projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
ac: use new LLVM 8 intrinsic when loading 16-bit values
[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 = LLVMTypeOf(num) == ctx->f64 ? ctx->f64_1 : ctx->f32_1;
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_mov(struct ac_llvm_context *ctx,
924 LLVMValueRef parameter,
925 LLVMValueRef llvm_chan,
926 LLVMValueRef attr_number,
927 LLVMValueRef params)
928 {
929 LLVMValueRef args[4];
930
931 args[0] = parameter;
932 args[1] = llvm_chan;
933 args[2] = attr_number;
934 args[3] = params;
935
936 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.mov",
937 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
938 }
939
940 LLVMValueRef
941 ac_build_gep_ptr(struct ac_llvm_context *ctx,
942 LLVMValueRef base_ptr,
943 LLVMValueRef index)
944 {
945 return LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
946 }
947
948 LLVMValueRef
949 ac_build_gep0(struct ac_llvm_context *ctx,
950 LLVMValueRef base_ptr,
951 LLVMValueRef index)
952 {
953 LLVMValueRef indices[2] = {
954 ctx->i32_0,
955 index,
956 };
957 return LLVMBuildGEP(ctx->builder, base_ptr, indices, 2, "");
958 }
959
960 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
961 LLVMValueRef index)
962 {
963 return LLVMBuildPointerCast(ctx->builder,
964 ac_build_gep0(ctx, ptr, index),
965 LLVMTypeOf(ptr), "");
966 }
967
968 void
969 ac_build_indexed_store(struct ac_llvm_context *ctx,
970 LLVMValueRef base_ptr, LLVMValueRef index,
971 LLVMValueRef value)
972 {
973 LLVMBuildStore(ctx->builder, value,
974 ac_build_gep0(ctx, base_ptr, index));
975 }
976
977 /**
978 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
979 * It's equivalent to doing a load from &base_ptr[index].
980 *
981 * \param base_ptr Where the array starts.
982 * \param index The element index into the array.
983 * \param uniform Whether the base_ptr and index can be assumed to be
984 * dynamically uniform (i.e. load to an SGPR)
985 * \param invariant Whether the load is invariant (no other opcodes affect it)
986 * \param no_unsigned_wraparound
987 * For all possible re-associations and re-distributions of an expression
988 * "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
989 * without inbounds in base_ptr), this parameter is true if "addr + offset"
990 * does not result in an unsigned integer wraparound. This is used for
991 * optimal code generation of 32-bit pointer arithmetic.
992 *
993 * For example, a 32-bit immediate offset that causes a 32-bit unsigned
994 * integer wraparound can't be an imm offset in s_load_dword, because
995 * the instruction performs "addr + offset" in 64 bits.
996 *
997 * Expected usage for bindless textures by chaining GEPs:
998 * // possible unsigned wraparound, don't use InBounds:
999 * ptr1 = LLVMBuildGEP(base_ptr, index);
1000 * image = load(ptr1); // becomes "s_load ptr1, 0"
1001 *
1002 * ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1003 * sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1004 */
1005 static LLVMValueRef
1006 ac_build_load_custom(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1007 LLVMValueRef index, bool uniform, bool invariant,
1008 bool no_unsigned_wraparound)
1009 {
1010 LLVMValueRef pointer, result;
1011 LLVMValueRef indices[2] = {ctx->i32_0, index};
1012
1013 if (no_unsigned_wraparound &&
1014 LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr)) == AC_ADDR_SPACE_CONST_32BIT)
1015 pointer = LLVMBuildInBoundsGEP(ctx->builder, base_ptr, indices, 2, "");
1016 else
1017 pointer = LLVMBuildGEP(ctx->builder, base_ptr, indices, 2, "");
1018
1019 if (uniform)
1020 LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
1021 result = LLVMBuildLoad(ctx->builder, pointer, "");
1022 if (invariant)
1023 LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
1024 return result;
1025 }
1026
1027 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1028 LLVMValueRef index)
1029 {
1030 return ac_build_load_custom(ctx, base_ptr, index, false, false, false);
1031 }
1032
1033 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
1034 LLVMValueRef base_ptr, LLVMValueRef index)
1035 {
1036 return ac_build_load_custom(ctx, base_ptr, index, false, true, false);
1037 }
1038
1039 /* This assumes that there is no unsigned integer wraparound during the address
1040 * computation, excluding all GEPs within base_ptr. */
1041 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
1042 LLVMValueRef base_ptr, LLVMValueRef index)
1043 {
1044 return ac_build_load_custom(ctx, base_ptr, index, true, true, true);
1045 }
1046
1047 /* See ac_build_load_custom() documentation. */
1048 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
1049 LLVMValueRef base_ptr, LLVMValueRef index)
1050 {
1051 return ac_build_load_custom(ctx, base_ptr, index, true, true, false);
1052 }
1053
1054 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1055 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1056 * or v4i32 (num_channels=3,4).
1057 */
1058 void
1059 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
1060 LLVMValueRef rsrc,
1061 LLVMValueRef vdata,
1062 unsigned num_channels,
1063 LLVMValueRef voffset,
1064 LLVMValueRef soffset,
1065 unsigned inst_offset,
1066 bool glc,
1067 bool slc,
1068 bool writeonly_memory,
1069 bool swizzle_enable_hint)
1070 {
1071 /* Split 3 channel stores, becase LLVM doesn't support 3-channel
1072 * intrinsics. */
1073 if (num_channels == 3) {
1074 LLVMValueRef v[3], v01;
1075
1076 for (int i = 0; i < 3; i++) {
1077 v[i] = LLVMBuildExtractElement(ctx->builder, vdata,
1078 LLVMConstInt(ctx->i32, i, 0), "");
1079 }
1080 v01 = ac_build_gather_values(ctx, v, 2);
1081
1082 ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset,
1083 soffset, inst_offset, glc, slc,
1084 writeonly_memory, swizzle_enable_hint);
1085 ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset,
1086 soffset, inst_offset + 8,
1087 glc, slc,
1088 writeonly_memory, swizzle_enable_hint);
1089 return;
1090 }
1091
1092 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1093 * (voffset is swizzled, but soffset isn't swizzled).
1094 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1095 */
1096 if (!swizzle_enable_hint) {
1097 LLVMValueRef offset = soffset;
1098
1099 static const char *types[] = {"f32", "v2f32", "v4f32"};
1100
1101 if (inst_offset)
1102 offset = LLVMBuildAdd(ctx->builder, offset,
1103 LLVMConstInt(ctx->i32, inst_offset, 0), "");
1104 if (voffset)
1105 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1106
1107 LLVMValueRef args[] = {
1108 ac_to_float(ctx, vdata),
1109 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
1110 ctx->i32_0,
1111 offset,
1112 LLVMConstInt(ctx->i1, glc, 0),
1113 LLVMConstInt(ctx->i1, slc, 0),
1114 };
1115
1116 char name[256];
1117 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.store.%s",
1118 types[CLAMP(num_channels, 1, 3) - 1]);
1119
1120 ac_build_intrinsic(ctx, name, ctx->voidt,
1121 args, ARRAY_SIZE(args),
1122 writeonly_memory ?
1123 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY :
1124 AC_FUNC_ATTR_WRITEONLY);
1125 return;
1126 }
1127
1128 static const unsigned dfmt[] = {
1129 V_008F0C_BUF_DATA_FORMAT_32,
1130 V_008F0C_BUF_DATA_FORMAT_32_32,
1131 V_008F0C_BUF_DATA_FORMAT_32_32_32,
1132 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1133 };
1134 static const char *types[] = {"i32", "v2i32", "v4i32"};
1135 LLVMValueRef args[] = {
1136 vdata,
1137 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
1138 ctx->i32_0,
1139 voffset ? voffset : ctx->i32_0,
1140 soffset,
1141 LLVMConstInt(ctx->i32, inst_offset, 0),
1142 LLVMConstInt(ctx->i32, dfmt[num_channels - 1], 0),
1143 LLVMConstInt(ctx->i32, V_008F0C_BUF_NUM_FORMAT_UINT, 0),
1144 LLVMConstInt(ctx->i1, glc, 0),
1145 LLVMConstInt(ctx->i1, slc, 0),
1146 };
1147 char name[256];
1148 snprintf(name, sizeof(name), "llvm.amdgcn.tbuffer.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 }
1157
1158 static LLVMValueRef
1159 ac_build_buffer_load_common(struct ac_llvm_context *ctx,
1160 LLVMValueRef rsrc,
1161 LLVMValueRef vindex,
1162 LLVMValueRef voffset,
1163 unsigned num_channels,
1164 bool glc,
1165 bool slc,
1166 bool can_speculate,
1167 bool use_format)
1168 {
1169 LLVMValueRef args[] = {
1170 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
1171 vindex ? vindex : ctx->i32_0,
1172 voffset,
1173 LLVMConstInt(ctx->i1, glc, 0),
1174 LLVMConstInt(ctx->i1, slc, 0)
1175 };
1176 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1177
1178 LLVMTypeRef types[] = {ctx->f32, ctx->v2f32, ctx->v4f32};
1179 const char *type_names[] = {"f32", "v2f32", "v4f32"};
1180 char name[256];
1181
1182 if (use_format) {
1183 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.format.%s",
1184 type_names[func]);
1185 } else {
1186 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.%s",
1187 type_names[func]);
1188 }
1189
1190 return ac_build_intrinsic(ctx, name, types[func], args,
1191 ARRAY_SIZE(args),
1192 ac_get_load_intr_attribs(can_speculate));
1193 }
1194
1195 static LLVMValueRef
1196 ac_build_llvm8_buffer_load_common(struct ac_llvm_context *ctx,
1197 LLVMValueRef rsrc,
1198 LLVMValueRef vindex,
1199 LLVMValueRef voffset,
1200 LLVMValueRef soffset,
1201 unsigned num_channels,
1202 bool glc,
1203 bool slc,
1204 bool can_speculate,
1205 bool use_format,
1206 bool structurized)
1207 {
1208 LLVMValueRef args[5];
1209 int idx = 0;
1210 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1211 if (structurized)
1212 args[idx++] = vindex ? vindex : ctx->i32_0;
1213 args[idx++] = voffset ? voffset : ctx->i32_0;
1214 args[idx++] = soffset ? soffset : ctx->i32_0;
1215 args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0);
1216 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1217
1218 LLVMTypeRef types[] = {ctx->f32, ctx->v2f32, ctx->v4f32};
1219 const char *type_names[] = {"f32", "v2f32", "v4f32"};
1220 const char *indexing_kind = structurized ? "struct" : "raw";
1221 char name[256];
1222
1223 if (use_format) {
1224 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.format.%s",
1225 indexing_kind, type_names[func]);
1226 } else {
1227 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.%s",
1228 indexing_kind, type_names[func]);
1229 }
1230
1231 return ac_build_intrinsic(ctx, name, types[func], args,
1232 idx,
1233 ac_get_load_intr_attribs(can_speculate));
1234 }
1235
1236 LLVMValueRef
1237 ac_build_buffer_load(struct ac_llvm_context *ctx,
1238 LLVMValueRef rsrc,
1239 int num_channels,
1240 LLVMValueRef vindex,
1241 LLVMValueRef voffset,
1242 LLVMValueRef soffset,
1243 unsigned inst_offset,
1244 unsigned glc,
1245 unsigned slc,
1246 bool can_speculate,
1247 bool allow_smem)
1248 {
1249 LLVMValueRef offset = LLVMConstInt(ctx->i32, inst_offset, 0);
1250 if (voffset)
1251 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1252 if (soffset)
1253 offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
1254
1255 if (allow_smem && !slc &&
1256 (!glc || (HAVE_LLVM >= 0x0800 && ctx->chip_class >= VI))) {
1257 assert(vindex == NULL);
1258
1259 LLVMValueRef result[8];
1260
1261 for (int i = 0; i < num_channels; i++) {
1262 if (i) {
1263 offset = LLVMBuildAdd(ctx->builder, offset,
1264 LLVMConstInt(ctx->i32, 4, 0), "");
1265 }
1266 const char *intrname =
1267 HAVE_LLVM >= 0x0800 ? "llvm.amdgcn.s.buffer.load.f32"
1268 : "llvm.SI.load.const.v4i32";
1269 unsigned num_args = HAVE_LLVM >= 0x0800 ? 3 : 2;
1270 LLVMValueRef args[3] = {
1271 rsrc,
1272 offset,
1273 glc ? ctx->i32_1 : ctx->i32_0,
1274 };
1275 result[i] = ac_build_intrinsic(ctx, intrname,
1276 ctx->f32, args, num_args,
1277 AC_FUNC_ATTR_READNONE |
1278 (HAVE_LLVM < 0x0800 ? AC_FUNC_ATTR_LEGACY : 0));
1279 }
1280 if (num_channels == 1)
1281 return result[0];
1282
1283 if (num_channels == 3)
1284 result[num_channels++] = LLVMGetUndef(ctx->f32);
1285 return ac_build_gather_values(ctx, result, num_channels);
1286 }
1287
1288 return ac_build_buffer_load_common(ctx, rsrc, vindex, offset,
1289 num_channels, glc, slc,
1290 can_speculate, false);
1291 }
1292
1293 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
1294 LLVMValueRef rsrc,
1295 LLVMValueRef vindex,
1296 LLVMValueRef voffset,
1297 unsigned num_channels,
1298 bool glc,
1299 bool can_speculate)
1300 {
1301 if (HAVE_LLVM >= 0x800) {
1302 return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
1303 num_channels, glc, false,
1304 can_speculate, true, true);
1305 }
1306 return ac_build_buffer_load_common(ctx, rsrc, vindex, voffset,
1307 num_channels, glc, false,
1308 can_speculate, true);
1309 }
1310
1311 LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
1312 LLVMValueRef rsrc,
1313 LLVMValueRef vindex,
1314 LLVMValueRef voffset,
1315 unsigned num_channels,
1316 bool glc,
1317 bool can_speculate)
1318 {
1319 if (HAVE_LLVM >= 0x800) {
1320 return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
1321 num_channels, glc, false,
1322 can_speculate, true, true);
1323 }
1324
1325 LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->builder, rsrc, LLVMConstInt(ctx->i32, 2, 0), "");
1326 LLVMValueRef stride = LLVMBuildExtractElement(ctx->builder, rsrc, ctx->i32_1, "");
1327 stride = LLVMBuildLShr(ctx->builder, stride, LLVMConstInt(ctx->i32, 16, 0), "");
1328
1329 LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->builder,
1330 LLVMBuildICmp(ctx->builder, LLVMIntUGT, elem_count, stride, ""),
1331 elem_count, stride, "");
1332
1333 LLVMValueRef new_rsrc = LLVMBuildInsertElement(ctx->builder, rsrc, new_elem_count,
1334 LLVMConstInt(ctx->i32, 2, 0), "");
1335
1336 return ac_build_buffer_load_common(ctx, new_rsrc, vindex, voffset,
1337 num_channels, glc, false,
1338 can_speculate, true);
1339 }
1340
1341 LLVMValueRef
1342 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
1343 LLVMValueRef rsrc,
1344 LLVMValueRef vindex,
1345 LLVMValueRef voffset,
1346 LLVMValueRef soffset,
1347 LLVMValueRef immoffset,
1348 LLVMValueRef glc)
1349 {
1350 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
1351 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1352 LLVMValueRef res;
1353
1354 if (HAVE_LLVM >= 0x0800) {
1355 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1356
1357 res = ac_build_llvm8_tbuffer_load(ctx, rsrc, vindex, voffset,
1358 soffset, 1, dfmt, nfmt, glc,
1359 false, true, true);
1360 } else {
1361 const char *name = "llvm.amdgcn.tbuffer.load.i32";
1362 LLVMTypeRef type = ctx->i32;
1363 LLVMValueRef params[] = {
1364 rsrc,
1365 vindex,
1366 voffset,
1367 soffset,
1368 immoffset,
1369 LLVMConstInt(ctx->i32, dfmt, false),
1370 LLVMConstInt(ctx->i32, nfmt, false),
1371 glc,
1372 ctx->i1false,
1373 };
1374 res = ac_build_intrinsic(ctx, name, type, params, 9, 0);
1375 }
1376
1377 return LLVMBuildTrunc(ctx->builder, res, ctx->i16, "");
1378 }
1379
1380 LLVMValueRef
1381 ac_build_llvm8_tbuffer_load(struct ac_llvm_context *ctx,
1382 LLVMValueRef rsrc,
1383 LLVMValueRef vindex,
1384 LLVMValueRef voffset,
1385 LLVMValueRef soffset,
1386 unsigned num_channels,
1387 unsigned dfmt,
1388 unsigned nfmt,
1389 bool glc,
1390 bool slc,
1391 bool can_speculate,
1392 bool structurized)
1393 {
1394 LLVMValueRef args[6];
1395 int idx = 0;
1396 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1397 if (structurized)
1398 args[idx++] = vindex ? vindex : ctx->i32_0;
1399 args[idx++] = voffset ? voffset : ctx->i32_0;
1400 args[idx++] = soffset ? soffset : ctx->i32_0;
1401 args[idx++] = LLVMConstInt(ctx->i32, dfmt | (nfmt << 4), 0);
1402 args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0);
1403 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1404
1405 LLVMTypeRef types[] = {ctx->i32, ctx->v2i32, ctx->v4i32};
1406 const char *type_names[] = {"i32", "v2i32", "v4i32"};
1407 const char *indexing_kind = structurized ? "struct" : "raw";
1408 char name[256];
1409
1410 snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.load.%s",
1411 indexing_kind, type_names[func]);
1412
1413 return ac_build_intrinsic(ctx, name, types[func], args,
1414 idx,
1415 ac_get_load_intr_attribs(can_speculate));
1416 }
1417
1418 /**
1419 * Set range metadata on an instruction. This can only be used on load and
1420 * call instructions. If you know an instruction can only produce the values
1421 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
1422 * \p lo is the minimum value inclusive.
1423 * \p hi is the maximum value exclusive.
1424 */
1425 static void set_range_metadata(struct ac_llvm_context *ctx,
1426 LLVMValueRef value, unsigned lo, unsigned hi)
1427 {
1428 LLVMValueRef range_md, md_args[2];
1429 LLVMTypeRef type = LLVMTypeOf(value);
1430 LLVMContextRef context = LLVMGetTypeContext(type);
1431
1432 md_args[0] = LLVMConstInt(type, lo, false);
1433 md_args[1] = LLVMConstInt(type, hi, false);
1434 range_md = LLVMMDNodeInContext(context, md_args, 2);
1435 LLVMSetMetadata(value, ctx->range_md_kind, range_md);
1436 }
1437
1438 LLVMValueRef
1439 ac_get_thread_id(struct ac_llvm_context *ctx)
1440 {
1441 LLVMValueRef tid;
1442
1443 LLVMValueRef tid_args[2];
1444 tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
1445 tid_args[1] = ctx->i32_0;
1446 tid_args[1] = ac_build_intrinsic(ctx,
1447 "llvm.amdgcn.mbcnt.lo", ctx->i32,
1448 tid_args, 2, AC_FUNC_ATTR_READNONE);
1449
1450 tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
1451 ctx->i32, tid_args,
1452 2, AC_FUNC_ATTR_READNONE);
1453 set_range_metadata(ctx, tid, 0, 64);
1454 return tid;
1455 }
1456
1457 /*
1458 * SI implements derivatives using the local data store (LDS)
1459 * All writes to the LDS happen in all executing threads at
1460 * the same time. TID is the Thread ID for the current
1461 * thread and is a value between 0 and 63, representing
1462 * the thread's position in the wavefront.
1463 *
1464 * For the pixel shader threads are grouped into quads of four pixels.
1465 * The TIDs of the pixels of a quad are:
1466 *
1467 * +------+------+
1468 * |4n + 0|4n + 1|
1469 * +------+------+
1470 * |4n + 2|4n + 3|
1471 * +------+------+
1472 *
1473 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
1474 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
1475 * the current pixel's column, and masking with 0xfffffffe yields the TID
1476 * of the left pixel of the current pixel's row.
1477 *
1478 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
1479 * adding 2 yields the TID of the pixel below the top pixel.
1480 */
1481 LLVMValueRef
1482 ac_build_ddxy(struct ac_llvm_context *ctx,
1483 uint32_t mask,
1484 int idx,
1485 LLVMValueRef val)
1486 {
1487 unsigned tl_lanes[4], trbl_lanes[4];
1488 LLVMValueRef tl, trbl;
1489 LLVMValueRef result;
1490
1491 for (unsigned i = 0; i < 4; ++i) {
1492 tl_lanes[i] = i & mask;
1493 trbl_lanes[i] = (i & mask) + idx;
1494 }
1495
1496 tl = ac_build_quad_swizzle(ctx, val,
1497 tl_lanes[0], tl_lanes[1],
1498 tl_lanes[2], tl_lanes[3]);
1499 trbl = ac_build_quad_swizzle(ctx, val,
1500 trbl_lanes[0], trbl_lanes[1],
1501 trbl_lanes[2], trbl_lanes[3]);
1502
1503 tl = LLVMBuildBitCast(ctx->builder, tl, ctx->f32, "");
1504 trbl = LLVMBuildBitCast(ctx->builder, trbl, ctx->f32, "");
1505 result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
1506
1507 result = ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.f32", ctx->f32,
1508 &result, 1, 0);
1509
1510 return result;
1511 }
1512
1513 void
1514 ac_build_sendmsg(struct ac_llvm_context *ctx,
1515 uint32_t msg,
1516 LLVMValueRef wave_id)
1517 {
1518 LLVMValueRef args[2];
1519 args[0] = LLVMConstInt(ctx->i32, msg, false);
1520 args[1] = wave_id;
1521 ac_build_intrinsic(ctx, "llvm.amdgcn.s.sendmsg", ctx->voidt, args, 2, 0);
1522 }
1523
1524 LLVMValueRef
1525 ac_build_imsb(struct ac_llvm_context *ctx,
1526 LLVMValueRef arg,
1527 LLVMTypeRef dst_type)
1528 {
1529 LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.amdgcn.sffbh.i32",
1530 dst_type, &arg, 1,
1531 AC_FUNC_ATTR_READNONE);
1532
1533 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
1534 * the index from LSB. Invert it by doing "31 - msb". */
1535 msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
1536 msb, "");
1537
1538 LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
1539 LLVMValueRef cond = LLVMBuildOr(ctx->builder,
1540 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
1541 arg, ctx->i32_0, ""),
1542 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
1543 arg, all_ones, ""), "");
1544
1545 return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
1546 }
1547
1548 LLVMValueRef
1549 ac_build_umsb(struct ac_llvm_context *ctx,
1550 LLVMValueRef arg,
1551 LLVMTypeRef dst_type)
1552 {
1553 const char *intrin_name;
1554 LLVMTypeRef type;
1555 LLVMValueRef highest_bit;
1556 LLVMValueRef zero;
1557 unsigned bitsize;
1558
1559 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(arg));
1560 switch (bitsize) {
1561 case 64:
1562 intrin_name = "llvm.ctlz.i64";
1563 type = ctx->i64;
1564 highest_bit = LLVMConstInt(ctx->i64, 63, false);
1565 zero = ctx->i64_0;
1566 break;
1567 case 32:
1568 intrin_name = "llvm.ctlz.i32";
1569 type = ctx->i32;
1570 highest_bit = LLVMConstInt(ctx->i32, 31, false);
1571 zero = ctx->i32_0;
1572 break;
1573 case 16:
1574 intrin_name = "llvm.ctlz.i16";
1575 type = ctx->i16;
1576 highest_bit = LLVMConstInt(ctx->i16, 15, false);
1577 zero = ctx->i16_0;
1578 break;
1579 default:
1580 unreachable(!"invalid bitsize");
1581 break;
1582 }
1583
1584 LLVMValueRef params[2] = {
1585 arg,
1586 ctx->i1true,
1587 };
1588
1589 LLVMValueRef msb = ac_build_intrinsic(ctx, intrin_name, type,
1590 params, 2,
1591 AC_FUNC_ATTR_READNONE);
1592
1593 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
1594 * the index from LSB. Invert it by doing "31 - msb". */
1595 msb = LLVMBuildSub(ctx->builder, highest_bit, msb, "");
1596 msb = LLVMBuildTruncOrBitCast(ctx->builder, msb, ctx->i32, "");
1597
1598 /* check for zero */
1599 return LLVMBuildSelect(ctx->builder,
1600 LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg, zero, ""),
1601 LLVMConstInt(ctx->i32, -1, true), msb, "");
1602 }
1603
1604 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
1605 LLVMValueRef b)
1606 {
1607 LLVMValueRef args[2] = {a, b};
1608 return ac_build_intrinsic(ctx, "llvm.minnum.f32", ctx->f32, args, 2,
1609 AC_FUNC_ATTR_READNONE);
1610 }
1611
1612 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
1613 LLVMValueRef b)
1614 {
1615 LLVMValueRef args[2] = {a, b};
1616 return ac_build_intrinsic(ctx, "llvm.maxnum.f32", ctx->f32, args, 2,
1617 AC_FUNC_ATTR_READNONE);
1618 }
1619
1620 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
1621 LLVMValueRef b)
1622 {
1623 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSLE, a, b, "");
1624 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
1625 }
1626
1627 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
1628 LLVMValueRef b)
1629 {
1630 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, a, b, "");
1631 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
1632 }
1633
1634 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a,
1635 LLVMValueRef b)
1636 {
1637 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntULE, a, b, "");
1638 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
1639 }
1640
1641 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value)
1642 {
1643 return ac_build_fmin(ctx, ac_build_fmax(ctx, value, ctx->f32_0),
1644 ctx->f32_1);
1645 }
1646
1647 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a)
1648 {
1649 LLVMValueRef args[9];
1650
1651 args[0] = LLVMConstInt(ctx->i32, a->target, 0);
1652 args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
1653
1654 if (a->compr) {
1655 LLVMTypeRef i16 = LLVMInt16TypeInContext(ctx->context);
1656 LLVMTypeRef v2i16 = LLVMVectorType(i16, 2);
1657
1658 args[2] = LLVMBuildBitCast(ctx->builder, a->out[0],
1659 v2i16, "");
1660 args[3] = LLVMBuildBitCast(ctx->builder, a->out[1],
1661 v2i16, "");
1662 args[4] = LLVMConstInt(ctx->i1, a->done, 0);
1663 args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
1664
1665 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16",
1666 ctx->voidt, args, 6, 0);
1667 } else {
1668 args[2] = a->out[0];
1669 args[3] = a->out[1];
1670 args[4] = a->out[2];
1671 args[5] = a->out[3];
1672 args[6] = LLVMConstInt(ctx->i1, a->done, 0);
1673 args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
1674
1675 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.f32",
1676 ctx->voidt, args, 8, 0);
1677 }
1678 }
1679
1680 void ac_build_export_null(struct ac_llvm_context *ctx)
1681 {
1682 struct ac_export_args args;
1683
1684 args.enabled_channels = 0x0; /* enabled channels */
1685 args.valid_mask = 1; /* whether the EXEC mask is valid */
1686 args.done = 1; /* DONE bit */
1687 args.target = V_008DFC_SQ_EXP_NULL;
1688 args.compr = 0; /* COMPR flag (0 = 32-bit export) */
1689 args.out[0] = LLVMGetUndef(ctx->f32); /* R */
1690 args.out[1] = LLVMGetUndef(ctx->f32); /* G */
1691 args.out[2] = LLVMGetUndef(ctx->f32); /* B */
1692 args.out[3] = LLVMGetUndef(ctx->f32); /* A */
1693
1694 ac_build_export(ctx, &args);
1695 }
1696
1697 static unsigned ac_num_coords(enum ac_image_dim dim)
1698 {
1699 switch (dim) {
1700 case ac_image_1d:
1701 return 1;
1702 case ac_image_2d:
1703 case ac_image_1darray:
1704 return 2;
1705 case ac_image_3d:
1706 case ac_image_cube:
1707 case ac_image_2darray:
1708 case ac_image_2dmsaa:
1709 return 3;
1710 case ac_image_2darraymsaa:
1711 return 4;
1712 default:
1713 unreachable("ac_num_coords: bad dim");
1714 }
1715 }
1716
1717 static unsigned ac_num_derivs(enum ac_image_dim dim)
1718 {
1719 switch (dim) {
1720 case ac_image_1d:
1721 case ac_image_1darray:
1722 return 2;
1723 case ac_image_2d:
1724 case ac_image_2darray:
1725 case ac_image_cube:
1726 return 4;
1727 case ac_image_3d:
1728 return 6;
1729 case ac_image_2dmsaa:
1730 case ac_image_2darraymsaa:
1731 default:
1732 unreachable("derivatives not supported");
1733 }
1734 }
1735
1736 static const char *get_atomic_name(enum ac_atomic_op op)
1737 {
1738 switch (op) {
1739 case ac_atomic_swap: return "swap";
1740 case ac_atomic_add: return "add";
1741 case ac_atomic_sub: return "sub";
1742 case ac_atomic_smin: return "smin";
1743 case ac_atomic_umin: return "umin";
1744 case ac_atomic_smax: return "smax";
1745 case ac_atomic_umax: return "umax";
1746 case ac_atomic_and: return "and";
1747 case ac_atomic_or: return "or";
1748 case ac_atomic_xor: return "xor";
1749 }
1750 unreachable("bad atomic op");
1751 }
1752
1753 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
1754 struct ac_image_args *a)
1755 {
1756 const char *overload[3] = { "", "", "" };
1757 unsigned num_overloads = 0;
1758 LLVMValueRef args[18];
1759 unsigned num_args = 0;
1760 enum ac_image_dim dim = a->dim;
1761
1762 assert(!a->lod || a->lod == ctx->i32_0 || a->lod == ctx->f32_0 ||
1763 !a->level_zero);
1764 assert((a->opcode != ac_image_get_resinfo && a->opcode != ac_image_load_mip &&
1765 a->opcode != ac_image_store_mip) ||
1766 a->lod);
1767 assert(a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
1768 (!a->compare && !a->offset));
1769 assert((a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
1770 a->opcode == ac_image_get_lod) ||
1771 !a->bias);
1772 assert((a->bias ? 1 : 0) +
1773 (a->lod ? 1 : 0) +
1774 (a->level_zero ? 1 : 0) +
1775 (a->derivs[0] ? 1 : 0) <= 1);
1776
1777 if (a->opcode == ac_image_get_lod) {
1778 switch (dim) {
1779 case ac_image_1darray:
1780 dim = ac_image_1d;
1781 break;
1782 case ac_image_2darray:
1783 case ac_image_cube:
1784 dim = ac_image_2d;
1785 break;
1786 default:
1787 break;
1788 }
1789 }
1790
1791 bool sample = a->opcode == ac_image_sample ||
1792 a->opcode == ac_image_gather4 ||
1793 a->opcode == ac_image_get_lod;
1794 bool atomic = a->opcode == ac_image_atomic ||
1795 a->opcode == ac_image_atomic_cmpswap;
1796 LLVMTypeRef coord_type = sample ? ctx->f32 : ctx->i32;
1797
1798 if (atomic || a->opcode == ac_image_store || a->opcode == ac_image_store_mip) {
1799 args[num_args++] = a->data[0];
1800 if (a->opcode == ac_image_atomic_cmpswap)
1801 args[num_args++] = a->data[1];
1802 }
1803
1804 if (!atomic)
1805 args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, false);
1806
1807 if (a->offset)
1808 args[num_args++] = ac_to_integer(ctx, a->offset);
1809 if (a->bias) {
1810 args[num_args++] = ac_to_float(ctx, a->bias);
1811 overload[num_overloads++] = ".f32";
1812 }
1813 if (a->compare)
1814 args[num_args++] = ac_to_float(ctx, a->compare);
1815 if (a->derivs[0]) {
1816 unsigned count = ac_num_derivs(dim);
1817 for (unsigned i = 0; i < count; ++i)
1818 args[num_args++] = ac_to_float(ctx, a->derivs[i]);
1819 overload[num_overloads++] = ".f32";
1820 }
1821 unsigned num_coords =
1822 a->opcode != ac_image_get_resinfo ? ac_num_coords(dim) : 0;
1823 for (unsigned i = 0; i < num_coords; ++i)
1824 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->coords[i], coord_type, "");
1825 if (a->lod)
1826 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->lod, coord_type, "");
1827 overload[num_overloads++] = sample ? ".f32" : ".i32";
1828
1829 args[num_args++] = a->resource;
1830 if (sample) {
1831 args[num_args++] = a->sampler;
1832 args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, false);
1833 }
1834
1835 args[num_args++] = ctx->i32_0; /* texfailctrl */
1836 args[num_args++] = LLVMConstInt(ctx->i32, a->cache_policy, false);
1837
1838 const char *name;
1839 const char *atomic_subop = "";
1840 switch (a->opcode) {
1841 case ac_image_sample: name = "sample"; break;
1842 case ac_image_gather4: name = "gather4"; break;
1843 case ac_image_load: name = "load"; break;
1844 case ac_image_load_mip: name = "load.mip"; break;
1845 case ac_image_store: name = "store"; break;
1846 case ac_image_store_mip: name = "store.mip"; break;
1847 case ac_image_atomic:
1848 name = "atomic.";
1849 atomic_subop = get_atomic_name(a->atomic);
1850 break;
1851 case ac_image_atomic_cmpswap:
1852 name = "atomic.";
1853 atomic_subop = "cmpswap";
1854 break;
1855 case ac_image_get_lod: name = "getlod"; break;
1856 case ac_image_get_resinfo: name = "getresinfo"; break;
1857 default: unreachable("invalid image opcode");
1858 }
1859
1860 const char *dimname;
1861 switch (dim) {
1862 case ac_image_1d: dimname = "1d"; break;
1863 case ac_image_2d: dimname = "2d"; break;
1864 case ac_image_3d: dimname = "3d"; break;
1865 case ac_image_cube: dimname = "cube"; break;
1866 case ac_image_1darray: dimname = "1darray"; break;
1867 case ac_image_2darray: dimname = "2darray"; break;
1868 case ac_image_2dmsaa: dimname = "2dmsaa"; break;
1869 case ac_image_2darraymsaa: dimname = "2darraymsaa"; break;
1870 default: unreachable("invalid dim");
1871 }
1872
1873 bool lod_suffix =
1874 a->lod && (a->opcode == ac_image_sample || a->opcode == ac_image_gather4);
1875 char intr_name[96];
1876 snprintf(intr_name, sizeof(intr_name),
1877 "llvm.amdgcn.image.%s%s" /* base name */
1878 "%s%s%s" /* sample/gather modifiers */
1879 ".%s.%s%s%s%s", /* dimension and type overloads */
1880 name, atomic_subop,
1881 a->compare ? ".c" : "",
1882 a->bias ? ".b" :
1883 lod_suffix ? ".l" :
1884 a->derivs[0] ? ".d" :
1885 a->level_zero ? ".lz" : "",
1886 a->offset ? ".o" : "",
1887 dimname,
1888 atomic ? "i32" : "v4f32",
1889 overload[0], overload[1], overload[2]);
1890
1891 LLVMTypeRef retty;
1892 if (atomic)
1893 retty = ctx->i32;
1894 else if (a->opcode == ac_image_store || a->opcode == ac_image_store_mip)
1895 retty = ctx->voidt;
1896 else
1897 retty = ctx->v4f32;
1898
1899 LLVMValueRef result =
1900 ac_build_intrinsic(ctx, intr_name, retty, args, num_args,
1901 a->attributes);
1902 if (!sample && retty == ctx->v4f32) {
1903 result = LLVMBuildBitCast(ctx->builder, result,
1904 ctx->v4i32, "");
1905 }
1906 return result;
1907 }
1908
1909 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
1910 LLVMValueRef args[2])
1911 {
1912 LLVMTypeRef v2f16 =
1913 LLVMVectorType(LLVMHalfTypeInContext(ctx->context), 2);
1914
1915 return ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz", v2f16,
1916 args, 2, AC_FUNC_ATTR_READNONE);
1917 }
1918
1919 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
1920 LLVMValueRef args[2])
1921 {
1922 LLVMValueRef res =
1923 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16",
1924 ctx->v2i16, args, 2,
1925 AC_FUNC_ATTR_READNONE);
1926 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1927 }
1928
1929 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
1930 LLVMValueRef args[2])
1931 {
1932 LLVMValueRef res =
1933 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16",
1934 ctx->v2i16, args, 2,
1935 AC_FUNC_ATTR_READNONE);
1936 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1937 }
1938
1939 /* The 8-bit and 10-bit clamping is for HW workarounds. */
1940 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
1941 LLVMValueRef args[2], unsigned bits, bool hi)
1942 {
1943 assert(bits == 8 || bits == 10 || bits == 16);
1944
1945 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
1946 bits == 8 ? 127 : bits == 10 ? 511 : 32767, 0);
1947 LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
1948 bits == 8 ? -128 : bits == 10 ? -512 : -32768, 0);
1949 LLVMValueRef max_alpha =
1950 bits != 10 ? max_rgb : ctx->i32_1;
1951 LLVMValueRef min_alpha =
1952 bits != 10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
1953
1954 /* Clamp. */
1955 if (bits != 16) {
1956 for (int i = 0; i < 2; i++) {
1957 bool alpha = hi && i == 1;
1958 args[i] = ac_build_imin(ctx, args[i],
1959 alpha ? max_alpha : max_rgb);
1960 args[i] = ac_build_imax(ctx, args[i],
1961 alpha ? min_alpha : min_rgb);
1962 }
1963 }
1964
1965 LLVMValueRef res =
1966 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16",
1967 ctx->v2i16, args, 2,
1968 AC_FUNC_ATTR_READNONE);
1969 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1970 }
1971
1972 /* The 8-bit and 10-bit clamping is for HW workarounds. */
1973 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
1974 LLVMValueRef args[2], unsigned bits, bool hi)
1975 {
1976 assert(bits == 8 || bits == 10 || bits == 16);
1977
1978 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
1979 bits == 8 ? 255 : bits == 10 ? 1023 : 65535, 0);
1980 LLVMValueRef max_alpha =
1981 bits != 10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
1982
1983 /* Clamp. */
1984 if (bits != 16) {
1985 for (int i = 0; i < 2; i++) {
1986 bool alpha = hi && i == 1;
1987 args[i] = ac_build_umin(ctx, args[i],
1988 alpha ? max_alpha : max_rgb);
1989 }
1990 }
1991
1992 LLVMValueRef res =
1993 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16",
1994 ctx->v2i16, args, 2,
1995 AC_FUNC_ATTR_READNONE);
1996 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
1997 }
1998
1999 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1)
2000 {
2001 return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1,
2002 &i1, 1, AC_FUNC_ATTR_READNONE);
2003 }
2004
2005 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1)
2006 {
2007 ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt,
2008 &i1, 1, 0);
2009 }
2010
2011 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
2012 LLVMValueRef offset, LLVMValueRef width,
2013 bool is_signed)
2014 {
2015 LLVMValueRef args[] = {
2016 input,
2017 offset,
2018 width,
2019 };
2020
2021 return ac_build_intrinsic(ctx,
2022 is_signed ? "llvm.amdgcn.sbfe.i32" :
2023 "llvm.amdgcn.ubfe.i32",
2024 ctx->i32, args, 3,
2025 AC_FUNC_ATTR_READNONE);
2026 }
2027
2028 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2029 LLVMValueRef s1, LLVMValueRef s2)
2030 {
2031 return LLVMBuildAdd(ctx->builder,
2032 LLVMBuildMul(ctx->builder, s0, s1, ""), s2, "");
2033 }
2034
2035 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2036 LLVMValueRef s1, LLVMValueRef s2)
2037 {
2038 return LLVMBuildFAdd(ctx->builder,
2039 LLVMBuildFMul(ctx->builder, s0, s1, ""), s2, "");
2040 }
2041
2042 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16)
2043 {
2044 LLVMValueRef args[1] = {
2045 LLVMConstInt(ctx->i32, simm16, false),
2046 };
2047 ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt",
2048 ctx->voidt, args, 1, 0);
2049 }
2050
2051 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
2052 unsigned bitsize)
2053 {
2054 LLVMTypeRef type;
2055 char *intr;
2056
2057 if (bitsize == 32) {
2058 intr = "llvm.floor.f32";
2059 type = ctx->f32;
2060 } else {
2061 intr = "llvm.floor.f64";
2062 type = ctx->f64;
2063 }
2064
2065 LLVMValueRef params[] = {
2066 src0,
2067 };
2068 LLVMValueRef floor = ac_build_intrinsic(ctx, intr, type, params, 1,
2069 AC_FUNC_ATTR_READNONE);
2070 return LLVMBuildFSub(ctx->builder, src0, floor, "");
2071 }
2072
2073 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2074 unsigned bitsize)
2075 {
2076 LLVMValueRef cmp, val, zero, one;
2077 LLVMTypeRef type;
2078
2079 switch (bitsize) {
2080 case 64:
2081 type = ctx->i64;
2082 zero = ctx->i64_0;
2083 one = ctx->i64_1;
2084 break;
2085 case 32:
2086 type = ctx->i32;
2087 zero = ctx->i32_0;
2088 one = ctx->i32_1;
2089 break;
2090 case 16:
2091 type = ctx->i16;
2092 zero = ctx->i16_0;
2093 one = ctx->i16_1;
2094 break;
2095 default:
2096 unreachable(!"invalid bitsize");
2097 break;
2098 }
2099
2100 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, src0, zero, "");
2101 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2102 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGE, val, zero, "");
2103 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstInt(type, -1, true), "");
2104 return val;
2105 }
2106
2107 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2108 unsigned bitsize)
2109 {
2110 LLVMValueRef cmp, val, zero, one;
2111 LLVMTypeRef type;
2112
2113 if (bitsize == 32) {
2114 type = ctx->f32;
2115 zero = ctx->f32_0;
2116 one = ctx->f32_1;
2117 } else {
2118 type = ctx->f64;
2119 zero = ctx->f64_0;
2120 one = ctx->f64_1;
2121 }
2122
2123 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src0, zero, "");
2124 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2125 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGE, val, zero, "");
2126 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstReal(type, -1.0), "");
2127 return val;
2128 }
2129
2130 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0)
2131 {
2132 LLVMValueRef result;
2133 unsigned bitsize;
2134
2135 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2136
2137 switch (bitsize) {
2138 case 64:
2139 result = ac_build_intrinsic(ctx, "llvm.ctpop.i64", ctx->i64,
2140 (LLVMValueRef []) { src0 }, 1,
2141 AC_FUNC_ATTR_READNONE);
2142
2143 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
2144 break;
2145 case 32:
2146 result = ac_build_intrinsic(ctx, "llvm.ctpop.i32", ctx->i32,
2147 (LLVMValueRef []) { src0 }, 1,
2148 AC_FUNC_ATTR_READNONE);
2149 break;
2150 case 16:
2151 result = ac_build_intrinsic(ctx, "llvm.ctpop.i16", ctx->i16,
2152 (LLVMValueRef []) { src0 }, 1,
2153 AC_FUNC_ATTR_READNONE);
2154 break;
2155 default:
2156 unreachable(!"invalid bitsize");
2157 break;
2158 }
2159
2160 return result;
2161 }
2162
2163 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
2164 LLVMValueRef src0)
2165 {
2166 LLVMValueRef result;
2167 unsigned bitsize;
2168
2169 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2170
2171 switch (bitsize) {
2172 case 32:
2173 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i32", ctx->i32,
2174 (LLVMValueRef []) { src0 }, 1,
2175 AC_FUNC_ATTR_READNONE);
2176 break;
2177 case 16:
2178 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i16", ctx->i16,
2179 (LLVMValueRef []) { src0 }, 1,
2180 AC_FUNC_ATTR_READNONE);
2181 break;
2182 default:
2183 unreachable(!"invalid bitsize");
2184 break;
2185 }
2186
2187 return result;
2188 }
2189
2190 #define AC_EXP_TARGET 0
2191 #define AC_EXP_ENABLED_CHANNELS 1
2192 #define AC_EXP_OUT0 2
2193
2194 enum ac_ir_type {
2195 AC_IR_UNDEF,
2196 AC_IR_CONST,
2197 AC_IR_VALUE,
2198 };
2199
2200 struct ac_vs_exp_chan
2201 {
2202 LLVMValueRef value;
2203 float const_float;
2204 enum ac_ir_type type;
2205 };
2206
2207 struct ac_vs_exp_inst {
2208 unsigned offset;
2209 LLVMValueRef inst;
2210 struct ac_vs_exp_chan chan[4];
2211 };
2212
2213 struct ac_vs_exports {
2214 unsigned num;
2215 struct ac_vs_exp_inst exp[VARYING_SLOT_MAX];
2216 };
2217
2218 /* Return true if the PARAM export has been eliminated. */
2219 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset,
2220 uint32_t num_outputs,
2221 struct ac_vs_exp_inst *exp)
2222 {
2223 unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
2224 bool is_zero[4] = {}, is_one[4] = {};
2225
2226 for (i = 0; i < 4; i++) {
2227 /* It's a constant expression. Undef outputs are eliminated too. */
2228 if (exp->chan[i].type == AC_IR_UNDEF) {
2229 is_zero[i] = true;
2230 is_one[i] = true;
2231 } else if (exp->chan[i].type == AC_IR_CONST) {
2232 if (exp->chan[i].const_float == 0)
2233 is_zero[i] = true;
2234 else if (exp->chan[i].const_float == 1)
2235 is_one[i] = true;
2236 else
2237 return false; /* other constant */
2238 } else
2239 return false;
2240 }
2241
2242 /* Only certain combinations of 0 and 1 can be eliminated. */
2243 if (is_zero[0] && is_zero[1] && is_zero[2])
2244 default_val = is_zero[3] ? 0 : 1;
2245 else if (is_one[0] && is_one[1] && is_one[2])
2246 default_val = is_zero[3] ? 2 : 3;
2247 else
2248 return false;
2249
2250 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
2251 LLVMInstructionEraseFromParent(exp->inst);
2252
2253 /* Change OFFSET to DEFAULT_VAL. */
2254 for (i = 0; i < num_outputs; i++) {
2255 if (vs_output_param_offset[i] == exp->offset) {
2256 vs_output_param_offset[i] =
2257 AC_EXP_PARAM_DEFAULT_VAL_0000 + default_val;
2258 break;
2259 }
2260 }
2261 return true;
2262 }
2263
2264 static bool ac_eliminate_duplicated_output(struct ac_llvm_context *ctx,
2265 uint8_t *vs_output_param_offset,
2266 uint32_t num_outputs,
2267 struct ac_vs_exports *processed,
2268 struct ac_vs_exp_inst *exp)
2269 {
2270 unsigned p, copy_back_channels = 0;
2271
2272 /* See if the output is already in the list of processed outputs.
2273 * The LLVMValueRef comparison relies on SSA.
2274 */
2275 for (p = 0; p < processed->num; p++) {
2276 bool different = false;
2277
2278 for (unsigned j = 0; j < 4; j++) {
2279 struct ac_vs_exp_chan *c1 = &processed->exp[p].chan[j];
2280 struct ac_vs_exp_chan *c2 = &exp->chan[j];
2281
2282 /* Treat undef as a match. */
2283 if (c2->type == AC_IR_UNDEF)
2284 continue;
2285
2286 /* If c1 is undef but c2 isn't, we can copy c2 to c1
2287 * and consider the instruction duplicated.
2288 */
2289 if (c1->type == AC_IR_UNDEF) {
2290 copy_back_channels |= 1 << j;
2291 continue;
2292 }
2293
2294 /* Test whether the channels are not equal. */
2295 if (c1->type != c2->type ||
2296 (c1->type == AC_IR_CONST &&
2297 c1->const_float != c2->const_float) ||
2298 (c1->type == AC_IR_VALUE &&
2299 c1->value != c2->value)) {
2300 different = true;
2301 break;
2302 }
2303 }
2304 if (!different)
2305 break;
2306
2307 copy_back_channels = 0;
2308 }
2309 if (p == processed->num)
2310 return false;
2311
2312 /* If a match was found, but the matching export has undef where the new
2313 * one has a normal value, copy the normal value to the undef channel.
2314 */
2315 struct ac_vs_exp_inst *match = &processed->exp[p];
2316
2317 /* Get current enabled channels mask. */
2318 LLVMValueRef arg = LLVMGetOperand(match->inst, AC_EXP_ENABLED_CHANNELS);
2319 unsigned enabled_channels = LLVMConstIntGetZExtValue(arg);
2320
2321 while (copy_back_channels) {
2322 unsigned chan = u_bit_scan(&copy_back_channels);
2323
2324 assert(match->chan[chan].type == AC_IR_UNDEF);
2325 LLVMSetOperand(match->inst, AC_EXP_OUT0 + chan,
2326 exp->chan[chan].value);
2327 match->chan[chan] = exp->chan[chan];
2328
2329 /* Update number of enabled channels because the original mask
2330 * is not always 0xf.
2331 */
2332 enabled_channels |= (1 << chan);
2333 LLVMSetOperand(match->inst, AC_EXP_ENABLED_CHANNELS,
2334 LLVMConstInt(ctx->i32, enabled_channels, 0));
2335 }
2336
2337 /* The PARAM export is duplicated. Kill it. */
2338 LLVMInstructionEraseFromParent(exp->inst);
2339
2340 /* Change OFFSET to the matching export. */
2341 for (unsigned i = 0; i < num_outputs; i++) {
2342 if (vs_output_param_offset[i] == exp->offset) {
2343 vs_output_param_offset[i] = match->offset;
2344 break;
2345 }
2346 }
2347 return true;
2348 }
2349
2350 void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
2351 LLVMValueRef main_fn,
2352 uint8_t *vs_output_param_offset,
2353 uint32_t num_outputs,
2354 uint8_t *num_param_exports)
2355 {
2356 LLVMBasicBlockRef bb;
2357 bool removed_any = false;
2358 struct ac_vs_exports exports;
2359
2360 exports.num = 0;
2361
2362 /* Process all LLVM instructions. */
2363 bb = LLVMGetFirstBasicBlock(main_fn);
2364 while (bb) {
2365 LLVMValueRef inst = LLVMGetFirstInstruction(bb);
2366
2367 while (inst) {
2368 LLVMValueRef cur = inst;
2369 inst = LLVMGetNextInstruction(inst);
2370 struct ac_vs_exp_inst exp;
2371
2372 if (LLVMGetInstructionOpcode(cur) != LLVMCall)
2373 continue;
2374
2375 LLVMValueRef callee = ac_llvm_get_called_value(cur);
2376
2377 if (!ac_llvm_is_function(callee))
2378 continue;
2379
2380 const char *name = LLVMGetValueName(callee);
2381 unsigned num_args = LLVMCountParams(callee);
2382
2383 /* Check if this is an export instruction. */
2384 if ((num_args != 9 && num_args != 8) ||
2385 (strcmp(name, "llvm.SI.export") &&
2386 strcmp(name, "llvm.amdgcn.exp.f32")))
2387 continue;
2388
2389 LLVMValueRef arg = LLVMGetOperand(cur, AC_EXP_TARGET);
2390 unsigned target = LLVMConstIntGetZExtValue(arg);
2391
2392 if (target < V_008DFC_SQ_EXP_PARAM)
2393 continue;
2394
2395 target -= V_008DFC_SQ_EXP_PARAM;
2396
2397 /* Parse the instruction. */
2398 memset(&exp, 0, sizeof(exp));
2399 exp.offset = target;
2400 exp.inst = cur;
2401
2402 for (unsigned i = 0; i < 4; i++) {
2403 LLVMValueRef v = LLVMGetOperand(cur, AC_EXP_OUT0 + i);
2404
2405 exp.chan[i].value = v;
2406
2407 if (LLVMIsUndef(v)) {
2408 exp.chan[i].type = AC_IR_UNDEF;
2409 } else if (LLVMIsAConstantFP(v)) {
2410 LLVMBool loses_info;
2411 exp.chan[i].type = AC_IR_CONST;
2412 exp.chan[i].const_float =
2413 LLVMConstRealGetDouble(v, &loses_info);
2414 } else {
2415 exp.chan[i].type = AC_IR_VALUE;
2416 }
2417 }
2418
2419 /* Eliminate constant and duplicated PARAM exports. */
2420 if (ac_eliminate_const_output(vs_output_param_offset,
2421 num_outputs, &exp) ||
2422 ac_eliminate_duplicated_output(ctx,
2423 vs_output_param_offset,
2424 num_outputs, &exports,
2425 &exp)) {
2426 removed_any = true;
2427 } else {
2428 exports.exp[exports.num++] = exp;
2429 }
2430 }
2431 bb = LLVMGetNextBasicBlock(bb);
2432 }
2433
2434 /* Remove holes in export memory due to removed PARAM exports.
2435 * This is done by renumbering all PARAM exports.
2436 */
2437 if (removed_any) {
2438 uint8_t old_offset[VARYING_SLOT_MAX];
2439 unsigned out, i;
2440
2441 /* Make a copy of the offsets. We need the old version while
2442 * we are modifying some of them. */
2443 memcpy(old_offset, vs_output_param_offset,
2444 sizeof(old_offset));
2445
2446 for (i = 0; i < exports.num; i++) {
2447 unsigned offset = exports.exp[i].offset;
2448
2449 /* Update vs_output_param_offset. Multiple outputs can
2450 * have the same offset.
2451 */
2452 for (out = 0; out < num_outputs; out++) {
2453 if (old_offset[out] == offset)
2454 vs_output_param_offset[out] = i;
2455 }
2456
2457 /* Change the PARAM offset in the instruction. */
2458 LLVMSetOperand(exports.exp[i].inst, AC_EXP_TARGET,
2459 LLVMConstInt(ctx->i32,
2460 V_008DFC_SQ_EXP_PARAM + i, 0));
2461 }
2462 *num_param_exports = exports.num;
2463 }
2464 }
2465
2466 void ac_init_exec_full_mask(struct ac_llvm_context *ctx)
2467 {
2468 LLVMValueRef full_mask = LLVMConstInt(ctx->i64, ~0ull, 0);
2469 ac_build_intrinsic(ctx,
2470 "llvm.amdgcn.init.exec", ctx->voidt,
2471 &full_mask, 1, AC_FUNC_ATTR_CONVERGENT);
2472 }
2473
2474 void ac_declare_lds_as_pointer(struct ac_llvm_context *ctx)
2475 {
2476 unsigned lds_size = ctx->chip_class >= CIK ? 65536 : 32768;
2477 ctx->lds = LLVMBuildIntToPtr(ctx->builder, ctx->i32_0,
2478 LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), AC_ADDR_SPACE_LDS),
2479 "lds");
2480 }
2481
2482 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
2483 LLVMValueRef dw_addr)
2484 {
2485 return ac_build_load(ctx, ctx->lds, dw_addr);
2486 }
2487
2488 void ac_lds_store(struct ac_llvm_context *ctx,
2489 LLVMValueRef dw_addr,
2490 LLVMValueRef value)
2491 {
2492 value = ac_to_integer(ctx, value);
2493 ac_build_indexed_store(ctx, ctx->lds,
2494 dw_addr, value);
2495 }
2496
2497 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
2498 LLVMTypeRef dst_type,
2499 LLVMValueRef src0)
2500 {
2501 unsigned src0_bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2502 const char *intrin_name;
2503 LLVMTypeRef type;
2504 LLVMValueRef zero;
2505
2506 switch (src0_bitsize) {
2507 case 64:
2508 intrin_name = "llvm.cttz.i64";
2509 type = ctx->i64;
2510 zero = ctx->i64_0;
2511 break;
2512 case 32:
2513 intrin_name = "llvm.cttz.i32";
2514 type = ctx->i32;
2515 zero = ctx->i32_0;
2516 break;
2517 case 16:
2518 intrin_name = "llvm.cttz.i16";
2519 type = ctx->i16;
2520 zero = ctx->i16_0;
2521 break;
2522 default:
2523 unreachable(!"invalid bitsize");
2524 }
2525
2526 LLVMValueRef params[2] = {
2527 src0,
2528
2529 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
2530 * add special code to check for x=0. The reason is that
2531 * the LLVM behavior for x=0 is different from what we
2532 * need here. However, LLVM also assumes that ffs(x) is
2533 * in [0, 31], but GLSL expects that ffs(0) = -1, so
2534 * a conditional assignment to handle 0 is still required.
2535 *
2536 * The hardware already implements the correct behavior.
2537 */
2538 ctx->i1true,
2539 };
2540
2541 LLVMValueRef lsb = ac_build_intrinsic(ctx, intrin_name, type,
2542 params, 2,
2543 AC_FUNC_ATTR_READNONE);
2544
2545 if (src0_bitsize == 64) {
2546 lsb = LLVMBuildTrunc(ctx->builder, lsb, ctx->i32, "");
2547 }
2548
2549 /* TODO: We need an intrinsic to skip this conditional. */
2550 /* Check for zero: */
2551 return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder,
2552 LLVMIntEQ, src0,
2553 zero, ""),
2554 LLVMConstInt(ctx->i32, -1, 0), lsb, "");
2555 }
2556
2557 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type)
2558 {
2559 return LLVMPointerType(LLVMArrayType(elem_type, 0),
2560 AC_ADDR_SPACE_CONST);
2561 }
2562
2563 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type)
2564 {
2565 return LLVMPointerType(LLVMArrayType(elem_type, 0),
2566 AC_ADDR_SPACE_CONST_32BIT);
2567 }
2568
2569 static struct ac_llvm_flow *
2570 get_current_flow(struct ac_llvm_context *ctx)
2571 {
2572 if (ctx->flow_depth > 0)
2573 return &ctx->flow[ctx->flow_depth - 1];
2574 return NULL;
2575 }
2576
2577 static struct ac_llvm_flow *
2578 get_innermost_loop(struct ac_llvm_context *ctx)
2579 {
2580 for (unsigned i = ctx->flow_depth; i > 0; --i) {
2581 if (ctx->flow[i - 1].loop_entry_block)
2582 return &ctx->flow[i - 1];
2583 }
2584 return NULL;
2585 }
2586
2587 static struct ac_llvm_flow *
2588 push_flow(struct ac_llvm_context *ctx)
2589 {
2590 struct ac_llvm_flow *flow;
2591
2592 if (ctx->flow_depth >= ctx->flow_depth_max) {
2593 unsigned new_max = MAX2(ctx->flow_depth << 1,
2594 AC_LLVM_INITIAL_CF_DEPTH);
2595
2596 ctx->flow = realloc(ctx->flow, new_max * sizeof(*ctx->flow));
2597 ctx->flow_depth_max = new_max;
2598 }
2599
2600 flow = &ctx->flow[ctx->flow_depth];
2601 ctx->flow_depth++;
2602
2603 flow->next_block = NULL;
2604 flow->loop_entry_block = NULL;
2605 return flow;
2606 }
2607
2608 static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base,
2609 int label_id)
2610 {
2611 char buf[32];
2612 snprintf(buf, sizeof(buf), "%s%d", base, label_id);
2613 LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
2614 }
2615
2616 /* Append a basic block at the level of the parent flow.
2617 */
2618 static LLVMBasicBlockRef append_basic_block(struct ac_llvm_context *ctx,
2619 const char *name)
2620 {
2621 assert(ctx->flow_depth >= 1);
2622
2623 if (ctx->flow_depth >= 2) {
2624 struct ac_llvm_flow *flow = &ctx->flow[ctx->flow_depth - 2];
2625
2626 return LLVMInsertBasicBlockInContext(ctx->context,
2627 flow->next_block, name);
2628 }
2629
2630 LLVMValueRef main_fn =
2631 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->builder));
2632 return LLVMAppendBasicBlockInContext(ctx->context, main_fn, name);
2633 }
2634
2635 /* Emit a branch to the given default target for the current block if
2636 * applicable -- that is, if the current block does not already contain a
2637 * branch from a break or continue.
2638 */
2639 static void emit_default_branch(LLVMBuilderRef builder,
2640 LLVMBasicBlockRef target)
2641 {
2642 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
2643 LLVMBuildBr(builder, target);
2644 }
2645
2646 void ac_build_bgnloop(struct ac_llvm_context *ctx, int label_id)
2647 {
2648 struct ac_llvm_flow *flow = push_flow(ctx);
2649 flow->loop_entry_block = append_basic_block(ctx, "LOOP");
2650 flow->next_block = append_basic_block(ctx, "ENDLOOP");
2651 set_basicblock_name(flow->loop_entry_block, "loop", label_id);
2652 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
2653 LLVMPositionBuilderAtEnd(ctx->builder, flow->loop_entry_block);
2654 }
2655
2656 void ac_build_break(struct ac_llvm_context *ctx)
2657 {
2658 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
2659 LLVMBuildBr(ctx->builder, flow->next_block);
2660 }
2661
2662 void ac_build_continue(struct ac_llvm_context *ctx)
2663 {
2664 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
2665 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
2666 }
2667
2668 void ac_build_else(struct ac_llvm_context *ctx, int label_id)
2669 {
2670 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
2671 LLVMBasicBlockRef endif_block;
2672
2673 assert(!current_branch->loop_entry_block);
2674
2675 endif_block = append_basic_block(ctx, "ENDIF");
2676 emit_default_branch(ctx->builder, endif_block);
2677
2678 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
2679 set_basicblock_name(current_branch->next_block, "else", label_id);
2680
2681 current_branch->next_block = endif_block;
2682 }
2683
2684 void ac_build_endif(struct ac_llvm_context *ctx, int label_id)
2685 {
2686 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
2687
2688 assert(!current_branch->loop_entry_block);
2689
2690 emit_default_branch(ctx->builder, current_branch->next_block);
2691 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
2692 set_basicblock_name(current_branch->next_block, "endif", label_id);
2693
2694 ctx->flow_depth--;
2695 }
2696
2697 void ac_build_endloop(struct ac_llvm_context *ctx, int label_id)
2698 {
2699 struct ac_llvm_flow *current_loop = get_current_flow(ctx);
2700
2701 assert(current_loop->loop_entry_block);
2702
2703 emit_default_branch(ctx->builder, current_loop->loop_entry_block);
2704
2705 LLVMPositionBuilderAtEnd(ctx->builder, current_loop->next_block);
2706 set_basicblock_name(current_loop->next_block, "endloop", label_id);
2707 ctx->flow_depth--;
2708 }
2709
2710 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id)
2711 {
2712 struct ac_llvm_flow *flow = push_flow(ctx);
2713 LLVMBasicBlockRef if_block;
2714
2715 if_block = append_basic_block(ctx, "IF");
2716 flow->next_block = append_basic_block(ctx, "ELSE");
2717 set_basicblock_name(if_block, "if", label_id);
2718 LLVMBuildCondBr(ctx->builder, cond, if_block, flow->next_block);
2719 LLVMPositionBuilderAtEnd(ctx->builder, if_block);
2720 }
2721
2722 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
2723 int label_id)
2724 {
2725 LLVMValueRef cond = LLVMBuildFCmp(ctx->builder, LLVMRealUNE,
2726 value, ctx->f32_0, "");
2727 ac_build_ifcc(ctx, cond, label_id);
2728 }
2729
2730 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
2731 int label_id)
2732 {
2733 LLVMValueRef cond = LLVMBuildICmp(ctx->builder, LLVMIntNE,
2734 ac_to_integer(ctx, value),
2735 ctx->i32_0, "");
2736 ac_build_ifcc(ctx, cond, label_id);
2737 }
2738
2739 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
2740 const char *name)
2741 {
2742 LLVMBuilderRef builder = ac->builder;
2743 LLVMBasicBlockRef current_block = LLVMGetInsertBlock(builder);
2744 LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
2745 LLVMBasicBlockRef first_block = LLVMGetEntryBasicBlock(function);
2746 LLVMValueRef first_instr = LLVMGetFirstInstruction(first_block);
2747 LLVMBuilderRef first_builder = LLVMCreateBuilderInContext(ac->context);
2748 LLVMValueRef res;
2749
2750 if (first_instr) {
2751 LLVMPositionBuilderBefore(first_builder, first_instr);
2752 } else {
2753 LLVMPositionBuilderAtEnd(first_builder, first_block);
2754 }
2755
2756 res = LLVMBuildAlloca(first_builder, type, name);
2757 LLVMDisposeBuilder(first_builder);
2758 return res;
2759 }
2760
2761 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac,
2762 LLVMTypeRef type, const char *name)
2763 {
2764 LLVMValueRef ptr = ac_build_alloca_undef(ac, type, name);
2765 LLVMBuildStore(ac->builder, LLVMConstNull(type), ptr);
2766 return ptr;
2767 }
2768
2769 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
2770 LLVMTypeRef type)
2771 {
2772 int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
2773 return LLVMBuildBitCast(ctx->builder, ptr,
2774 LLVMPointerType(type, addr_space), "");
2775 }
2776
2777 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
2778 unsigned count)
2779 {
2780 unsigned num_components = ac_get_llvm_num_components(value);
2781 if (count == num_components)
2782 return value;
2783
2784 LLVMValueRef masks[MAX2(count, 2)];
2785 masks[0] = ctx->i32_0;
2786 masks[1] = ctx->i32_1;
2787 for (unsigned i = 2; i < count; i++)
2788 masks[i] = LLVMConstInt(ctx->i32, i, false);
2789
2790 if (count == 1)
2791 return LLVMBuildExtractElement(ctx->builder, value, masks[0],
2792 "");
2793
2794 LLVMValueRef swizzle = LLVMConstVector(masks, count);
2795 return LLVMBuildShuffleVector(ctx->builder, value, value, swizzle, "");
2796 }
2797
2798 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
2799 unsigned rshift, unsigned bitwidth)
2800 {
2801 LLVMValueRef value = param;
2802 if (rshift)
2803 value = LLVMBuildLShr(ctx->builder, value,
2804 LLVMConstInt(ctx->i32, rshift, false), "");
2805
2806 if (rshift + bitwidth < 32) {
2807 unsigned mask = (1 << bitwidth) - 1;
2808 value = LLVMBuildAnd(ctx->builder, value,
2809 LLVMConstInt(ctx->i32, mask, false), "");
2810 }
2811 return value;
2812 }
2813
2814 /* Adjust the sample index according to FMASK.
2815 *
2816 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
2817 * which is the identity mapping. Each nibble says which physical sample
2818 * should be fetched to get that sample.
2819 *
2820 * For example, 0x11111100 means there are only 2 samples stored and
2821 * the second sample covers 3/4 of the pixel. When reading samples 0
2822 * and 1, return physical sample 0 (determined by the first two 0s
2823 * in FMASK), otherwise return physical sample 1.
2824 *
2825 * The sample index should be adjusted as follows:
2826 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
2827 */
2828 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
2829 LLVMValueRef *addr, bool is_array_tex)
2830 {
2831 struct ac_image_args fmask_load = {};
2832 fmask_load.opcode = ac_image_load;
2833 fmask_load.resource = fmask;
2834 fmask_load.dmask = 0xf;
2835 fmask_load.dim = is_array_tex ? ac_image_2darray : ac_image_2d;
2836
2837 fmask_load.coords[0] = addr[0];
2838 fmask_load.coords[1] = addr[1];
2839 if (is_array_tex)
2840 fmask_load.coords[2] = addr[2];
2841
2842 LLVMValueRef fmask_value = ac_build_image_opcode(ac, &fmask_load);
2843 fmask_value = LLVMBuildExtractElement(ac->builder, fmask_value,
2844 ac->i32_0, "");
2845
2846 /* Apply the formula. */
2847 unsigned sample_chan = is_array_tex ? 3 : 2;
2848 LLVMValueRef final_sample;
2849 final_sample = LLVMBuildMul(ac->builder, addr[sample_chan],
2850 LLVMConstInt(ac->i32, 4, 0), "");
2851 final_sample = LLVMBuildLShr(ac->builder, fmask_value, final_sample, "");
2852 /* Mask the sample index by 0x7, because 0x8 means an unknown value
2853 * with EQAA, so those will map to 0. */
2854 final_sample = LLVMBuildAnd(ac->builder, final_sample,
2855 LLVMConstInt(ac->i32, 0x7, 0), "");
2856
2857 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
2858 * resource descriptor is 0 (invalid).
2859 */
2860 LLVMValueRef tmp;
2861 tmp = LLVMBuildBitCast(ac->builder, fmask, ac->v8i32, "");
2862 tmp = LLVMBuildExtractElement(ac->builder, tmp, ac->i32_1, "");
2863 tmp = LLVMBuildICmp(ac->builder, LLVMIntNE, tmp, ac->i32_0, "");
2864
2865 /* Replace the MSAA sample index. */
2866 addr[sample_chan] = LLVMBuildSelect(ac->builder, tmp, final_sample,
2867 addr[sample_chan], "");
2868 }
2869
2870 static LLVMValueRef
2871 _ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
2872 {
2873 ac_build_optimization_barrier(ctx, &src);
2874 return ac_build_intrinsic(ctx,
2875 lane == NULL ? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
2876 LLVMTypeOf(src), (LLVMValueRef []) {
2877 src, lane },
2878 lane == NULL ? 1 : 2,
2879 AC_FUNC_ATTR_READNONE |
2880 AC_FUNC_ATTR_CONVERGENT);
2881 }
2882
2883 /**
2884 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
2885 * @param ctx
2886 * @param src
2887 * @param lane - id of the lane or NULL for the first active lane
2888 * @return value of the lane
2889 */
2890 LLVMValueRef
2891 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
2892 {
2893 LLVMTypeRef src_type = LLVMTypeOf(src);
2894 src = ac_to_integer(ctx, src);
2895 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
2896 LLVMValueRef ret;
2897
2898 if (bits == 32) {
2899 ret = _ac_build_readlane(ctx, src, lane);
2900 } else {
2901 assert(bits % 32 == 0);
2902 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
2903 LLVMValueRef src_vector =
2904 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
2905 ret = LLVMGetUndef(vec_type);
2906 for (unsigned i = 0; i < bits / 32; i++) {
2907 src = LLVMBuildExtractElement(ctx->builder, src_vector,
2908 LLVMConstInt(ctx->i32, i, 0), "");
2909 LLVMValueRef ret_comp = _ac_build_readlane(ctx, src, lane);
2910 ret = LLVMBuildInsertElement(ctx->builder, ret, ret_comp,
2911 LLVMConstInt(ctx->i32, i, 0), "");
2912 }
2913 }
2914 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
2915 }
2916
2917 LLVMValueRef
2918 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane)
2919 {
2920 /* TODO: Use the actual instruction when LLVM adds an intrinsic for it.
2921 */
2922 LLVMValueRef pred = LLVMBuildICmp(ctx->builder, LLVMIntEQ, lane,
2923 ac_get_thread_id(ctx), "");
2924 return LLVMBuildSelect(ctx->builder, pred, value, src, "");
2925 }
2926
2927 LLVMValueRef
2928 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask)
2929 {
2930 LLVMValueRef mask_vec = LLVMBuildBitCast(ctx->builder, mask,
2931 LLVMVectorType(ctx->i32, 2),
2932 "");
2933 LLVMValueRef mask_lo = LLVMBuildExtractElement(ctx->builder, mask_vec,
2934 ctx->i32_0, "");
2935 LLVMValueRef mask_hi = LLVMBuildExtractElement(ctx->builder, mask_vec,
2936 ctx->i32_1, "");
2937 LLVMValueRef val =
2938 ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
2939 (LLVMValueRef []) { mask_lo, ctx->i32_0 },
2940 2, AC_FUNC_ATTR_READNONE);
2941 val = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi", ctx->i32,
2942 (LLVMValueRef []) { mask_hi, val },
2943 2, AC_FUNC_ATTR_READNONE);
2944 return val;
2945 }
2946
2947 enum dpp_ctrl {
2948 _dpp_quad_perm = 0x000,
2949 _dpp_row_sl = 0x100,
2950 _dpp_row_sr = 0x110,
2951 _dpp_row_rr = 0x120,
2952 dpp_wf_sl1 = 0x130,
2953 dpp_wf_rl1 = 0x134,
2954 dpp_wf_sr1 = 0x138,
2955 dpp_wf_rr1 = 0x13C,
2956 dpp_row_mirror = 0x140,
2957 dpp_row_half_mirror = 0x141,
2958 dpp_row_bcast15 = 0x142,
2959 dpp_row_bcast31 = 0x143
2960 };
2961
2962 static inline enum dpp_ctrl
2963 dpp_quad_perm(unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
2964 {
2965 assert(lane0 < 4 && lane1 < 4 && lane2 < 4 && lane3 < 4);
2966 return _dpp_quad_perm | lane0 | (lane1 << 2) | (lane2 << 4) | (lane3 << 6);
2967 }
2968
2969 static inline enum dpp_ctrl
2970 dpp_row_sl(unsigned amount)
2971 {
2972 assert(amount > 0 && amount < 16);
2973 return _dpp_row_sl | amount;
2974 }
2975
2976 static inline enum dpp_ctrl
2977 dpp_row_sr(unsigned amount)
2978 {
2979 assert(amount > 0 && amount < 16);
2980 return _dpp_row_sr | amount;
2981 }
2982
2983 static LLVMValueRef
2984 _ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
2985 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
2986 bool bound_ctrl)
2987 {
2988 return ac_build_intrinsic(ctx, "llvm.amdgcn.update.dpp.i32",
2989 LLVMTypeOf(old),
2990 (LLVMValueRef[]) {
2991 old, src,
2992 LLVMConstInt(ctx->i32, dpp_ctrl, 0),
2993 LLVMConstInt(ctx->i32, row_mask, 0),
2994 LLVMConstInt(ctx->i32, bank_mask, 0),
2995 LLVMConstInt(ctx->i1, bound_ctrl, 0) },
2996 6, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
2997 }
2998
2999 static LLVMValueRef
3000 ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3001 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3002 bool bound_ctrl)
3003 {
3004 LLVMTypeRef src_type = LLVMTypeOf(src);
3005 src = ac_to_integer(ctx, src);
3006 old = ac_to_integer(ctx, old);
3007 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3008 LLVMValueRef ret;
3009 if (bits == 32) {
3010 ret = _ac_build_dpp(ctx, old, src, dpp_ctrl, row_mask,
3011 bank_mask, bound_ctrl);
3012 } else {
3013 assert(bits % 32 == 0);
3014 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3015 LLVMValueRef src_vector =
3016 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3017 LLVMValueRef old_vector =
3018 LLVMBuildBitCast(ctx->builder, old, vec_type, "");
3019 ret = LLVMGetUndef(vec_type);
3020 for (unsigned i = 0; i < bits / 32; i++) {
3021 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3022 LLVMConstInt(ctx->i32, i,
3023 0), "");
3024 old = LLVMBuildExtractElement(ctx->builder, old_vector,
3025 LLVMConstInt(ctx->i32, i,
3026 0), "");
3027 LLVMValueRef ret_comp = _ac_build_dpp(ctx, old, src,
3028 dpp_ctrl,
3029 row_mask,
3030 bank_mask,
3031 bound_ctrl);
3032 ret = LLVMBuildInsertElement(ctx->builder, ret,
3033 ret_comp,
3034 LLVMConstInt(ctx->i32, i,
3035 0), "");
3036 }
3037 }
3038 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3039 }
3040
3041 static inline unsigned
3042 ds_pattern_bitmode(unsigned and_mask, unsigned or_mask, unsigned xor_mask)
3043 {
3044 assert(and_mask < 32 && or_mask < 32 && xor_mask < 32);
3045 return and_mask | (or_mask << 5) | (xor_mask << 10);
3046 }
3047
3048 static LLVMValueRef
3049 _ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
3050 {
3051 return ac_build_intrinsic(ctx, "llvm.amdgcn.ds.swizzle",
3052 LLVMTypeOf(src), (LLVMValueRef []) {
3053 src, LLVMConstInt(ctx->i32, mask, 0) },
3054 2, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3055 }
3056
3057 LLVMValueRef
3058 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
3059 {
3060 LLVMTypeRef src_type = LLVMTypeOf(src);
3061 src = ac_to_integer(ctx, src);
3062 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3063 LLVMValueRef ret;
3064 if (bits == 32) {
3065 ret = _ac_build_ds_swizzle(ctx, src, mask);
3066 } else {
3067 assert(bits % 32 == 0);
3068 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3069 LLVMValueRef src_vector =
3070 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3071 ret = LLVMGetUndef(vec_type);
3072 for (unsigned i = 0; i < bits / 32; i++) {
3073 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3074 LLVMConstInt(ctx->i32, i,
3075 0), "");
3076 LLVMValueRef ret_comp = _ac_build_ds_swizzle(ctx, src,
3077 mask);
3078 ret = LLVMBuildInsertElement(ctx->builder, ret,
3079 ret_comp,
3080 LLVMConstInt(ctx->i32, i,
3081 0), "");
3082 }
3083 }
3084 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3085 }
3086
3087 static LLVMValueRef
3088 ac_build_wwm(struct ac_llvm_context *ctx, LLVMValueRef src)
3089 {
3090 char name[32], type[8];
3091 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
3092 snprintf(name, sizeof(name), "llvm.amdgcn.wwm.%s", type);
3093 return ac_build_intrinsic(ctx, name, LLVMTypeOf(src),
3094 (LLVMValueRef []) { src }, 1,
3095 AC_FUNC_ATTR_READNONE);
3096 }
3097
3098 static LLVMValueRef
3099 ac_build_set_inactive(struct ac_llvm_context *ctx, LLVMValueRef src,
3100 LLVMValueRef inactive)
3101 {
3102 char name[33], type[8];
3103 LLVMTypeRef src_type = LLVMTypeOf(src);
3104 src = ac_to_integer(ctx, src);
3105 inactive = ac_to_integer(ctx, inactive);
3106 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
3107 snprintf(name, sizeof(name), "llvm.amdgcn.set.inactive.%s", type);
3108 LLVMValueRef ret =
3109 ac_build_intrinsic(ctx, name,
3110 LLVMTypeOf(src), (LLVMValueRef []) {
3111 src, inactive }, 2,
3112 AC_FUNC_ATTR_READNONE |
3113 AC_FUNC_ATTR_CONVERGENT);
3114 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3115 }
3116
3117 static LLVMValueRef
3118 get_reduction_identity(struct ac_llvm_context *ctx, nir_op op, unsigned type_size)
3119 {
3120 if (type_size == 4) {
3121 switch (op) {
3122 case nir_op_iadd: return ctx->i32_0;
3123 case nir_op_fadd: return ctx->f32_0;
3124 case nir_op_imul: return ctx->i32_1;
3125 case nir_op_fmul: return ctx->f32_1;
3126 case nir_op_imin: return LLVMConstInt(ctx->i32, INT32_MAX, 0);
3127 case nir_op_umin: return LLVMConstInt(ctx->i32, UINT32_MAX, 0);
3128 case nir_op_fmin: return LLVMConstReal(ctx->f32, INFINITY);
3129 case nir_op_imax: return LLVMConstInt(ctx->i32, INT32_MIN, 0);
3130 case nir_op_umax: return ctx->i32_0;
3131 case nir_op_fmax: return LLVMConstReal(ctx->f32, -INFINITY);
3132 case nir_op_iand: return LLVMConstInt(ctx->i32, -1, 0);
3133 case nir_op_ior: return ctx->i32_0;
3134 case nir_op_ixor: return ctx->i32_0;
3135 default:
3136 unreachable("bad reduction intrinsic");
3137 }
3138 } else { /* type_size == 64bit */
3139 switch (op) {
3140 case nir_op_iadd: return ctx->i64_0;
3141 case nir_op_fadd: return ctx->f64_0;
3142 case nir_op_imul: return ctx->i64_1;
3143 case nir_op_fmul: return ctx->f64_1;
3144 case nir_op_imin: return LLVMConstInt(ctx->i64, INT64_MAX, 0);
3145 case nir_op_umin: return LLVMConstInt(ctx->i64, UINT64_MAX, 0);
3146 case nir_op_fmin: return LLVMConstReal(ctx->f64, INFINITY);
3147 case nir_op_imax: return LLVMConstInt(ctx->i64, INT64_MIN, 0);
3148 case nir_op_umax: return ctx->i64_0;
3149 case nir_op_fmax: return LLVMConstReal(ctx->f64, -INFINITY);
3150 case nir_op_iand: return LLVMConstInt(ctx->i64, -1, 0);
3151 case nir_op_ior: return ctx->i64_0;
3152 case nir_op_ixor: return ctx->i64_0;
3153 default:
3154 unreachable("bad reduction intrinsic");
3155 }
3156 }
3157 }
3158
3159 static LLVMValueRef
3160 ac_build_alu_op(struct ac_llvm_context *ctx, LLVMValueRef lhs, LLVMValueRef rhs, nir_op op)
3161 {
3162 bool _64bit = ac_get_type_size(LLVMTypeOf(lhs)) == 8;
3163 switch (op) {
3164 case nir_op_iadd: return LLVMBuildAdd(ctx->builder, lhs, rhs, "");
3165 case nir_op_fadd: return LLVMBuildFAdd(ctx->builder, lhs, rhs, "");
3166 case nir_op_imul: return LLVMBuildMul(ctx->builder, lhs, rhs, "");
3167 case nir_op_fmul: return LLVMBuildFMul(ctx->builder, lhs, rhs, "");
3168 case nir_op_imin: return LLVMBuildSelect(ctx->builder,
3169 LLVMBuildICmp(ctx->builder, LLVMIntSLT, lhs, rhs, ""),
3170 lhs, rhs, "");
3171 case nir_op_umin: return LLVMBuildSelect(ctx->builder,
3172 LLVMBuildICmp(ctx->builder, LLVMIntULT, lhs, rhs, ""),
3173 lhs, rhs, "");
3174 case nir_op_fmin: return ac_build_intrinsic(ctx,
3175 _64bit ? "llvm.minnum.f64" : "llvm.minnum.f32",
3176 _64bit ? ctx->f64 : ctx->f32,
3177 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
3178 case nir_op_imax: return LLVMBuildSelect(ctx->builder,
3179 LLVMBuildICmp(ctx->builder, LLVMIntSGT, lhs, rhs, ""),
3180 lhs, rhs, "");
3181 case nir_op_umax: return LLVMBuildSelect(ctx->builder,
3182 LLVMBuildICmp(ctx->builder, LLVMIntUGT, lhs, rhs, ""),
3183 lhs, rhs, "");
3184 case nir_op_fmax: return ac_build_intrinsic(ctx,
3185 _64bit ? "llvm.maxnum.f64" : "llvm.maxnum.f32",
3186 _64bit ? ctx->f64 : ctx->f32,
3187 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
3188 case nir_op_iand: return LLVMBuildAnd(ctx->builder, lhs, rhs, "");
3189 case nir_op_ior: return LLVMBuildOr(ctx->builder, lhs, rhs, "");
3190 case nir_op_ixor: return LLVMBuildXor(ctx->builder, lhs, rhs, "");
3191 default:
3192 unreachable("bad reduction intrinsic");
3193 }
3194 }
3195
3196 /**
3197 * \param maxprefix specifies that the result only needs to be correct for a
3198 * prefix of this many threads
3199 *
3200 * TODO: add inclusive and excluse scan functions for SI chip class.
3201 */
3202 static LLVMValueRef
3203 ac_build_scan(struct ac_llvm_context *ctx, nir_op op, LLVMValueRef src, LLVMValueRef identity,
3204 unsigned maxprefix)
3205 {
3206 LLVMValueRef result, tmp;
3207 result = src;
3208 if (maxprefix <= 1)
3209 return result;
3210 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
3211 result = ac_build_alu_op(ctx, result, tmp, op);
3212 if (maxprefix <= 2)
3213 return result;
3214 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(2), 0xf, 0xf, false);
3215 result = ac_build_alu_op(ctx, result, tmp, op);
3216 if (maxprefix <= 3)
3217 return result;
3218 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(3), 0xf, 0xf, false);
3219 result = ac_build_alu_op(ctx, result, tmp, op);
3220 if (maxprefix <= 4)
3221 return result;
3222 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(4), 0xf, 0xe, false);
3223 result = ac_build_alu_op(ctx, result, tmp, op);
3224 if (maxprefix <= 8)
3225 return result;
3226 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(8), 0xf, 0xc, false);
3227 result = ac_build_alu_op(ctx, result, tmp, op);
3228 if (maxprefix <= 16)
3229 return result;
3230 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
3231 result = ac_build_alu_op(ctx, result, tmp, op);
3232 if (maxprefix <= 32)
3233 return result;
3234 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
3235 result = ac_build_alu_op(ctx, result, tmp, op);
3236 return result;
3237 }
3238
3239 LLVMValueRef
3240 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
3241 {
3242 LLVMValueRef result;
3243
3244 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
3245 LLVMBuilderRef builder = ctx->builder;
3246 src = LLVMBuildZExt(builder, src, ctx->i32, "");
3247 result = ac_build_ballot(ctx, src);
3248 result = ac_build_mbcnt(ctx, result);
3249 result = LLVMBuildAdd(builder, result, src, "");
3250 return result;
3251 }
3252
3253 ac_build_optimization_barrier(ctx, &src);
3254
3255 LLVMValueRef identity =
3256 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
3257 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
3258 LLVMTypeOf(identity), "");
3259 result = ac_build_scan(ctx, op, result, identity, 64);
3260
3261 return ac_build_wwm(ctx, result);
3262 }
3263
3264 LLVMValueRef
3265 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
3266 {
3267 LLVMValueRef result;
3268
3269 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
3270 LLVMBuilderRef builder = ctx->builder;
3271 src = LLVMBuildZExt(builder, src, ctx->i32, "");
3272 result = ac_build_ballot(ctx, src);
3273 result = ac_build_mbcnt(ctx, result);
3274 return result;
3275 }
3276
3277 ac_build_optimization_barrier(ctx, &src);
3278
3279 LLVMValueRef identity =
3280 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
3281 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
3282 LLVMTypeOf(identity), "");
3283 result = ac_build_dpp(ctx, identity, result, dpp_wf_sr1, 0xf, 0xf, false);
3284 result = ac_build_scan(ctx, op, result, identity, 64);
3285
3286 return ac_build_wwm(ctx, result);
3287 }
3288
3289 LLVMValueRef
3290 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size)
3291 {
3292 if (cluster_size == 1) return src;
3293 ac_build_optimization_barrier(ctx, &src);
3294 LLVMValueRef result, swap;
3295 LLVMValueRef identity = get_reduction_identity(ctx, op,
3296 ac_get_type_size(LLVMTypeOf(src)));
3297 result = LLVMBuildBitCast(ctx->builder,
3298 ac_build_set_inactive(ctx, src, identity),
3299 LLVMTypeOf(identity), "");
3300 swap = ac_build_quad_swizzle(ctx, result, 1, 0, 3, 2);
3301 result = ac_build_alu_op(ctx, result, swap, op);
3302 if (cluster_size == 2) return ac_build_wwm(ctx, result);
3303
3304 swap = ac_build_quad_swizzle(ctx, result, 2, 3, 0, 1);
3305 result = ac_build_alu_op(ctx, result, swap, op);
3306 if (cluster_size == 4) return ac_build_wwm(ctx, result);
3307
3308 if (ctx->chip_class >= VI)
3309 swap = ac_build_dpp(ctx, identity, result, dpp_row_half_mirror, 0xf, 0xf, false);
3310 else
3311 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x04));
3312 result = ac_build_alu_op(ctx, result, swap, op);
3313 if (cluster_size == 8) return ac_build_wwm(ctx, result);
3314
3315 if (ctx->chip_class >= VI)
3316 swap = ac_build_dpp(ctx, identity, result, dpp_row_mirror, 0xf, 0xf, false);
3317 else
3318 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x08));
3319 result = ac_build_alu_op(ctx, result, swap, op);
3320 if (cluster_size == 16) return ac_build_wwm(ctx, result);
3321
3322 if (ctx->chip_class >= VI && cluster_size != 32)
3323 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
3324 else
3325 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x10));
3326 result = ac_build_alu_op(ctx, result, swap, op);
3327 if (cluster_size == 32) return ac_build_wwm(ctx, result);
3328
3329 if (ctx->chip_class >= VI) {
3330 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
3331 result = ac_build_alu_op(ctx, result, swap, op);
3332 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 63, 0));
3333 return ac_build_wwm(ctx, result);
3334 } else {
3335 swap = ac_build_readlane(ctx, result, ctx->i32_0);
3336 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 32, 0));
3337 result = ac_build_alu_op(ctx, result, swap, op);
3338 return ac_build_wwm(ctx, result);
3339 }
3340 }
3341
3342 /**
3343 * "Top half" of a scan that reduces per-wave values across an entire
3344 * workgroup.
3345 *
3346 * The source value must be present in the highest lane of the wave, and the
3347 * highest lane must be live.
3348 */
3349 void
3350 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3351 {
3352 if (ws->maxwaves <= 1)
3353 return;
3354
3355 const LLVMValueRef i32_63 = LLVMConstInt(ctx->i32, 63, false);
3356 LLVMBuilderRef builder = ctx->builder;
3357 LLVMValueRef tid = ac_get_thread_id(ctx);
3358 LLVMValueRef tmp;
3359
3360 tmp = LLVMBuildICmp(builder, LLVMIntEQ, tid, i32_63, "");
3361 ac_build_ifcc(ctx, tmp, 1000);
3362 LLVMBuildStore(builder, ws->src, LLVMBuildGEP(builder, ws->scratch, &ws->waveidx, 1, ""));
3363 ac_build_endif(ctx, 1000);
3364 }
3365
3366 /**
3367 * "Bottom half" of a scan that reduces per-wave values across an entire
3368 * workgroup.
3369 *
3370 * The caller must place a barrier between the top and bottom halves.
3371 */
3372 void
3373 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3374 {
3375 const LLVMTypeRef type = LLVMTypeOf(ws->src);
3376 const LLVMValueRef identity =
3377 get_reduction_identity(ctx, ws->op, ac_get_type_size(type));
3378
3379 if (ws->maxwaves <= 1) {
3380 ws->result_reduce = ws->src;
3381 ws->result_inclusive = ws->src;
3382 ws->result_exclusive = identity;
3383 return;
3384 }
3385 assert(ws->maxwaves <= 32);
3386
3387 LLVMBuilderRef builder = ctx->builder;
3388 LLVMValueRef tid = ac_get_thread_id(ctx);
3389 LLVMBasicBlockRef bbs[2];
3390 LLVMValueRef phivalues_scan[2];
3391 LLVMValueRef tmp, tmp2;
3392
3393 bbs[0] = LLVMGetInsertBlock(builder);
3394 phivalues_scan[0] = LLVMGetUndef(type);
3395
3396 if (ws->enable_reduce)
3397 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->numwaves, "");
3398 else if (ws->enable_inclusive)
3399 tmp = LLVMBuildICmp(builder, LLVMIntULE, tid, ws->waveidx, "");
3400 else
3401 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->waveidx, "");
3402 ac_build_ifcc(ctx, tmp, 1001);
3403 {
3404 tmp = LLVMBuildLoad(builder, LLVMBuildGEP(builder, ws->scratch, &tid, 1, ""), "");
3405
3406 ac_build_optimization_barrier(ctx, &tmp);
3407
3408 bbs[1] = LLVMGetInsertBlock(builder);
3409 phivalues_scan[1] = ac_build_scan(ctx, ws->op, tmp, identity, ws->maxwaves);
3410 }
3411 ac_build_endif(ctx, 1001);
3412
3413 const LLVMValueRef scan = ac_build_phi(ctx, type, 2, phivalues_scan, bbs);
3414
3415 if (ws->enable_reduce) {
3416 tmp = LLVMBuildSub(builder, ws->numwaves, ctx->i32_1, "");
3417 ws->result_reduce = ac_build_readlane(ctx, scan, tmp);
3418 }
3419 if (ws->enable_inclusive)
3420 ws->result_inclusive = ac_build_readlane(ctx, scan, ws->waveidx);
3421 if (ws->enable_exclusive) {
3422 tmp = LLVMBuildSub(builder, ws->waveidx, ctx->i32_1, "");
3423 tmp = ac_build_readlane(ctx, scan, tmp);
3424 tmp2 = LLVMBuildICmp(builder, LLVMIntEQ, ws->waveidx, ctx->i32_0, "");
3425 ws->result_exclusive = LLVMBuildSelect(builder, tmp2, identity, tmp, "");
3426 }
3427 }
3428
3429 /**
3430 * Inclusive scan of a per-wave value across an entire workgroup.
3431 *
3432 * This implies an s_barrier instruction.
3433 *
3434 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
3435 * of the workgroup are live. (This requirement cannot easily be relaxed in a
3436 * useful manner because of the barrier in the algorithm.)
3437 */
3438 void
3439 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3440 {
3441 ac_build_wg_wavescan_top(ctx, ws);
3442 ac_build_s_barrier(ctx);
3443 ac_build_wg_wavescan_bottom(ctx, ws);
3444 }
3445
3446 /**
3447 * "Top half" of a scan that reduces per-thread values across an entire
3448 * workgroup.
3449 *
3450 * All lanes must be active when this code runs.
3451 */
3452 void
3453 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3454 {
3455 if (ws->enable_exclusive) {
3456 ws->extra = ac_build_exclusive_scan(ctx, ws->src, ws->op);
3457 if (LLVMTypeOf(ws->src) == ctx->i1 && ws->op == nir_op_iadd)
3458 ws->src = LLVMBuildZExt(ctx->builder, ws->src, ctx->i32, "");
3459 ws->src = ac_build_alu_op(ctx, ws->extra, ws->src, ws->op);
3460 } else {
3461 ws->src = ac_build_inclusive_scan(ctx, ws->src, ws->op);
3462 }
3463
3464 bool enable_inclusive = ws->enable_inclusive;
3465 bool enable_exclusive = ws->enable_exclusive;
3466 ws->enable_inclusive = false;
3467 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
3468 ac_build_wg_wavescan_top(ctx, ws);
3469 ws->enable_inclusive = enable_inclusive;
3470 ws->enable_exclusive = enable_exclusive;
3471 }
3472
3473 /**
3474 * "Bottom half" of a scan that reduces per-thread values across an entire
3475 * workgroup.
3476 *
3477 * The caller must place a barrier between the top and bottom halves.
3478 */
3479 void
3480 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3481 {
3482 bool enable_inclusive = ws->enable_inclusive;
3483 bool enable_exclusive = ws->enable_exclusive;
3484 ws->enable_inclusive = false;
3485 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
3486 ac_build_wg_wavescan_bottom(ctx, ws);
3487 ws->enable_inclusive = enable_inclusive;
3488 ws->enable_exclusive = enable_exclusive;
3489
3490 /* ws->result_reduce is already the correct value */
3491 if (ws->enable_inclusive)
3492 ws->result_inclusive = ac_build_alu_op(ctx, ws->result_exclusive, ws->src, ws->op);
3493 if (ws->enable_exclusive)
3494 ws->result_exclusive = ac_build_alu_op(ctx, ws->result_exclusive, ws->extra, ws->op);
3495 }
3496
3497 /**
3498 * A scan that reduces per-thread values across an entire workgroup.
3499 *
3500 * The caller must ensure that all lanes are active when this code runs
3501 * (WWM is insufficient!), because there is an implied barrier.
3502 */
3503 void
3504 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
3505 {
3506 ac_build_wg_scan_top(ctx, ws);
3507 ac_build_s_barrier(ctx);
3508 ac_build_wg_scan_bottom(ctx, ws);
3509 }
3510
3511 LLVMValueRef
3512 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
3513 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
3514 {
3515 unsigned mask = dpp_quad_perm(lane0, lane1, lane2, lane3);
3516 if (ctx->chip_class >= VI) {
3517 return ac_build_dpp(ctx, src, src, mask, 0xf, 0xf, false);
3518 } else {
3519 return ac_build_ds_swizzle(ctx, src, (1 << 15) | mask);
3520 }
3521 }
3522
3523 LLVMValueRef
3524 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index)
3525 {
3526 index = LLVMBuildMul(ctx->builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
3527 return ac_build_intrinsic(ctx,
3528 "llvm.amdgcn.ds.bpermute", ctx->i32,
3529 (LLVMValueRef []) {index, src}, 2,
3530 AC_FUNC_ATTR_READNONE |
3531 AC_FUNC_ATTR_CONVERGENT);
3532 }