 case nir_op_ior: return ctx->i32_0;
3151 case nir_op_ixor: return ctx->i32_0;
3152 default:
3153 unreachable("bad reduction intrinsic");
3154 }
3155 } else { /* type_size == 64bit */
3156 switch (op) {
3157 case nir_op_iadd: return ctx->i64_0;
3158 case nir_op_fadd: return ctx->f64_0;
3159 case nir_op_imul: return ctx->i64_1;
3160 case nir_op_fmul: return ctx->f64_1;
3161 case nir_op_imin: return LLVMConstInt(ctx->i64, INT64_MAX, 0);
3162 case nir_op_umin: return LLVMConstInt(ctx->i64, UINT64_MAX, 0);
3163 case nir_op_fmin: return LLVMConstReal(ctx->f64, INFINITY);
3164 case nir_op_imax: return LLVMConstInt(ctx->i64, INT64_MIN, 0);
3165 case nir_op_umax: return ctx->i64_0;
3166 case nir_op_fmax: return LLVMConstReal(ctx->f64, -INFINITY);
3167 case nir_op_iand: return LLVMConstInt(ctx->i64, -1, 0);
3168 case nir_op_ior: return ctx->i64_0;
3169 case nir_op_ixor: return ctx->i64_0;
3170 default:
3171 unreachable("bad reduction intrinsic");
3172 }
3173 }
3174 }
3175
3176 static LLVMValueRef
3177 ac_build_alu_op(struct ac_llvm_context *ctx, LLVMValueRef lhs, LLVMValueRef rhs, nir_op op)
3178 {
3179 bool _64bit = ac_get_type_size(LLVMTypeOf(lhs)) == 8;
3180 switch (op) {
3181 case nir_op_iadd: return LLVMBuildAdd(ctx->builder, lhs, rhs, "");
3182 case nir_op_fadd: return LLVMBuildFAdd(ctx->builder, lhs, rhs, "");
3183 case nir_op_imul: return LLVMBuildMul(ctx->builder, lhs, rhs, "");
3184 case nir_op_fmul: return LLVMBuildFMul(ctx->builder, lhs, rhs, "");
3185 case nir_op_imin: return LLVMBuildSelect(ctx->builder,
3186 LLVMBuildICmp(ctx->builder, LLVMIntSLT, lhs, rhs, ""),
3187 lhs, rhs, "");
3188 case nir_op_umin: return LLVMBuildSelect(ctx->builder,
3189 LLVMBuildICmp(ctx->builder, LLVMIntULT, lhs, rhs, ""),
3190 lhs, rhs, "");
3191 case nir_op_fmin: return ac_build_intrinsic(ctx,
3192 _64bit ? "llvm.minnum.f64" : "llvm.minnum.f32",
3193 _64bit ? ctx->f64 : ctx->f32,
3194 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
3195 case nir_op_imax: return LLVMBuildSelect(ctx->builder,
3196 LLVMBuildICmp(ctx->builder, LLVMIntSGT, lhs, rhs, ""),
3197 lhs, rhs, "");
3198 case nir_op_umax: return LLVMBuildSelect(ctx->builder,
3199 LLVMBuildICmp(ctx->builder, LLVMIntUGT, lhs, rhs, ""),
3200 lhs, rhs, "");
3201 case nir_op_fmax: return ac_build_intrinsic(ctx,
3202 _64bit ? "llvm.maxnum.f64" : "llvm.maxnum.f32",
3203 _64bit ? ctx->f64 : ctx->f32,
3204 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
3205 case nir_op_iand: return LLVMBuildAnd(ctx->builder, lhs, rhs, "");
3206 case nir_op_ior: return LLVMBuildOr(ctx->builder, lhs, rhs, "");
3207 case nir_op_ixor: return LLVMBuildXor(ctx->builder, lhs, rhs, "");
3208 default:
3209 unreachable("bad reduction intrinsic");
3210 }
3211 }
3212
3213 /**
3214 * \param maxprefix specifies that the result only needs to be correct for a
3215 * prefix of this many threads
3216 *
3217 * TODO: add inclusive and excluse scan functions for SI chip class.
3218 */
3219 static LLVMValueRef
3220 ac_build_scan(struct ac_llvm_context *ctx, nir_op op, LLVMValueRef src, LLVMValueRef identity,
3221 unsigned maxprefix)
3222 {
3223 LLVMValueRef result, tmp;
3224 result = src;
3225 if (maxprefix <= 1)
3226 return result;
3227 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
3228 result = ac_build_alu_op(ctx, result, tmp, op);
3229 if (maxprefix <= 2)
3230 return result;
3231 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(2), 0xf, 0xf, false);
3232 result = ac_build_alu_op(ctx, result, tmp, op);
3233 if (maxprefix <= 3)
3234 return result;
3235 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(3), 0xf, 0xf, false);
3236 result = ac_build_alu_op(ctx, result, tmp, op);
3237 if (maxprefix <= 4)
3238 return result;
3239 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(4), 0xf, 0xe, false);
3240 result = ac_build_alu_op(ctx, result, tmp, op);
3241 if (maxprefix <= 8)
3242 return result;
3243 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(8), 0xf, 0xc, false);
3244 result = ac_build_alu_op(ctx, result, tmp, op);
3245 if (maxprefix <= 16)
3246 return result;
3247 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
3248 result = ac_build_alu_op(ctx, result, tmp, op);
3249 if (maxprefix <= 32)
3250 return result;
3251 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
3252 result = ac_build_alu_op(ctx, result, tmp, op);
3253 return result;
3254 }
3255
3256 LLVMValueRef
3257 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
3258 {
3259 LLVMValueRef result;
3260
3261 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
3262 LLVMBuilderRef builder = ctx->builder;
3263 src = LLVMBuildZExt(builder, src, ctx->i32, "");
3264 result = ac_build_ballot(ctx, src);
3265 result = ac_build_mbcnt(ctx, result);
3266 result = LLVMBuildAdd(builder, result, src, "");
3267 return result;
3268 }
3269
3270 ac_build_optimization_barrier(ctx, &src);
3271
3272 LLVMValueRef identity =
3273 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
3274 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
3275 LLVMTypeOf(identity), "");
3276 result = ac_build_scan(ctx, op, result, identity, 64);
3277
3278 return ac_build_wwm(ctx, result);
3279 }
3280
3281 LLVMValueRef
3282 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
3283 {
3284 LLVMValueRef result;
3285
3286 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
3287 LLVMBuilderRef builder = ctx->builder;
3288 src = LLVMBuildZExt(builder, src, ctx->i32, "");
3289 result = ac_build_ballot(ctx, src);
3290 result = ac_build_mbcnt(ctx, result);
3291 return result;
3292 }
3293
3294 ac_build_optimization_barrier(ctx, &src);
3295
3296 LLVMValueRef identity =
3297 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
3298 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
3299 LLVMTypeOf(identity), "");
3300 result = ac_build_dpp(ctx, identity, result, dpp_wf_sr1, 0xf, 0xf, false);
3301 result = ac_build_scan(ctx, op, result, identity, 64);
3302
3303 return ac_build_wwm(ctx, result);
3304 }
3305
3306 LLVMValueRef
3307 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size)
3308 {
3309 if (cluster_size == 1) return src;
3310 ac_build_optimization_barrier(ctx, &src);
3311 LLVMValueRef result, swap;
3312 LLVMValueRef identity = get_reduction_identity(ctx, op,
3313 ac_get_type_size(LLVMTypeOf(src)));
3314 result = LLVMBuildBitCast(ctx->builder,
3315 ac_build_set_inactive(ctx, src, identity),
3316 LLVMTypeOf(identity), "");
3317 swap = ac_build_quad_swizzle(ctx, result, 1, 0, 3, 2);
3318 result = ac_build_alu_op(ctx, result, swap, op);
3319 if (cluster_size == 2) return ac_build_wwm(ctx, result);
3320
3321 swap = ac_build_quad_swizzle(ctx, result, 2, 3, 0, 1);
3322 result = ac_build_alu_op(ctx, result, swap, op);
3323 if (cluster_size == 4) return ac_build_wwm(ctx, result);
3324
3325 if (ctx->chip_class >= VI)
3326 swap = ac_build_dpp(ctx, identity, result, dpp_row_half_mirror, 0xf, 0xf, false);
3327 else
3328 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x04));
3329 result = ac_build_alu_op(ctx, result, swap, op);
3330 if (cluster_size == 8) return ac_build_wwm(ctx, result);
3331
3332 if (ctx->chip_class >= VI)
3333 swap = ac_build_dpp(ctx, identity, result, dpp_row_mirror, 0xf, 0xf, false);
3334 else
3335 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x08));
3336 result = ac_build_alu_op(ctx, result, swap, op);
3337 if (cluster_size == 16) return ac_build_wwm(ctx, result);
3338
3339 if (ctx->chip_class >= VI && cluster_size != 32)
3340 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
3341 else
3342 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x10));
3343 result = ac_build_alu_op(ctx, result, swap, op);
3344 if (cluster_size == 32) return ac_build_wwm(ctx, result);
3345
3346 if (ctx->chip_class >= VI) {
3347 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
3348 result = ac_build_alu_op(ctx, result, swap, op);
3349 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 63, 0));
3350 return ac_build_wwm(ctx, result);
3351 } else {
3352 swap = ac_build_readlane(ctx, result, ctx->i32_0);
3353 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 32, 0));
3354 result = ac_build_alu_op(ctx, result, swap, op);
3355 return ac_build_wwm(ctx, result);
3356 }
3357 }
3358
3359 /**
3360 * "Top half" of a scan that reduces per-wave values across an entire
3361 * workgroup.
3362 *
3363 * The source value must be present in the highest lane of the wave, and the
3364 * highest lane must be live.
3365 */
3366 void
3367 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3368 {
3369 if (ws->maxwaves <= 1)
3370 return;
3371
3372 const LLVMValueRef i32_63 = LLVMConstInt(ctx->i32, 63, false);
3373 LLVMBuilderRef builder = ctx->builder;
3374 LLVMValueRef tid = ac_get_thread_id(ctx);
3375 LLVMValueRef tmp;
3376
3377 tmp = LLVMBuildICmp(builder, LLVMIntEQ, tid, i32_63, "");
3378 ac_build_ifcc(ctx, tmp, 1000);
3379 LLVMBuildStore(builder, ws->src, LLVMBuildGEP(builder, ws->scratch, &ws->waveidx, 1, ""));
3380 ac_build_endif(ctx, 1000);
3381 }
3382
3383 /**
3384 * "Bottom half" of a scan that reduces per-wave values across an entire
3385 * workgroup.
3386 *
3387 * The caller must place a barrier between the top and bottom halves.
3388 */
3389 void
3390 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3391 {
3392 const LLVMTypeRef type = LLVMTypeOf(ws->src);
3393 const LLVMValueRef identity =
3394 get_reduction_identity(ctx, ws->op, ac_get_type_size(type));
3395
3396 if (ws->maxwaves <= 1) {
3397 ws->result_reduce = ws->src;
3398 ws->result_inclusive = ws->src;
3399 ws->result_exclusive = identity;
3400 return;
3401 }
3402 assert(ws->maxwaves <= 32);
3403
3404 LLVMBuilderRef builder = ctx->builder;
3405 LLVMValueRef tid = ac_get_thread_id(ctx);
3406 LLVMBasicBlockRef bbs[2];
3407 LLVMValueRef phivalues_scan[2];
3408 LLVMValueRef tmp, tmp2;
3409
3410 bbs[0] = LLVMGetInsertBlock(builder);
3411 phivalues_scan[0] = LLVMGetUndef(type);
3412
3413 if (ws->enable_reduce)
3414 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->numwaves, "");
3415 else if (ws->enable_inclusive)
3416 tmp = LLVMBuildICmp(builder, LLVMIntULE, tid, ws->waveidx, "");
3417 else
3418 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->waveidx, "");
3419 ac_build_ifcc(ctx, tmp, 1001);
3420 {
3421 tmp = LLVMBuildLoad(builder, LLVMBuildGEP(builder, ws->scratch, &tid, 1, ""), "");
3422
3423 ac_build_optimization_barrier(ctx, &tmp);
3424
3425 bbs[1] = LLVMGetInsertBlock(builder);
3426 phivalues_scan[1] = ac_build_scan(ctx, ws->op, tmp, identity, ws->maxwaves);
3427 }
3428 ac_build_endif(ctx, 1001);
3429
3430 const LLVMValueRef scan = ac_build_phi(ctx, type, 2, phivalues_scan, bbs);
3431
3432 if (ws->enable_reduce) {
3433 tmp = LLVMBuildSub(builder, ws->numwaves, ctx->i32_1, "");
3434 ws->result_reduce = ac_build_readlane(ctx, scan, tmp);
3435 }
3436 if (ws->enable_inclusive)
3437 ws->result_inclusive = ac_build_readlane(ctx, scan, ws->waveidx);
3438 if (ws->enable_exclusive) {
3439 tmp = LLVMBuildSub(builder, ws->waveidx, ctx->i32_1, "");
3440 tmp = ac_build_readlane(ctx, scan, tmp);
3441 tmp2 = LLVMBuildICmp(builder, LLVMIntEQ, ws->waveidx, ctx->i32_0, "");
3442 ws->result_exclusive = LLVMBuildSelect(builder, tmp2, identity, tmp, "");
3443 }
3444 }
3445
3446 /**
3447 * Inclusive scan of a per-wave value across an entire workgroup.
3448 *
3449 * This implies an s_barrier instruction.
3450 *
3451 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
3452 * of the workgroup are live. (This requirement cannot easily be relaxed in a
3453 * useful manner because of the barrier in the algorithm.)
3454 */
3455 void
3456 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3457 {
3458 ac_build_wg_wavescan_top(ctx, ws);
3459 ac_build_s_barrier(ctx);
3460 ac_build_wg_wavescan_bottom(ctx, ws);
3461 }
3462
3463 /**
3464 * "Top half" of a scan that reduces per-thread values across an entire
3465 * workgroup.
3466 *
3467 * All lanes must be active when this code runs.
3468 */
3469 void
3470 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3471 {
3472 if (ws->enable_exclusive) {
3473 ws->extra = ac_build_exclusive_scan(ctx, ws->src, ws->op);
3474 if (LLVMTypeOf(ws->src) == ctx->i1 && ws->op == nir_op_iadd)
3475 ws->src = LLVMBuildZExt(ctx->builder, ws->src, ctx->i32, "");
3476 ws->src = ac_build_alu_op(ctx, ws->extra, ws->src, ws->op);
3477 } else {
3478 ws->src = ac_build_inclusive_scan(ctx, ws->src, ws->op);
3479 }
3480
3481 bool enable_inclusive = ws->enable_inclusive;
3482 bool enable_exclusive = ws->enable_exclusive;
3483 ws->enable_inclusive = false;
3484 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
3485 ac_build_wg_wavescan_top(ctx, ws);
3486 ws->enable_inclusive = enable_inclusive;
3487 ws->enable_exclusive = enable_exclusive;
3488 }
3489
3490 /**
3491 * "Bottom half" of a scan that reduces per-thread values across an entire
3492 * workgroup.
3493 *
3494 * The caller must place a barrier between the top and bottom halves.
3495 */
3496 void
3497 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3498 {
3499 bool enable_inclusive = ws->enable_inclusive;
3500 bool enable_exclusive = ws->enable_exclusive;
3501 ws->enable_inclusive = false;
3502 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
3503 ac_build_wg_wavescan_bottom(ctx, ws);
3504 ws->enable_inclusive = enable_inclusive;
3505 ws->enable_exclusive = enable_exclusive;
3506
3507 /* ws->result_reduce is already the correct value */
3508 if (ws->enable_inclusive)
3509 ws->result_inclusive = ac_build_alu_op(ctx, ws->result_exclusive, ws->src, ws->op);
3510 if (ws->enable_exclusive)
3511 ws->result_exclusive = ac_build_alu_op(ctx, ws->result_exclusive, ws->extra, ws->op);
3512 }
3513
3514 /**
3515 * A scan that reduces per-thread values across an entire workgroup.
3516 *
3517 * The caller must ensure that all lanes are active when this code runs
3518 * (WWM is insufficient!), because there is an implied barrier.
3519 */
3520 void
3521 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3522 {
3523 ac_build_wg_scan_top(ctx, ws);
3524 ac_build_s_barrier(ctx);
3525 ac_build_wg_scan_bottom(ctx, ws);
3526 }
3527
3528 LLVMValueRef
3529 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
3530 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
3531 {
3532 unsigned mask = dpp_quad_perm(lane0, lane1, lane2, lane3);
3533 if (ctx->chip_class >= VI) {
3534 return ac_build_dpp(ctx, src, src, mask, 0xf, 0xf, false);
3535 } else {
3536 return ac_build_ds_swizzle(ctx, src, (1 << 15) | mask);
3537 }
3538 }
3539
3540 LLVMValueRef
3541 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index)
3542 {
3543 index = LLVMBuildMul(ctx->builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
3544 return ac_build_intrinsic(ctx,
3545 "llvm.amdgcn.ds.bpermute", ctx->i32,
3546 (LLVMValueRef []) {index, src}, 2,
3547 AC_FUNC_ATTR_READNONE |
3548 AC_FUNC_ATTR_CONVERGENT);
3549 }