projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
ac/surface/gfx10: allow "rotated" micro mode
[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->v3f32 = LLVMVectorType(ctx->f32, 3);
88 ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
89 ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
90
91 ctx->i8_0 = LLVMConstInt(ctx->i8, 0, false);
92 ctx->i8_1 = LLVMConstInt(ctx->i8, 1, false);
93 ctx->i16_0 = LLVMConstInt(ctx->i16, 0, false);
94 ctx->i16_1 = LLVMConstInt(ctx->i16, 1, false);
95 ctx->i32_0 = LLVMConstInt(ctx->i32, 0, false);
96 ctx->i32_1 = LLVMConstInt(ctx->i32, 1, false);
97 ctx->i64_0 = LLVMConstInt(ctx->i64, 0, false);
98 ctx->i64_1 = LLVMConstInt(ctx->i64, 1, false);
99 ctx->f16_0 = LLVMConstReal(ctx->f16, 0.0);
100 ctx->f16_1 = LLVMConstReal(ctx->f16, 1.0);
101 ctx->f32_0 = LLVMConstReal(ctx->f32, 0.0);
102 ctx->f32_1 = LLVMConstReal(ctx->f32, 1.0);
103 ctx->f64_0 = LLVMConstReal(ctx->f64, 0.0);
104 ctx->f64_1 = LLVMConstReal(ctx->f64, 1.0);
105
106 ctx->i1false = LLVMConstInt(ctx->i1, 0, false);
107 ctx->i1true = LLVMConstInt(ctx->i1, 1, false);
108
109 ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context,
110 "range", 5);
111
112 ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context,
113 "invariant.load", 14);
114
115 ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->context, "fpmath", 6);
116
117 args[0] = LLVMConstReal(ctx->f32, 2.5);
118 ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->context, args, 1);
119
120 ctx->uniform_md_kind = LLVMGetMDKindIDInContext(ctx->context,
121 "amdgpu.uniform", 14);
122
123 ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
124 }
125
126 void
127 ac_llvm_context_dispose(struct ac_llvm_context *ctx)
128 {
129 free(ctx->flow);
130 ctx->flow = NULL;
131 ctx->flow_depth_max = 0;
132 }
133
134 int
135 ac_get_llvm_num_components(LLVMValueRef value)
136 {
137 LLVMTypeRef type = LLVMTypeOf(value);
138 unsigned num_components = LLVMGetTypeKind(type) == LLVMVectorTypeKind
139 ? LLVMGetVectorSize(type)
140 : 1;
141 return num_components;
142 }
143
144 LLVMValueRef
145 ac_llvm_extract_elem(struct ac_llvm_context *ac,
146 LLVMValueRef value,
147 int index)
148 {
149 if (LLVMGetTypeKind(LLVMTypeOf(value)) != LLVMVectorTypeKind) {
150 assert(index == 0);
151 return value;
152 }
153
154 return LLVMBuildExtractElement(ac->builder, value,
155 LLVMConstInt(ac->i32, index, false), "");
156 }
157
158 int
159 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type)
160 {
161 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
162 type = LLVMGetElementType(type);
163
164 if (LLVMGetTypeKind(type) == LLVMIntegerTypeKind)
165 return LLVMGetIntTypeWidth(type);
166
167 if (type == ctx->f16)
168 return 16;
169 if (type == ctx->f32)
170 return 32;
171 if (type == ctx->f64)
172 return 64;
173
174 unreachable("Unhandled type kind in get_elem_bits");
175 }
176
177 unsigned
178 ac_get_type_size(LLVMTypeRef type)
179 {
180 LLVMTypeKind kind = LLVMGetTypeKind(type);
181
182 switch (kind) {
183 case LLVMIntegerTypeKind:
184 return LLVMGetIntTypeWidth(type) / 8;
185 case LLVMHalfTypeKind:
186 return 2;
187 case LLVMFloatTypeKind:
188 return 4;
189 case LLVMDoubleTypeKind:
190 return 8;
191 case LLVMPointerTypeKind:
192 if (LLVMGetPointerAddressSpace(type) == AC_ADDR_SPACE_CONST_32BIT)
193 return 4;
194 return 8;
195 case LLVMVectorTypeKind:
196 return LLVMGetVectorSize(type) *
197 ac_get_type_size(LLVMGetElementType(type));
198 case LLVMArrayTypeKind:
199 return LLVMGetArrayLength(type) *
200 ac_get_type_size(LLVMGetElementType(type));
201 default:
202 assert(0);
203 return 0;
204 }
205 }
206
207 static LLVMTypeRef to_integer_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
208 {
209 if (t == ctx->i8)
210 return ctx->i8;
211 else if (t == ctx->f16 || t == ctx->i16)
212 return ctx->i16;
213 else if (t == ctx->f32 || t == ctx->i32)
214 return ctx->i32;
215 else if (t == ctx->f64 || t == ctx->i64)
216 return ctx->i64;
217 else
218 unreachable("Unhandled integer size");
219 }
220
221 LLVMTypeRef
222 ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
223 {
224 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
225 LLVMTypeRef elem_type = LLVMGetElementType(t);
226 return LLVMVectorType(to_integer_type_scalar(ctx, elem_type),
227 LLVMGetVectorSize(t));
228 }
229 if (LLVMGetTypeKind(t) == LLVMPointerTypeKind) {
230 switch (LLVMGetPointerAddressSpace(t)) {
231 case AC_ADDR_SPACE_GLOBAL:
232 return ctx->i64;
233 case AC_ADDR_SPACE_LDS:
234 return ctx->i32;
235 default:
236 unreachable("unhandled address space");
237 }
238 }
239 return to_integer_type_scalar(ctx, t);
240 }
241
242 LLVMValueRef
243 ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v)
244 {
245 LLVMTypeRef type = LLVMTypeOf(v);
246 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
247 return LLVMBuildPtrToInt(ctx->builder, v, ac_to_integer_type(ctx, type), "");
248 }
249 return LLVMBuildBitCast(ctx->builder, v, ac_to_integer_type(ctx, type), "");
250 }
251
252 LLVMValueRef
253 ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v)
254 {
255 LLVMTypeRef type = LLVMTypeOf(v);
256 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind)
257 return v;
258 return ac_to_integer(ctx, v);
259 }
260
261 static LLVMTypeRef to_float_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
262 {
263 if (t == ctx->i8)
264 return ctx->i8;
265 else if (t == ctx->i16 || t == ctx->f16)
266 return ctx->f16;
267 else if (t == ctx->i32 || t == ctx->f32)
268 return ctx->f32;
269 else if (t == ctx->i64 || t == ctx->f64)
270 return ctx->f64;
271 else
272 unreachable("Unhandled float size");
273 }
274
275 LLVMTypeRef
276 ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
277 {
278 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
279 LLVMTypeRef elem_type = LLVMGetElementType(t);
280 return LLVMVectorType(to_float_type_scalar(ctx, elem_type),
281 LLVMGetVectorSize(t));
282 }
283 return to_float_type_scalar(ctx, t);
284 }
285
286 LLVMValueRef
287 ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v)
288 {
289 LLVMTypeRef type = LLVMTypeOf(v);
290 return LLVMBuildBitCast(ctx->builder, v, ac_to_float_type(ctx, type), "");
291 }
292
293
294 LLVMValueRef
295 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
296 LLVMTypeRef return_type, LLVMValueRef *params,
297 unsigned param_count, unsigned attrib_mask)
298 {
299 LLVMValueRef function, call;
300 bool set_callsite_attrs = !(attrib_mask & AC_FUNC_ATTR_LEGACY);
301
302 function = LLVMGetNamedFunction(ctx->module, name);
303 if (!function) {
304 LLVMTypeRef param_types[32], function_type;
305 unsigned i;
306
307 assert(param_count <= 32);
308
309 for (i = 0; i < param_count; ++i) {
310 assert(params[i]);
311 param_types[i] = LLVMTypeOf(params[i]);
312 }
313 function_type =
314 LLVMFunctionType(return_type, param_types, param_count, 0);
315 function = LLVMAddFunction(ctx->module, name, function_type);
316
317 LLVMSetFunctionCallConv(function, LLVMCCallConv);
318 LLVMSetLinkage(function, LLVMExternalLinkage);
319
320 if (!set_callsite_attrs)
321 ac_add_func_attributes(ctx->context, function, attrib_mask);
322 }
323
324 call = LLVMBuildCall(ctx->builder, function, params, param_count, "");
325 if (set_callsite_attrs)
326 ac_add_func_attributes(ctx->context, call, attrib_mask);
327 return call;
328 }
329
330 /**
331 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
332 * intrinsic names).
333 */
334 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize)
335 {
336 LLVMTypeRef elem_type = type;
337
338 assert(bufsize >= 8);
339
340 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
341 int ret = snprintf(buf, bufsize, "v%u",
342 LLVMGetVectorSize(type));
343 if (ret < 0) {
344 char *type_name = LLVMPrintTypeToString(type);
345 fprintf(stderr, "Error building type name for: %s\n",
346 type_name);
347 return;
348 }
349 elem_type = LLVMGetElementType(type);
350 buf += ret;
351 bufsize -= ret;
352 }
353 switch (LLVMGetTypeKind(elem_type)) {
354 default: break;
355 case LLVMIntegerTypeKind:
356 snprintf(buf, bufsize, "i%d", LLVMGetIntTypeWidth(elem_type));
357 break;
358 case LLVMHalfTypeKind:
359 snprintf(buf, bufsize, "f16");
360 break;
361 case LLVMFloatTypeKind:
362 snprintf(buf, bufsize, "f32");
363 break;
364 case LLVMDoubleTypeKind:
365 snprintf(buf, bufsize, "f64");
366 break;
367 }
368 }
369
370 /**
371 * Helper function that builds an LLVM IR PHI node and immediately adds
372 * incoming edges.
373 */
374 LLVMValueRef
375 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
376 unsigned count_incoming, LLVMValueRef *values,
377 LLVMBasicBlockRef *blocks)
378 {
379 LLVMValueRef phi = LLVMBuildPhi(ctx->builder, type, "");
380 LLVMAddIncoming(phi, values, blocks, count_incoming);
381 return phi;
382 }
383
384 void ac_build_s_barrier(struct ac_llvm_context *ctx)
385 {
386 ac_build_intrinsic(ctx, "llvm.amdgcn.s.barrier", ctx->voidt, NULL,
387 0, AC_FUNC_ATTR_CONVERGENT);
388 }
389
390 /* Prevent optimizations (at least of memory accesses) across the current
391 * point in the program by emitting empty inline assembly that is marked as
392 * having side effects.
393 *
394 * Optionally, a value can be passed through the inline assembly to prevent
395 * LLVM from hoisting calls to ReadNone functions.
396 */
397 void
398 ac_build_optimization_barrier(struct ac_llvm_context *ctx,
399 LLVMValueRef *pvgpr)
400 {
401 static int counter = 0;
402
403 LLVMBuilderRef builder = ctx->builder;
404 char code[16];
405
406 snprintf(code, sizeof(code), "; %d", p_atomic_inc_return(&counter));
407
408 if (!pvgpr) {
409 LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
410 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "", true, false);
411 LLVMBuildCall(builder, inlineasm, NULL, 0, "");
412 } else {
413 LLVMTypeRef ftype = LLVMFunctionType(ctx->i32, &ctx->i32, 1, false);
414 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "=v,0", true, false);
415 LLVMValueRef vgpr = *pvgpr;
416 LLVMTypeRef vgpr_type = LLVMTypeOf(vgpr);
417 unsigned vgpr_size = ac_get_type_size(vgpr_type);
418 LLVMValueRef vgpr0;
419
420 assert(vgpr_size % 4 == 0);
421
422 vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
423 vgpr0 = LLVMBuildExtractElement(builder, vgpr, ctx->i32_0, "");
424 vgpr0 = LLVMBuildCall(builder, inlineasm, &vgpr0, 1, "");
425 vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
426 vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
427
428 *pvgpr = vgpr;
429 }
430 }
431
432 LLVMValueRef
433 ac_build_shader_clock(struct ac_llvm_context *ctx)
434 {
435 LLVMValueRef tmp = ac_build_intrinsic(ctx, "llvm.readcyclecounter",
436 ctx->i64, NULL, 0, 0);
437 return LLVMBuildBitCast(ctx->builder, tmp, ctx->v2i32, "");
438 }
439
440 LLVMValueRef
441 ac_build_ballot(struct ac_llvm_context *ctx,
442 LLVMValueRef value)
443 {
444 const char *name = HAVE_LLVM >= 0x900 ? "llvm.amdgcn.icmp.i64.i32" : "llvm.amdgcn.icmp.i32";
445 LLVMValueRef args[3] = {
446 value,
447 ctx->i32_0,
448 LLVMConstInt(ctx->i32, LLVMIntNE, 0)
449 };
450
451 /* We currently have no other way to prevent LLVM from lifting the icmp
452 * calls to a dominating basic block.
453 */
454 ac_build_optimization_barrier(ctx, &args[0]);
455
456 args[0] = ac_to_integer(ctx, args[0]);
457
458 return ac_build_intrinsic(ctx, name,
459 ctx->i64, args, 3,
460 AC_FUNC_ATTR_NOUNWIND |
461 AC_FUNC_ATTR_READNONE |
462 AC_FUNC_ATTR_CONVERGENT);
463 }
464
465 LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
466 LLVMValueRef value)
467 {
468 const char *name = HAVE_LLVM >= 0x900 ? "llvm.amdgcn.icmp.i64.i1" : "llvm.amdgcn.icmp.i1";
469 LLVMValueRef args[3] = {
470 value,
471 ctx->i1false,
472 LLVMConstInt(ctx->i32, LLVMIntNE, 0),
473 };
474
475 assert(HAVE_LLVM >= 0x0800);
476 return ac_build_intrinsic(ctx, name, ctx->i64, args, 3,
477 AC_FUNC_ATTR_NOUNWIND |
478 AC_FUNC_ATTR_READNONE |
479 AC_FUNC_ATTR_CONVERGENT);
480 }
481
482 LLVMValueRef
483 ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value)
484 {
485 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
486 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
487 return LLVMBuildICmp(ctx->builder, LLVMIntEQ, vote_set, active_set, "");
488 }
489
490 LLVMValueRef
491 ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value)
492 {
493 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
494 return LLVMBuildICmp(ctx->builder, LLVMIntNE, vote_set,
495 LLVMConstInt(ctx->i64, 0, 0), "");
496 }
497
498 LLVMValueRef
499 ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value)
500 {
501 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
502 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
503
504 LLVMValueRef all = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
505 vote_set, active_set, "");
506 LLVMValueRef none = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
507 vote_set,
508 LLVMConstInt(ctx->i64, 0, 0), "");
509 return LLVMBuildOr(ctx->builder, all, none, "");
510 }
511
512 LLVMValueRef
513 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
514 unsigned value_count, unsigned component)
515 {
516 LLVMValueRef vec = NULL;
517
518 if (value_count == 1) {
519 return values[component];
520 } else if (!value_count)
521 unreachable("value_count is 0");
522
523 for (unsigned i = component; i < value_count + component; i++) {
524 LLVMValueRef value = values[i];
525
526 if (i == component)
527 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
528 LLVMValueRef index = LLVMConstInt(ctx->i32, i - component, false);
529 vec = LLVMBuildInsertElement(ctx->builder, vec, value, index, "");
530 }
531 return vec;
532 }
533
534 LLVMValueRef
535 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
536 LLVMValueRef *values,
537 unsigned value_count,
538 unsigned value_stride,
539 bool load,
540 bool always_vector)
541 {
542 LLVMBuilderRef builder = ctx->builder;
543 LLVMValueRef vec = NULL;
544 unsigned i;
545
546 if (value_count == 1 && !always_vector) {
547 if (load)
548 return LLVMBuildLoad(builder, values[0], "");
549 return values[0];
550 } else if (!value_count)
551 unreachable("value_count is 0");
552
553 for (i = 0; i < value_count; i++) {
554 LLVMValueRef value = values[i * value_stride];
555 if (load)
556 value = LLVMBuildLoad(builder, value, "");
557
558 if (!i)
559 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
560 LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
561 vec = LLVMBuildInsertElement(builder, vec, value, index, "");
562 }
563 return vec;
564 }
565
566 LLVMValueRef
567 ac_build_gather_values(struct ac_llvm_context *ctx,
568 LLVMValueRef *values,
569 unsigned value_count)
570 {
571 return ac_build_gather_values_extended(ctx, values, value_count, 1, false, false);
572 }
573
574 /* Expand a scalar or vector to <dst_channels x type> by filling the remaining
575 * channels with undef. Extract at most src_channels components from the input.
576 */
577 static LLVMValueRef
578 ac_build_expand(struct ac_llvm_context *ctx,
579 LLVMValueRef value,
580 unsigned src_channels,
581 unsigned dst_channels)
582 {
583 LLVMTypeRef elemtype;
584 LLVMValueRef chan[dst_channels];
585
586 if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMVectorTypeKind) {
587 unsigned vec_size = LLVMGetVectorSize(LLVMTypeOf(value));
588
589 if (src_channels == dst_channels && vec_size == dst_channels)
590 return value;
591
592 src_channels = MIN2(src_channels, vec_size);
593
594 for (unsigned i = 0; i < src_channels; i++)
595 chan[i] = ac_llvm_extract_elem(ctx, value, i);
596
597 elemtype = LLVMGetElementType(LLVMTypeOf(value));
598 } else {
599 if (src_channels) {
600 assert(src_channels == 1);
601 chan[0] = value;
602 }
603 elemtype = LLVMTypeOf(value);
604 }
605
606 for (unsigned i = src_channels; i < dst_channels; i++)
607 chan[i] = LLVMGetUndef(elemtype);
608
609 return ac_build_gather_values(ctx, chan, dst_channels);
610 }
611
612 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
613 * with undef. Extract at most num_channels components from the input.
614 */
615 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
616 LLVMValueRef value,
617 unsigned num_channels)
618 {
619 return ac_build_expand(ctx, value, num_channels, 4);
620 }
621
622 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value)
623 {
624 unsigned type_size = ac_get_type_size(LLVMTypeOf(value));
625 const char *name;
626
627 if (type_size == 2)
628 name = "llvm.rint.f16";
629 else if (type_size == 4)
630 name = "llvm.rint.f32";
631 else
632 name = "llvm.rint.f64";
633
634 return ac_build_intrinsic(ctx, name, LLVMTypeOf(value), &value, 1,
635 AC_FUNC_ATTR_READNONE);
636 }
637
638 LLVMValueRef
639 ac_build_fdiv(struct ac_llvm_context *ctx,
640 LLVMValueRef num,
641 LLVMValueRef den)
642 {
643 /* If we do (num / den), LLVM >= 7.0 does:
644 * return num * v_rcp_f32(den * (fabs(den) > 0x1.0p+96f ? 0x1.0p-32f : 1.0f));
645 *
646 * If we do (num * (1 / den)), LLVM does:
647 * return num * v_rcp_f32(den);
648 */
649 LLVMValueRef one = LLVMConstReal(LLVMTypeOf(num), 1.0);
650 LLVMValueRef rcp = LLVMBuildFDiv(ctx->builder, one, den, "");
651 LLVMValueRef ret = LLVMBuildFMul(ctx->builder, num, rcp, "");
652
653 /* Use v_rcp_f32 instead of precise division. */
654 if (!LLVMIsConstant(ret))
655 LLVMSetMetadata(ret, ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
656 return ret;
657 }
658
659 /* See fast_idiv_by_const.h. */
660 /* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
661 LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
662 LLVMValueRef num,
663 LLVMValueRef multiplier,
664 LLVMValueRef pre_shift,
665 LLVMValueRef post_shift,
666 LLVMValueRef increment)
667 {
668 LLVMBuilderRef builder = ctx->builder;
669
670 num = LLVMBuildLShr(builder, num, pre_shift, "");
671 num = LLVMBuildMul(builder,
672 LLVMBuildZExt(builder, num, ctx->i64, ""),
673 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
674 num = LLVMBuildAdd(builder, num,
675 LLVMBuildZExt(builder, increment, ctx->i64, ""), "");
676 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
677 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
678 return LLVMBuildLShr(builder, num, post_shift, "");
679 }
680
681 /* See fast_idiv_by_const.h. */
682 /* If num != UINT_MAX, this more efficient version can be used. */
683 /* Set: increment = util_fast_udiv_info::increment; */
684 LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
685 LLVMValueRef num,
686 LLVMValueRef multiplier,
687 LLVMValueRef pre_shift,
688 LLVMValueRef post_shift,
689 LLVMValueRef increment)
690 {
691 LLVMBuilderRef builder = ctx->builder;
692
693 num = LLVMBuildLShr(builder, num, pre_shift, "");
694 num = LLVMBuildNUWAdd(builder, num, increment, "");
695 num = LLVMBuildMul(builder,
696 LLVMBuildZExt(builder, num, ctx->i64, ""),
697 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
698 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
699 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
700 return LLVMBuildLShr(builder, num, post_shift, "");
701 }
702
703 /* See fast_idiv_by_const.h. */
704 /* Both operands must fit in 31 bits and the divisor must not be 1. */
705 LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
706 LLVMValueRef num,
707 LLVMValueRef multiplier,
708 LLVMValueRef post_shift)
709 {
710 LLVMBuilderRef builder = ctx->builder;
711
712 num = LLVMBuildMul(builder,
713 LLVMBuildZExt(builder, num, ctx->i64, ""),
714 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
715 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
716 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
717 return LLVMBuildLShr(builder, num, post_shift, "");
718 }
719
720 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
721 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
722 * already multiplied by two. id is the cube face number.
723 */
724 struct cube_selection_coords {
725 LLVMValueRef stc[2];
726 LLVMValueRef ma;
727 LLVMValueRef id;
728 };
729
730 static void
731 build_cube_intrinsic(struct ac_llvm_context *ctx,
732 LLVMValueRef in[3],
733 struct cube_selection_coords *out)
734 {
735 LLVMTypeRef f32 = ctx->f32;
736
737 out->stc[1] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubetc",
738 f32, in, 3, AC_FUNC_ATTR_READNONE);
739 out->stc[0] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubesc",
740 f32, in, 3, AC_FUNC_ATTR_READNONE);
741 out->ma = ac_build_intrinsic(ctx, "llvm.amdgcn.cubema",
742 f32, in, 3, AC_FUNC_ATTR_READNONE);
743 out->id = ac_build_intrinsic(ctx, "llvm.amdgcn.cubeid",
744 f32, in, 3, AC_FUNC_ATTR_READNONE);
745 }
746
747 /**
748 * Build a manual selection sequence for cube face sc/tc coordinates and
749 * major axis vector (multiplied by 2 for consistency) for the given
750 * vec3 \p coords, for the face implied by \p selcoords.
751 *
752 * For the major axis, we always adjust the sign to be in the direction of
753 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
754 * the selcoords major axis.
755 */
756 static void build_cube_select(struct ac_llvm_context *ctx,
757 const struct cube_selection_coords *selcoords,
758 const LLVMValueRef *coords,
759 LLVMValueRef *out_st,
760 LLVMValueRef *out_ma)
761 {
762 LLVMBuilderRef builder = ctx->builder;
763 LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
764 LLVMValueRef is_ma_positive;
765 LLVMValueRef sgn_ma;
766 LLVMValueRef is_ma_z, is_not_ma_z;
767 LLVMValueRef is_ma_y;
768 LLVMValueRef is_ma_x;
769 LLVMValueRef sgn;
770 LLVMValueRef tmp;
771
772 is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
773 selcoords->ma, LLVMConstReal(f32, 0.0), "");
774 sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
775 LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
776
777 is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
778 is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
779 is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
780 LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
781 is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
782
783 /* Select sc */
784 tmp = LLVMBuildSelect(builder, is_ma_x, coords[2], coords[0], "");
785 sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
786 LLVMBuildSelect(builder, is_ma_z, sgn_ma,
787 LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
788 out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
789
790 /* Select tc */
791 tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
792 sgn = LLVMBuildSelect(builder, is_ma_y, sgn_ma,
793 LLVMConstReal(f32, -1.0), "");
794 out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
795
796 /* Select ma */
797 tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
798 LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
799 tmp = ac_build_intrinsic(ctx, "llvm.fabs.f32",
800 ctx->f32, &tmp, 1, AC_FUNC_ATTR_READNONE);
801 *out_ma = LLVMBuildFMul(builder, tmp, LLVMConstReal(f32, 2.0), "");
802 }
803
804 void
805 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
806 bool is_deriv, bool is_array, bool is_lod,
807 LLVMValueRef *coords_arg,
808 LLVMValueRef *derivs_arg)
809 {
810
811 LLVMBuilderRef builder = ctx->builder;
812 struct cube_selection_coords selcoords;
813 LLVMValueRef coords[3];
814 LLVMValueRef invma;
815
816 if (is_array && !is_lod) {
817 LLVMValueRef tmp = ac_build_round(ctx, coords_arg[3]);
818
819 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
820 *
821 * "For Array forms, the array layer used will be
822 *
823 * max(0, min(d−1, floor(layer+0.5)))
824 *
825 * where d is the depth of the texture array and layer
826 * comes from the component indicated in the tables below.
827 * Workaroudn for an issue where the layer is taken from a
828 * helper invocation which happens to fall on a different
829 * layer due to extrapolation."
830 *
831 * GFX8 and earlier attempt to implement this in hardware by
832 * clamping the value of coords[2] = (8 * layer) + face.
833 * Unfortunately, this means that the we end up with the wrong
834 * face when clamping occurs.
835 *
836 * Clamp the layer earlier to work around the issue.
837 */
838 if (ctx->chip_class <= GFX8) {
839 LLVMValueRef ge0;
840 ge0 = LLVMBuildFCmp(builder, LLVMRealOGE, tmp, ctx->f32_0, "");
841 tmp = LLVMBuildSelect(builder, ge0, tmp, ctx->f32_0, "");
842 }
843
844 coords_arg[3] = tmp;
845 }
846
847 build_cube_intrinsic(ctx, coords_arg, &selcoords);
848
849 invma = ac_build_intrinsic(ctx, "llvm.fabs.f32",
850 ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
851 invma = ac_build_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
852
853 for (int i = 0; i < 2; ++i)
854 coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
855
856 coords[2] = selcoords.id;
857
858 if (is_deriv && derivs_arg) {
859 LLVMValueRef derivs[4];
860 int axis;
861
862 /* Convert cube derivatives to 2D derivatives. */
863 for (axis = 0; axis < 2; axis++) {
864 LLVMValueRef deriv_st[2];
865 LLVMValueRef deriv_ma;
866
867 /* Transform the derivative alongside the texture
868 * coordinate. Mathematically, the correct formula is
869 * as follows. Assume we're projecting onto the +Z face
870 * and denote by dx/dh the derivative of the (original)
871 * X texture coordinate with respect to horizontal
872 * window coordinates. The projection onto the +Z face
873 * plane is:
874 *
875 * f(x,z) = x/z
876 *
877 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
878 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
879 *
880 * This motivatives the implementation below.
881 *
882 * Whether this actually gives the expected results for
883 * apps that might feed in derivatives obtained via
884 * finite differences is anyone's guess. The OpenGL spec
885 * seems awfully quiet about how textureGrad for cube
886 * maps should be handled.
887 */
888 build_cube_select(ctx, &selcoords, &derivs_arg[axis * 3],
889 deriv_st, &deriv_ma);
890
891 deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
892
893 for (int i = 0; i < 2; ++i)
894 derivs[axis * 2 + i] =
895 LLVMBuildFSub(builder,
896 LLVMBuildFMul(builder, deriv_st[i], invma, ""),
897 LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
898 }
899
900 memcpy(derivs_arg, derivs, sizeof(derivs));
901 }
902
903 /* Shift the texture coordinate. This must be applied after the
904 * derivative calculation.
905 */
906 for (int i = 0; i < 2; ++i)
907 coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
908
909 if (is_array) {
910 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
911 /* coords_arg.w component - array_index for cube arrays */
912 coords[2] = ac_build_fmad(ctx, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), coords[2]);
913 }
914
915 memcpy(coords_arg, coords, sizeof(coords));
916 }
917
918
919 LLVMValueRef
920 ac_build_fs_interp(struct ac_llvm_context *ctx,
921 LLVMValueRef llvm_chan,
922 LLVMValueRef attr_number,
923 LLVMValueRef params,
924 LLVMValueRef i,
925 LLVMValueRef j)
926 {
927 LLVMValueRef args[5];
928 LLVMValueRef p1;
929
930 args[0] = i;
931 args[1] = llvm_chan;
932 args[2] = attr_number;
933 args[3] = params;
934
935 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1",
936 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
937
938 args[0] = p1;
939 args[1] = j;
940 args[2] = llvm_chan;
941 args[3] = attr_number;
942 args[4] = params;
943
944 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2",
945 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
946 }
947
948 LLVMValueRef
949 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
950 LLVMValueRef llvm_chan,
951 LLVMValueRef attr_number,
952 LLVMValueRef params,
953 LLVMValueRef i,
954 LLVMValueRef j)
955 {
956 LLVMValueRef args[6];
957 LLVMValueRef p1;
958
959 args[0] = i;
960 args[1] = llvm_chan;
961 args[2] = attr_number;
962 args[3] = ctx->i1false;
963 args[4] = params;
964
965 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1.f16",
966 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
967
968 args[0] = p1;
969 args[1] = j;
970 args[2] = llvm_chan;
971 args[3] = attr_number;
972 args[4] = ctx->i1false;
973 args[5] = params;
974
975 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2.f16",
976 ctx->f16, args, 6, AC_FUNC_ATTR_READNONE);
977 }
978
979 LLVMValueRef
980 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
981 LLVMValueRef parameter,
982 LLVMValueRef llvm_chan,
983 LLVMValueRef attr_number,
984 LLVMValueRef params)
985 {
986 LLVMValueRef args[4];
987
988 args[0] = parameter;
989 args[1] = llvm_chan;
990 args[2] = attr_number;
991 args[3] = params;
992
993 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.mov",
994 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
995 }
996
997 LLVMValueRef
998 ac_build_gep_ptr(struct ac_llvm_context *ctx,
999 LLVMValueRef base_ptr,
1000 LLVMValueRef index)
1001 {
1002 return LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
1003 }
1004
1005 LLVMValueRef
1006 ac_build_gep0(struct ac_llvm_context *ctx,
1007 LLVMValueRef base_ptr,
1008 LLVMValueRef index)
1009 {
1010 LLVMValueRef indices[2] = {
1011 ctx->i32_0,
1012 index,
1013 };
1014 return LLVMBuildGEP(ctx->builder, base_ptr, indices, 2, "");
1015 }
1016
1017 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
1018 LLVMValueRef index)
1019 {
1020 return LLVMBuildPointerCast(ctx->builder,
1021 LLVMBuildGEP(ctx->builder, ptr, &index, 1, ""),
1022 LLVMTypeOf(ptr), "");
1023 }
1024
1025 void
1026 ac_build_indexed_store(struct ac_llvm_context *ctx,
1027 LLVMValueRef base_ptr, LLVMValueRef index,
1028 LLVMValueRef value)
1029 {
1030 LLVMBuildStore(ctx->builder, value,
1031 ac_build_gep0(ctx, base_ptr, index));
1032 }
1033
1034 /**
1035 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
1036 * It's equivalent to doing a load from &base_ptr[index].
1037 *
1038 * \param base_ptr Where the array starts.
1039 * \param index The element index into the array.
1040 * \param uniform Whether the base_ptr and index can be assumed to be
1041 * dynamically uniform (i.e. load to an SGPR)
1042 * \param invariant Whether the load is invariant (no other opcodes affect it)
1043 * \param no_unsigned_wraparound
1044 * For all possible re-associations and re-distributions of an expression
1045 * "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
1046 * without inbounds in base_ptr), this parameter is true if "addr + offset"
1047 * does not result in an unsigned integer wraparound. This is used for
1048 * optimal code generation of 32-bit pointer arithmetic.
1049 *
1050 * For example, a 32-bit immediate offset that causes a 32-bit unsigned
1051 * integer wraparound can't be an imm offset in s_load_dword, because
1052 * the instruction performs "addr + offset" in 64 bits.
1053 *
1054 * Expected usage for bindless textures by chaining GEPs:
1055 * // possible unsigned wraparound, don't use InBounds:
1056 * ptr1 = LLVMBuildGEP(base_ptr, index);
1057 * image = load(ptr1); // becomes "s_load ptr1, 0"
1058 *
1059 * ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1060 * sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1061 */
1062 static LLVMValueRef
1063 ac_build_load_custom(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1064 LLVMValueRef index, bool uniform, bool invariant,
1065 bool no_unsigned_wraparound)
1066 {
1067 LLVMValueRef pointer, result;
1068
1069 if (no_unsigned_wraparound &&
1070 LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr)) == AC_ADDR_SPACE_CONST_32BIT)
1071 pointer = LLVMBuildInBoundsGEP(ctx->builder, base_ptr, &index, 1, "");
1072 else
1073 pointer = LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
1074
1075 if (uniform)
1076 LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
1077 result = LLVMBuildLoad(ctx->builder, pointer, "");
1078 if (invariant)
1079 LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
1080 return result;
1081 }
1082
1083 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1084 LLVMValueRef index)
1085 {
1086 return ac_build_load_custom(ctx, base_ptr, index, false, false, false);
1087 }
1088
1089 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
1090 LLVMValueRef base_ptr, LLVMValueRef index)
1091 {
1092 return ac_build_load_custom(ctx, base_ptr, index, false, true, false);
1093 }
1094
1095 /* This assumes that there is no unsigned integer wraparound during the address
1096 * computation, excluding all GEPs within base_ptr. */
1097 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
1098 LLVMValueRef base_ptr, LLVMValueRef index)
1099 {
1100 return ac_build_load_custom(ctx, base_ptr, index, true, true, true);
1101 }
1102
1103 /* See ac_build_load_custom() documentation. */
1104 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
1105 LLVMValueRef base_ptr, LLVMValueRef index)
1106 {
1107 return ac_build_load_custom(ctx, base_ptr, index, true, true, false);
1108 }
1109
1110 static void
1111 ac_build_llvm7_buffer_store_common(struct ac_llvm_context *ctx,
1112 LLVMValueRef rsrc,
1113 LLVMValueRef data,
1114 LLVMValueRef vindex,
1115 LLVMValueRef voffset,
1116 unsigned num_channels,
1117 bool glc,
1118 bool slc,
1119 bool use_format)
1120 {
1121 LLVMValueRef args[] = {
1122 data,
1123 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
1124 vindex ? vindex : ctx->i32_0,
1125 voffset,
1126 LLVMConstInt(ctx->i1, glc, 0),
1127 LLVMConstInt(ctx->i1, slc, 0)
1128 };
1129 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1130
1131 const char *type_names[] = {"f32", "v2f32", "v4f32"};
1132 char name[256];
1133
1134 if (use_format) {
1135 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.store.format.%s",
1136 type_names[func]);
1137 } else {
1138 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.store.%s",
1139 type_names[func]);
1140 }
1141
1142 ac_build_intrinsic(ctx, name, ctx->voidt, args, ARRAY_SIZE(args),
1143 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
1144 }
1145
1146 static void
1147 ac_build_llvm8_buffer_store_common(struct ac_llvm_context *ctx,
1148 LLVMValueRef rsrc,
1149 LLVMValueRef data,
1150 LLVMValueRef vindex,
1151 LLVMValueRef voffset,
1152 LLVMValueRef soffset,
1153 unsigned num_channels,
1154 LLVMTypeRef return_channel_type,
1155 bool glc,
1156 bool slc,
1157 bool use_format,
1158 bool structurized)
1159 {
1160 LLVMValueRef args[6];
1161 int idx = 0;
1162 args[idx++] = data;
1163 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1164 if (structurized)
1165 args[idx++] = vindex ? vindex : ctx->i32_0;
1166 args[idx++] = voffset ? voffset : ctx->i32_0;
1167 args[idx++] = soffset ? soffset : ctx->i32_0;
1168 args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0);
1169 unsigned func = !ac_has_vec3_support(ctx->chip_class, use_format) && num_channels == 3 ? 4 : num_channels;
1170 const char *indexing_kind = structurized ? "struct" : "raw";
1171 char name[256], type_name[8];
1172
1173 LLVMTypeRef type = func > 1 ? LLVMVectorType(return_channel_type, func) : return_channel_type;
1174 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1175
1176 if (use_format) {
1177 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.format.%s",
1178 indexing_kind, type_name);
1179 } else {
1180 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.%s",
1181 indexing_kind, type_name);
1182 }
1183
1184 ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
1185 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
1186 }
1187
1188 void
1189 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
1190 LLVMValueRef rsrc,
1191 LLVMValueRef data,
1192 LLVMValueRef vindex,
1193 LLVMValueRef voffset,
1194 unsigned num_channels,
1195 bool glc,
1196 bool slc)
1197 {
1198 if (HAVE_LLVM >= 0x800) {
1199 ac_build_llvm8_buffer_store_common(ctx, rsrc, data, vindex,
1200 voffset, NULL, num_channels,
1201 ctx->f32, glc, slc,
1202 true, true);
1203 } else {
1204 ac_build_llvm7_buffer_store_common(ctx, rsrc, data, vindex, voffset,
1205 num_channels, glc, slc,
1206 true);
1207 }
1208 }
1209
1210 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1211 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1212 * or v4i32 (num_channels=3,4).
1213 */
1214 void
1215 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
1216 LLVMValueRef rsrc,
1217 LLVMValueRef vdata,
1218 unsigned num_channels,
1219 LLVMValueRef voffset,
1220 LLVMValueRef soffset,
1221 unsigned inst_offset,
1222 bool glc,
1223 bool slc,
1224 bool swizzle_enable_hint)
1225 {
1226 /* Split 3 channel stores, because only LLVM 9+ support 3-channel
1227 * intrinsics. */
1228 if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false)) {
1229 LLVMValueRef v[3], v01;
1230
1231 for (int i = 0; i < 3; i++) {
1232 v[i] = LLVMBuildExtractElement(ctx->builder, vdata,
1233 LLVMConstInt(ctx->i32, i, 0), "");
1234 }
1235 v01 = ac_build_gather_values(ctx, v, 2);
1236
1237 ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset,
1238 soffset, inst_offset, glc, slc,
1239 swizzle_enable_hint);
1240 ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset,
1241 soffset, inst_offset + 8,
1242 glc, slc,
1243 swizzle_enable_hint);
1244 return;
1245 }
1246
1247 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1248 * (voffset is swizzled, but soffset isn't swizzled).
1249 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1250 */
1251 if (!swizzle_enable_hint) {
1252 LLVMValueRef offset = soffset;
1253
1254 if (inst_offset)
1255 offset = LLVMBuildAdd(ctx->builder, offset,
1256 LLVMConstInt(ctx->i32, inst_offset, 0), "");
1257
1258 if (HAVE_LLVM >= 0x800) {
1259 ac_build_llvm8_buffer_store_common(ctx, rsrc,
1260 ac_to_float(ctx, vdata),
1261 ctx->i32_0,
1262 voffset, offset,
1263 num_channels,
1264 ctx->f32,
1265 glc, slc,
1266 false, false);
1267 } else {
1268 if (voffset)
1269 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1270
1271 ac_build_llvm7_buffer_store_common(ctx, rsrc,
1272 ac_to_float(ctx, vdata),
1273 ctx->i32_0, offset,
1274 num_channels, glc, slc,
1275 false);
1276 }
1277 return;
1278 }
1279
1280 static const unsigned dfmts[] = {
1281 V_008F0C_BUF_DATA_FORMAT_32,
1282 V_008F0C_BUF_DATA_FORMAT_32_32,
1283 V_008F0C_BUF_DATA_FORMAT_32_32_32,
1284 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1285 };
1286 unsigned dfmt = dfmts[num_channels - 1];
1287 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1288 LLVMValueRef immoffset = LLVMConstInt(ctx->i32, inst_offset, 0);
1289
1290 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1291 immoffset, num_channels, dfmt, nfmt, glc,
1292 slc);
1293 }
1294
1295 static LLVMValueRef
1296 ac_build_llvm7_buffer_load_common(struct ac_llvm_context *ctx,
1297 LLVMValueRef rsrc,
1298 LLVMValueRef vindex,
1299 LLVMValueRef voffset,
1300 unsigned num_channels,
1301 bool glc,
1302 bool slc,
1303 bool can_speculate,
1304 bool use_format)
1305 {
1306 LLVMValueRef args[] = {
1307 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
1308 vindex ? vindex : ctx->i32_0,
1309 voffset,
1310 LLVMConstInt(ctx->i1, glc, 0),
1311 LLVMConstInt(ctx->i1, slc, 0)
1312 };
1313 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1314
1315 LLVMTypeRef types[] = {ctx->f32, ctx->v2f32, ctx->v4f32};
1316 const char *type_names[] = {"f32", "v2f32", "v4f32"};
1317 char name[256];
1318
1319 if (use_format) {
1320 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.format.%s",
1321 type_names[func]);
1322 } else {
1323 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.%s",
1324 type_names[func]);
1325 }
1326
1327 return ac_build_intrinsic(ctx, name, types[func], args,
1328 ARRAY_SIZE(args),
1329 ac_get_load_intr_attribs(can_speculate));
1330 }
1331
1332 static LLVMValueRef
1333 ac_build_llvm8_buffer_load_common(struct ac_llvm_context *ctx,
1334 LLVMValueRef rsrc,
1335 LLVMValueRef vindex,
1336 LLVMValueRef voffset,
1337 LLVMValueRef soffset,
1338 unsigned num_channels,
1339 LLVMTypeRef channel_type,
1340 bool glc,
1341 bool slc,
1342 bool can_speculate,
1343 bool use_format,
1344 bool structurized)
1345 {
1346 LLVMValueRef args[5];
1347 int idx = 0;
1348 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1349 if (structurized)
1350 args[idx++] = vindex ? vindex : ctx->i32_0;
1351 args[idx++] = voffset ? voffset : ctx->i32_0;
1352 args[idx++] = soffset ? soffset : ctx->i32_0;
1353 args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0);
1354 unsigned func = !ac_has_vec3_support(ctx->chip_class, use_format) && num_channels == 3 ? 4 : num_channels;
1355 const char *indexing_kind = structurized ? "struct" : "raw";
1356 char name[256], type_name[8];
1357
1358 LLVMTypeRef type = func > 1 ? LLVMVectorType(channel_type, func) : channel_type;
1359 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1360
1361 if (use_format) {
1362 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.format.%s",
1363 indexing_kind, type_name);
1364 } else {
1365 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.%s",
1366 indexing_kind, type_name);
1367 }
1368
1369 return ac_build_intrinsic(ctx, name, type, args, idx,
1370 ac_get_load_intr_attribs(can_speculate));
1371 }
1372
1373 LLVMValueRef
1374 ac_build_buffer_load(struct ac_llvm_context *ctx,
1375 LLVMValueRef rsrc,
1376 int num_channels,
1377 LLVMValueRef vindex,
1378 LLVMValueRef voffset,
1379 LLVMValueRef soffset,
1380 unsigned inst_offset,
1381 unsigned glc,
1382 unsigned slc,
1383 bool can_speculate,
1384 bool allow_smem)
1385 {
1386 LLVMValueRef offset = LLVMConstInt(ctx->i32, inst_offset, 0);
1387 if (voffset)
1388 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1389 if (soffset)
1390 offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
1391
1392 if (allow_smem && !slc &&
1393 (!glc || (HAVE_LLVM >= 0x0800 && ctx->chip_class >= GFX8))) {
1394 assert(vindex == NULL);
1395
1396 LLVMValueRef result[8];
1397
1398 for (int i = 0; i < num_channels; i++) {
1399 if (i) {
1400 offset = LLVMBuildAdd(ctx->builder, offset,
1401 LLVMConstInt(ctx->i32, 4, 0), "");
1402 }
1403 const char *intrname =
1404 HAVE_LLVM >= 0x0800 ? "llvm.amdgcn.s.buffer.load.f32"
1405 : "llvm.SI.load.const.v4i32";
1406 unsigned num_args = HAVE_LLVM >= 0x0800 ? 3 : 2;
1407 LLVMValueRef args[3] = {
1408 rsrc,
1409 offset,
1410 glc ? ctx->i32_1 : ctx->i32_0,
1411 };
1412 result[i] = ac_build_intrinsic(ctx, intrname,
1413 ctx->f32, args, num_args,
1414 AC_FUNC_ATTR_READNONE |
1415 (HAVE_LLVM < 0x0800 ? AC_FUNC_ATTR_LEGACY : 0));
1416 }
1417 if (num_channels == 1)
1418 return result[0];
1419
1420 if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false))
1421 result[num_channels++] = LLVMGetUndef(ctx->f32);
1422 return ac_build_gather_values(ctx, result, num_channels);
1423 }
1424
1425 if (HAVE_LLVM >= 0x0800) {
1426 return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex,
1427 offset, ctx->i32_0,
1428 num_channels, ctx->f32,
1429 glc, slc,
1430 can_speculate, false,
1431 false);
1432 }
1433
1434 return ac_build_llvm7_buffer_load_common(ctx, rsrc, vindex, offset,
1435 num_channels, glc, slc,
1436 can_speculate, false);
1437 }
1438
1439 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
1440 LLVMValueRef rsrc,
1441 LLVMValueRef vindex,
1442 LLVMValueRef voffset,
1443 unsigned num_channels,
1444 bool glc,
1445 bool can_speculate)
1446 {
1447 if (HAVE_LLVM >= 0x800) {
1448 return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
1449 num_channels, ctx->f32,
1450 glc, false,
1451 can_speculate, true, true);
1452 }
1453 return ac_build_llvm7_buffer_load_common(ctx, rsrc, vindex, voffset,
1454 num_channels, glc, false,
1455 can_speculate, true);
1456 }
1457
1458 LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
1459 LLVMValueRef rsrc,
1460 LLVMValueRef vindex,
1461 LLVMValueRef voffset,
1462 unsigned num_channels,
1463 bool glc,
1464 bool can_speculate)
1465 {
1466 if (HAVE_LLVM >= 0x800) {
1467 return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
1468 num_channels, ctx->f32,
1469 glc, false,
1470 can_speculate, true, true);
1471 }
1472
1473 LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->builder, rsrc, LLVMConstInt(ctx->i32, 2, 0), "");
1474 LLVMValueRef stride = LLVMBuildExtractElement(ctx->builder, rsrc, ctx->i32_1, "");
1475 stride = LLVMBuildLShr(ctx->builder, stride, LLVMConstInt(ctx->i32, 16, 0), "");
1476
1477 LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->builder,
1478 LLVMBuildICmp(ctx->builder, LLVMIntUGT, elem_count, stride, ""),
1479 elem_count, stride, "");
1480
1481 LLVMValueRef new_rsrc = LLVMBuildInsertElement(ctx->builder, rsrc, new_elem_count,
1482 LLVMConstInt(ctx->i32, 2, 0), "");
1483
1484 return ac_build_llvm7_buffer_load_common(ctx, new_rsrc, vindex, voffset,
1485 num_channels, glc, false,
1486 can_speculate, true);
1487 }
1488
1489 /// Translate a (dfmt, nfmt) pair into a chip-appropriate combined format
1490 /// value for LLVM8+ tbuffer intrinsics.
1491 static unsigned
1492 ac_get_tbuffer_format(struct ac_llvm_context *ctx,
1493 unsigned dfmt, unsigned nfmt)
1494 {
1495 if (ctx->chip_class >= GFX10) {
1496 unsigned format;
1497 switch (dfmt) {
1498 default: unreachable("bad dfmt");
1499 case V_008F0C_BUF_DATA_FORMAT_8: format = V_008F0C_IMG_FORMAT_8_UINT; break;
1500 case V_008F0C_BUF_DATA_FORMAT_8_8: format = V_008F0C_IMG_FORMAT_8_8_UINT; break;
1501 case V_008F0C_BUF_DATA_FORMAT_8_8_8_8: format = V_008F0C_IMG_FORMAT_8_8_8_8_UINT; break;
1502 case V_008F0C_BUF_DATA_FORMAT_16: format = V_008F0C_IMG_FORMAT_16_UINT; break;
1503 case V_008F0C_BUF_DATA_FORMAT_16_16: format = V_008F0C_IMG_FORMAT_16_16_UINT; break;
1504 case V_008F0C_BUF_DATA_FORMAT_16_16_16_16: format = V_008F0C_IMG_FORMAT_16_16_16_16_UINT; break;
1505 case V_008F0C_BUF_DATA_FORMAT_32: format = V_008F0C_IMG_FORMAT_32_UINT; break;
1506 case V_008F0C_BUF_DATA_FORMAT_32_32: format = V_008F0C_IMG_FORMAT_32_32_UINT; break;
1507 case V_008F0C_BUF_DATA_FORMAT_32_32_32_32: format = V_008F0C_IMG_FORMAT_32_32_32_32_UINT; break;
1508 }
1509
1510 // Use the regularity properties of the combined format enum.
1511 //
1512 // Note: float is incompatible with 8-bit data formats,
1513 // [us]{norm,scaled} are incomparible with 32-bit data formats.
1514 // [us]scaled are not writable.
1515 switch (nfmt) {
1516 case V_008F0C_BUF_NUM_FORMAT_UNORM: format -= 4; break;
1517 case V_008F0C_BUF_NUM_FORMAT_SNORM: format -= 3; break;
1518 case V_008F0C_BUF_NUM_FORMAT_USCALED: format -= 2; break;
1519 case V_008F0C_BUF_NUM_FORMAT_SSCALED: format -= 1; break;
1520 default: unreachable("bad nfmt");
1521 case V_008F0C_BUF_NUM_FORMAT_UINT: break;
1522 case V_008F0C_BUF_NUM_FORMAT_SINT: format += 1; break;
1523 case V_008F0C_BUF_NUM_FORMAT_FLOAT: format += 2; break;
1524 }
1525
1526 return format;
1527 } else {
1528 return dfmt | (nfmt << 4);
1529 }
1530 }
1531
1532 static LLVMValueRef
1533 ac_build_llvm8_tbuffer_load(struct ac_llvm_context *ctx,
1534 LLVMValueRef rsrc,
1535 LLVMValueRef vindex,
1536 LLVMValueRef voffset,
1537 LLVMValueRef soffset,
1538 unsigned num_channels,
1539 unsigned dfmt,
1540 unsigned nfmt,
1541 bool glc,
1542 bool slc,
1543 bool can_speculate,
1544 bool structurized)
1545 {
1546 LLVMValueRef args[6];
1547 int idx = 0;
1548 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1549 if (structurized)
1550 args[idx++] = vindex ? vindex : ctx->i32_0;
1551 args[idx++] = voffset ? voffset : ctx->i32_0;
1552 args[idx++] = soffset ? soffset : ctx->i32_0;
1553 args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx, dfmt, nfmt), 0);
1554 args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0);
1555 unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
1556 const char *indexing_kind = structurized ? "struct" : "raw";
1557 char name[256], type_name[8];
1558
1559 LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
1560 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1561
1562 snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.load.%s",
1563 indexing_kind, type_name);
1564
1565 return ac_build_intrinsic(ctx, name, type, args, idx,
1566 ac_get_load_intr_attribs(can_speculate));
1567 }
1568
1569 static LLVMValueRef
1570 ac_build_tbuffer_load(struct ac_llvm_context *ctx,
1571 LLVMValueRef rsrc,
1572 LLVMValueRef vindex,
1573 LLVMValueRef voffset,
1574 LLVMValueRef soffset,
1575 LLVMValueRef immoffset,
1576 unsigned num_channels,
1577 unsigned dfmt,
1578 unsigned nfmt,
1579 bool glc,
1580 bool slc,
1581 bool can_speculate,
1582 bool structurized) /* only matters for LLVM 8+ */
1583 {
1584 if (HAVE_LLVM >= 0x800) {
1585 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1586
1587 return ac_build_llvm8_tbuffer_load(ctx, rsrc, vindex, voffset,
1588 soffset, num_channels,
1589 dfmt, nfmt, glc, slc,
1590 can_speculate, structurized);
1591 }
1592
1593 LLVMValueRef args[] = {
1594 rsrc,
1595 vindex ? vindex : ctx->i32_0,
1596 voffset,
1597 soffset,
1598 immoffset,
1599 LLVMConstInt(ctx->i32, dfmt, false),
1600 LLVMConstInt(ctx->i32, nfmt, false),
1601 LLVMConstInt(ctx->i1, glc, false),
1602 LLVMConstInt(ctx->i1, slc, false),
1603 };
1604 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1605 LLVMTypeRef types[] = {ctx->i32, ctx->v2i32, ctx->v4i32};
1606 const char *type_names[] = {"i32", "v2i32", "v4i32"};
1607 char name[256];
1608
1609 snprintf(name, sizeof(name), "llvm.amdgcn.tbuffer.load.%s",
1610 type_names[func]);
1611
1612 return ac_build_intrinsic(ctx, name, types[func], args, 9,
1613 ac_get_load_intr_attribs(can_speculate));
1614 }
1615
1616 LLVMValueRef
1617 ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
1618 LLVMValueRef rsrc,
1619 LLVMValueRef vindex,
1620 LLVMValueRef voffset,
1621 LLVMValueRef soffset,
1622 LLVMValueRef immoffset,
1623 unsigned num_channels,
1624 unsigned dfmt,
1625 unsigned nfmt,
1626 bool glc,
1627 bool slc,
1628 bool can_speculate)
1629 {
1630 return ac_build_tbuffer_load(ctx, rsrc, vindex, voffset, soffset,
1631 immoffset, num_channels, dfmt, nfmt, glc,
1632 slc, can_speculate, true);
1633 }
1634
1635 LLVMValueRef
1636 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
1637 LLVMValueRef rsrc,
1638 LLVMValueRef voffset,
1639 LLVMValueRef soffset,
1640 LLVMValueRef immoffset,
1641 unsigned num_channels,
1642 unsigned dfmt,
1643 unsigned nfmt,
1644 bool glc,
1645 bool slc,
1646 bool can_speculate)
1647 {
1648 return ac_build_tbuffer_load(ctx, rsrc, NULL, voffset, soffset,
1649 immoffset, num_channels, dfmt, nfmt, glc,
1650 slc, can_speculate, false);
1651 }
1652
1653 LLVMValueRef
1654 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
1655 LLVMValueRef rsrc,
1656 LLVMValueRef voffset,
1657 LLVMValueRef soffset,
1658 LLVMValueRef immoffset,
1659 bool glc)
1660 {
1661 LLVMValueRef res;
1662
1663 if (HAVE_LLVM >= 0x900) {
1664 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1665
1666 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1667 res = ac_build_llvm8_buffer_load_common(ctx, rsrc, NULL,
1668 voffset, soffset,
1669 1, ctx->i16, glc, false,
1670 false, false, false);
1671 } else {
1672 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
1673 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1674
1675 res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
1676 immoffset, 1, dfmt, nfmt, glc, false,
1677 false);
1678
1679 res = LLVMBuildTrunc(ctx->builder, res, ctx->i16, "");
1680 }
1681
1682 return res;
1683 }
1684
1685 LLVMValueRef
1686 ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
1687 LLVMValueRef rsrc,
1688 LLVMValueRef voffset,
1689 LLVMValueRef soffset,
1690 LLVMValueRef immoffset,
1691 bool glc)
1692 {
1693 LLVMValueRef res;
1694
1695 if (HAVE_LLVM >= 0x900) {
1696 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1697
1698 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1699 res = ac_build_llvm8_buffer_load_common(ctx, rsrc, NULL,
1700 voffset, soffset,
1701 1, ctx->i8, glc, false,
1702 false, false, false);
1703 } else {
1704 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
1705 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1706
1707 res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
1708 immoffset, 1, dfmt, nfmt, glc, false,
1709 false);
1710
1711 res = LLVMBuildTrunc(ctx->builder, res, ctx->i8, "");
1712 }
1713
1714 return res;
1715 }
1716
1717 /**
1718 * Convert an 11- or 10-bit unsigned floating point number to an f32.
1719 *
1720 * The input exponent is expected to be biased analogous to IEEE-754, i.e. by
1721 * 2^(exp_bits-1) - 1 (as defined in OpenGL and other graphics APIs).
1722 */
1723 static LLVMValueRef
1724 ac_ufN_to_float(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned exp_bits, unsigned mant_bits)
1725 {
1726 assert(LLVMTypeOf(src) == ctx->i32);
1727
1728 LLVMValueRef tmp;
1729 LLVMValueRef mantissa;
1730 mantissa = LLVMBuildAnd(ctx->builder, src, LLVMConstInt(ctx->i32, (1 << mant_bits) - 1, false), "");
1731
1732 /* Converting normal numbers is just a shift + correcting the exponent bias */
1733 unsigned normal_shift = 23 - mant_bits;
1734 unsigned bias_shift = 127 - ((1 << (exp_bits - 1)) - 1);
1735 LLVMValueRef shifted, normal;
1736
1737 shifted = LLVMBuildShl(ctx->builder, src, LLVMConstInt(ctx->i32, normal_shift, false), "");
1738 normal = LLVMBuildAdd(ctx->builder, shifted, LLVMConstInt(ctx->i32, bias_shift << 23, false), "");
1739
1740 /* Converting nan/inf numbers is the same, but with a different exponent update */
1741 LLVMValueRef naninf;
1742 naninf = LLVMBuildOr(ctx->builder, normal, LLVMConstInt(ctx->i32, 0xff << 23, false), "");
1743
1744 /* Converting denormals is the complex case: determine the leading zeros of the
1745 * mantissa to obtain the correct shift for the mantissa and exponent correction.
1746 */
1747 LLVMValueRef denormal;
1748 LLVMValueRef params[2] = {
1749 mantissa,
1750 ctx->i1true, /* result can be undef when arg is 0 */
1751 };
1752 LLVMValueRef ctlz = ac_build_intrinsic(ctx, "llvm.ctlz.i32", ctx->i32,
1753 params, 2, AC_FUNC_ATTR_READNONE);
1754
1755 /* Shift such that the leading 1 ends up as the LSB of the exponent field. */
1756 tmp = LLVMBuildSub(ctx->builder, ctlz, LLVMConstInt(ctx->i32, 8, false), "");
1757 denormal = LLVMBuildShl(ctx->builder, mantissa, tmp, "");
1758
1759 unsigned denormal_exp = bias_shift + (32 - mant_bits) - 1;
1760 tmp = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, denormal_exp, false), ctlz, "");
1761 tmp = LLVMBuildShl(ctx->builder, tmp, LLVMConstInt(ctx->i32, 23, false), "");
1762 denormal = LLVMBuildAdd(ctx->builder, denormal, tmp, "");
1763
1764 /* Select the final result. */
1765 LLVMValueRef result;
1766
1767 tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src,
1768 LLVMConstInt(ctx->i32, ((1 << exp_bits) - 1) << mant_bits, false), "");
1769 result = LLVMBuildSelect(ctx->builder, tmp, naninf, normal, "");
1770
1771 tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src,
1772 LLVMConstInt(ctx->i32, 1 << mant_bits, false), "");
1773 result = LLVMBuildSelect(ctx->builder, tmp, result, denormal, "");
1774
1775 tmp = LLVMBuildICmp(ctx->builder, LLVMIntNE, src, ctx->i32_0, "");
1776 result = LLVMBuildSelect(ctx->builder, tmp, result, ctx->i32_0, "");
1777
1778 return ac_to_float(ctx, result);
1779 }
1780
1781 /**
1782 * Generate a fully general open coded buffer format fetch with all required
1783 * fixups suitable for vertex fetch, using non-format buffer loads.
1784 *
1785 * Some combinations of argument values have special interpretations:
1786 * - size = 8 bytes, format = fixed indicates PIPE_FORMAT_R11G11B10_FLOAT
1787 * - size = 8 bytes, format != {float,fixed} indicates a 2_10_10_10 data format
1788 *
1789 * \param log_size log(size of channel in bytes)
1790 * \param num_channels number of channels (1 to 4)
1791 * \param format AC_FETCH_FORMAT_xxx value
1792 * \param reverse whether XYZ channels are reversed
1793 * \param known_aligned whether the source is known to be aligned to hardware's
1794 * effective element size for loading the given format
1795 * (note: this means dword alignment for 8_8_8_8, 16_16, etc.)
1796 * \param rsrc buffer resource descriptor
1797 * \return the resulting vector of floats or integers bitcast to <4 x i32>
1798 */
1799 LLVMValueRef
1800 ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
1801 unsigned log_size,
1802 unsigned num_channels,
1803 unsigned format,
1804 bool reverse,
1805 bool known_aligned,
1806 LLVMValueRef rsrc,
1807 LLVMValueRef vindex,
1808 LLVMValueRef voffset,
1809 LLVMValueRef soffset,
1810 bool glc,
1811 bool slc,
1812 bool can_speculate)
1813 {
1814 LLVMValueRef tmp;
1815 unsigned load_log_size = log_size;
1816 unsigned load_num_channels = num_channels;
1817 if (log_size == 3) {
1818 load_log_size = 2;
1819 if (format == AC_FETCH_FORMAT_FLOAT) {
1820 load_num_channels = 2 * num_channels;
1821 } else {
1822 load_num_channels = 1; /* 10_11_11 or 2_10_10_10 */
1823 }
1824 }
1825
1826 int log_recombine = 0;
1827 if (ctx->chip_class == GFX6 && !known_aligned) {
1828 /* Avoid alignment restrictions by loading one byte at a time. */
1829 load_num_channels <<= load_log_size;
1830 log_recombine = load_log_size;
1831 load_log_size = 0;
1832 } else if (load_num_channels == 2 || load_num_channels == 4) {
1833 log_recombine = -util_logbase2(load_num_channels);
1834 load_num_channels = 1;
1835 load_log_size += -log_recombine;
1836 }
1837
1838 assert(load_log_size >= 2 || HAVE_LLVM >= 0x0900);
1839
1840 LLVMValueRef loads[32]; /* up to 32 bytes */
1841 for (unsigned i = 0; i < load_num_channels; ++i) {
1842 tmp = LLVMBuildAdd(ctx->builder, soffset,
1843 LLVMConstInt(ctx->i32, i << load_log_size, false), "");
1844 if (HAVE_LLVM >= 0x0800) {
1845 LLVMTypeRef channel_type = load_log_size == 0 ? ctx->i8 :
1846 load_log_size == 1 ? ctx->i16 : ctx->i32;
1847 unsigned num_channels = 1 << (MAX2(load_log_size, 2) - 2);
1848 loads[i] = ac_build_llvm8_buffer_load_common(
1849 ctx, rsrc, vindex, voffset, tmp,
1850 num_channels, channel_type, glc, slc,
1851 can_speculate, false, true);
1852 } else {
1853 tmp = LLVMBuildAdd(ctx->builder, voffset, tmp, "");
1854 loads[i] = ac_build_llvm7_buffer_load_common(
1855 ctx, rsrc, vindex, tmp,
1856 1 << (load_log_size - 2), glc, slc, can_speculate, false);
1857 }
1858 if (load_log_size >= 2)
1859 loads[i] = ac_to_integer(ctx, loads[i]);
1860 }
1861
1862 if (log_recombine > 0) {
1863 /* Recombine bytes if necessary (GFX6 only) */
1864 LLVMTypeRef dst_type = log_recombine == 2 ? ctx->i32 : ctx->i16;
1865
1866 for (unsigned src = 0, dst = 0; src < load_num_channels; ++dst) {
1867 LLVMValueRef accum = NULL;
1868 for (unsigned i = 0; i < (1 << log_recombine); ++i, ++src) {
1869 tmp = LLVMBuildZExt(ctx->builder, loads[src], dst_type, "");
1870 if (i == 0) {
1871 accum = tmp;
1872 } else {
1873 tmp = LLVMBuildShl(ctx->builder, tmp,
1874 LLVMConstInt(dst_type, 8 * i, false), "");
1875 accum = LLVMBuildOr(ctx->builder, accum, tmp, "");
1876 }
1877 }
1878 loads[dst] = accum;
1879 }
1880 } else if (log_recombine < 0) {
1881 /* Split vectors of dwords */
1882 if (load_log_size > 2) {
1883 assert(load_num_channels == 1);
1884 LLVMValueRef loaded = loads[0];
1885 unsigned log_split = load_log_size - 2;
1886 log_recombine += log_split;
1887 load_num_channels = 1 << log_split;
1888 load_log_size = 2;
1889 for (unsigned i = 0; i < load_num_channels; ++i) {
1890 tmp = LLVMConstInt(ctx->i32, i, false);
1891 loads[i] = LLVMBuildExtractElement(ctx->builder, loaded, tmp, "");
1892 }
1893 }
1894
1895 /* Further split dwords and shorts if required */
1896 if (log_recombine < 0) {
1897 for (unsigned src = load_num_channels,
1898 dst = load_num_channels << -log_recombine;
1899 src > 0; --src) {
1900 unsigned dst_bits = 1 << (3 + load_log_size + log_recombine);
1901 LLVMTypeRef dst_type = LLVMIntTypeInContext(ctx->context, dst_bits);
1902 LLVMValueRef loaded = loads[src - 1];
1903 LLVMTypeRef loaded_type = LLVMTypeOf(loaded);
1904 for (unsigned i = 1 << -log_recombine; i > 0; --i, --dst) {
1905 tmp = LLVMConstInt(loaded_type, dst_bits * (i - 1), false);
1906 tmp = LLVMBuildLShr(ctx->builder, loaded, tmp, "");
1907 loads[dst - 1] = LLVMBuildTrunc(ctx->builder, tmp, dst_type, "");
1908 }
1909 }
1910 }
1911 }
1912
1913 if (log_size == 3) {
1914 if (format == AC_FETCH_FORMAT_FLOAT) {
1915 for (unsigned i = 0; i < num_channels; ++i) {
1916 tmp = ac_build_gather_values(ctx, &loads[2 * i], 2);
1917 loads[i] = LLVMBuildBitCast(ctx->builder, tmp, ctx->f64, "");
1918 }
1919 } else if (format == AC_FETCH_FORMAT_FIXED) {
1920 /* 10_11_11_FLOAT */
1921 LLVMValueRef data = loads[0];
1922 LLVMValueRef i32_2047 = LLVMConstInt(ctx->i32, 2047, false);
1923 LLVMValueRef r = LLVMBuildAnd(ctx->builder, data, i32_2047, "");
1924 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 11, false), "");
1925 LLVMValueRef g = LLVMBuildAnd(ctx->builder, tmp, i32_2047, "");
1926 LLVMValueRef b = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 22, false), "");
1927
1928 loads[0] = ac_to_integer(ctx, ac_ufN_to_float(ctx, r, 5, 6));
1929 loads[1] = ac_to_integer(ctx, ac_ufN_to_float(ctx, g, 5, 6));
1930 loads[2] = ac_to_integer(ctx, ac_ufN_to_float(ctx, b, 5, 5));
1931
1932 num_channels = 3;
1933 log_size = 2;
1934 format = AC_FETCH_FORMAT_FLOAT;
1935 } else {
1936 /* 2_10_10_10 data formats */
1937 LLVMValueRef data = loads[0];
1938 LLVMTypeRef i10 = LLVMIntTypeInContext(ctx->context, 10);
1939 LLVMTypeRef i2 = LLVMIntTypeInContext(ctx->context, 2);
1940 loads[0] = LLVMBuildTrunc(ctx->builder, data, i10, "");
1941 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 10, false), "");
1942 loads[1] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
1943 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 20, false), "");
1944 loads[2] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
1945 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 30, false), "");
1946 loads[3] = LLVMBuildTrunc(ctx->builder, tmp, i2, "");
1947
1948 num_channels = 4;
1949 }
1950 }
1951
1952 if (format == AC_FETCH_FORMAT_FLOAT) {
1953 if (log_size != 2) {
1954 for (unsigned chan = 0; chan < num_channels; ++chan) {
1955 tmp = ac_to_float(ctx, loads[chan]);
1956 if (log_size == 3)
1957 tmp = LLVMBuildFPTrunc(ctx->builder, tmp, ctx->f32, "");
1958 else if (log_size == 1)
1959 tmp = LLVMBuildFPExt(ctx->builder, tmp, ctx->f32, "");
1960 loads[chan] = ac_to_integer(ctx, tmp);
1961 }
1962 }
1963 } else if (format == AC_FETCH_FORMAT_UINT) {
1964 if (log_size != 2) {
1965 for (unsigned chan = 0; chan < num_channels; ++chan)
1966 loads[chan] = LLVMBuildZExt(ctx->builder, loads[chan], ctx->i32, "");
1967 }
1968 } else if (format == AC_FETCH_FORMAT_SINT) {
1969 if (log_size != 2) {
1970 for (unsigned chan = 0; chan < num_channels; ++chan)
1971 loads[chan] = LLVMBuildSExt(ctx->builder, loads[chan], ctx->i32, "");
1972 }
1973 } else {
1974 bool unsign = format == AC_FETCH_FORMAT_UNORM ||
1975 format == AC_FETCH_FORMAT_USCALED ||
1976 format == AC_FETCH_FORMAT_UINT;
1977
1978 for (unsigned chan = 0; chan < num_channels; ++chan) {
1979 if (unsign) {
1980 tmp = LLVMBuildUIToFP(ctx->builder, loads[chan], ctx->f32, "");
1981 } else {
1982 tmp = LLVMBuildSIToFP(ctx->builder, loads[chan], ctx->f32, "");
1983 }
1984
1985 LLVMValueRef scale = NULL;
1986 if (format == AC_FETCH_FORMAT_FIXED) {
1987 assert(log_size == 2);
1988 scale = LLVMConstReal(ctx->f32, 1.0 / 0x10000);
1989 } else if (format == AC_FETCH_FORMAT_UNORM) {
1990 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
1991 scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << bits) - 1));
1992 } else if (format == AC_FETCH_FORMAT_SNORM) {
1993 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
1994 scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << (bits - 1)) - 1));
1995 }
1996 if (scale)
1997 tmp = LLVMBuildFMul(ctx->builder, tmp, scale, "");
1998
1999 if (format == AC_FETCH_FORMAT_SNORM) {
2000 /* Clamp to [-1, 1] */
2001 LLVMValueRef neg_one = LLVMConstReal(ctx->f32, -1.0);
2002 LLVMValueRef clamp =
2003 LLVMBuildFCmp(ctx->builder, LLVMRealULT, tmp, neg_one, "");
2004 tmp = LLVMBuildSelect(ctx->builder, clamp, neg_one, tmp, "");
2005 }
2006
2007 loads[chan] = ac_to_integer(ctx, tmp);
2008 }
2009 }
2010
2011 while (num_channels < 4) {
2012 if (format == AC_FETCH_FORMAT_UINT || format == AC_FETCH_FORMAT_SINT) {
2013 loads[num_channels] = num_channels == 3 ? ctx->i32_1 : ctx->i32_0;
2014 } else {
2015 loads[num_channels] = ac_to_integer(ctx, num_channels == 3 ? ctx->f32_1 : ctx->f32_0);
2016 }
2017 num_channels++;
2018 }
2019
2020 if (reverse) {
2021 tmp = loads[0];
2022 loads[0] = loads[2];
2023 loads[2] = tmp;
2024 }
2025
2026 return ac_build_gather_values(ctx, loads, 4);
2027 }
2028
2029 static void
2030 ac_build_llvm8_tbuffer_store(struct ac_llvm_context *ctx,
2031 LLVMValueRef rsrc,
2032 LLVMValueRef vdata,
2033 LLVMValueRef vindex,
2034 LLVMValueRef voffset,
2035 LLVMValueRef soffset,
2036 unsigned num_channels,
2037 unsigned dfmt,
2038 unsigned nfmt,
2039 bool glc,
2040 bool slc,
2041 bool structurized)
2042 {
2043 LLVMValueRef args[7];
2044 int idx = 0;
2045 args[idx++] = vdata;
2046 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
2047 if (structurized)
2048 args[idx++] = vindex ? vindex : ctx->i32_0;
2049 args[idx++] = voffset ? voffset : ctx->i32_0;
2050 args[idx++] = soffset ? soffset : ctx->i32_0;
2051 args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx, dfmt, nfmt), 0);
2052 args[idx++] = LLVMConstInt(ctx->i32, (glc ? 1 : 0) + (slc ? 2 : 0), 0);
2053 unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
2054 const char *indexing_kind = structurized ? "struct" : "raw";
2055 char name[256], type_name[8];
2056
2057 LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
2058 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
2059
2060 snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.store.%s",
2061 indexing_kind, type_name);
2062
2063 ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
2064 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
2065 }
2066
2067 static void
2068 ac_build_tbuffer_store(struct ac_llvm_context *ctx,
2069 LLVMValueRef rsrc,
2070 LLVMValueRef vdata,
2071 LLVMValueRef vindex,
2072 LLVMValueRef voffset,
2073 LLVMValueRef soffset,
2074 LLVMValueRef immoffset,
2075 unsigned num_channels,
2076 unsigned dfmt,
2077 unsigned nfmt,
2078 bool glc,
2079 bool slc,
2080 bool structurized) /* only matters for LLVM 8+ */
2081 {
2082 if (HAVE_LLVM >= 0x800) {
2083 voffset = LLVMBuildAdd(ctx->builder,
2084 voffset ? voffset : ctx->i32_0,
2085 immoffset, "");
2086
2087 ac_build_llvm8_tbuffer_store(ctx, rsrc, vdata, vindex, voffset,
2088 soffset, num_channels, dfmt, nfmt,
2089 glc, slc, structurized);
2090 } else {
2091 LLVMValueRef params[] = {
2092 vdata,
2093 rsrc,
2094 vindex ? vindex : ctx->i32_0,
2095 voffset ? voffset : ctx->i32_0,
2096 soffset ? soffset : ctx->i32_0,
2097 immoffset,
2098 LLVMConstInt(ctx->i32, dfmt, false),
2099 LLVMConstInt(ctx->i32, nfmt, false),
2100 LLVMConstInt(ctx->i1, glc, false),
2101 LLVMConstInt(ctx->i1, slc, false),
2102 };
2103 unsigned func = CLAMP(num_channels, 1, 3) - 1;
2104 const char *type_names[] = {"i32", "v2i32", "v4i32"};
2105 char name[256];
2106
2107 snprintf(name, sizeof(name), "llvm.amdgcn.tbuffer.store.%s",
2108 type_names[func]);
2109
2110 ac_build_intrinsic(ctx, name, ctx->voidt, params, 10,
2111 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
2112 }
2113 }
2114
2115 void
2116 ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
2117 LLVMValueRef rsrc,
2118 LLVMValueRef vdata,
2119 LLVMValueRef vindex,
2120 LLVMValueRef voffset,
2121 LLVMValueRef soffset,
2122 LLVMValueRef immoffset,
2123 unsigned num_channels,
2124 unsigned dfmt,
2125 unsigned nfmt,
2126 bool glc,
2127 bool slc)
2128 {
2129 ac_build_tbuffer_store(ctx, rsrc, vdata, vindex, voffset, soffset,
2130 immoffset, num_channels, dfmt, nfmt, glc, slc,
2131 true);
2132 }
2133
2134 void
2135 ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
2136 LLVMValueRef rsrc,
2137 LLVMValueRef vdata,
2138 LLVMValueRef voffset,
2139 LLVMValueRef soffset,
2140 LLVMValueRef immoffset,
2141 unsigned num_channels,
2142 unsigned dfmt,
2143 unsigned nfmt,
2144 bool glc,
2145 bool slc)
2146 {
2147 ac_build_tbuffer_store(ctx, rsrc, vdata, NULL, voffset, soffset,
2148 immoffset, num_channels, dfmt, nfmt, glc, slc,
2149 false);
2150 }
2151
2152 void
2153 ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
2154 LLVMValueRef rsrc,
2155 LLVMValueRef vdata,
2156 LLVMValueRef voffset,
2157 LLVMValueRef soffset,
2158 bool glc)
2159 {
2160 vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i16, "");
2161
2162 if (HAVE_LLVM >= 0x900) {
2163 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
2164 ac_build_llvm8_buffer_store_common(ctx, rsrc, vdata, NULL,
2165 voffset, soffset, 1,
2166 ctx->i16, glc, false,
2167 false, false);
2168 } else {
2169 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
2170 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
2171
2172 vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
2173
2174 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
2175 ctx->i32_0, 1, dfmt, nfmt, glc,
2176 false);
2177 }
2178 }
2179
2180 void
2181 ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
2182 LLVMValueRef rsrc,
2183 LLVMValueRef vdata,
2184 LLVMValueRef voffset,
2185 LLVMValueRef soffset,
2186 bool glc)
2187 {
2188 vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i8, "");
2189
2190 if (HAVE_LLVM >= 0x900) {
2191 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
2192 ac_build_llvm8_buffer_store_common(ctx, rsrc, vdata, NULL,
2193 voffset, soffset, 1,
2194 ctx->i8, glc, false,
2195 false, false);
2196 } else {
2197 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
2198 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
2199
2200 vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
2201
2202 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
2203 ctx->i32_0, 1, dfmt, nfmt, glc, false);
2204 }
2205 }
2206 /**
2207 * Set range metadata on an instruction. This can only be used on load and
2208 * call instructions. If you know an instruction can only produce the values
2209 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
2210 * \p lo is the minimum value inclusive.
2211 * \p hi is the maximum value exclusive.
2212 */
2213 static void set_range_metadata(struct ac_llvm_context *ctx,
2214 LLVMValueRef value, unsigned lo, unsigned hi)
2215 {
2216 LLVMValueRef range_md, md_args[2];
2217 LLVMTypeRef type = LLVMTypeOf(value);
2218 LLVMContextRef context = LLVMGetTypeContext(type);
2219
2220 md_args[0] = LLVMConstInt(type, lo, false);
2221 md_args[1] = LLVMConstInt(type, hi, false);
2222 range_md = LLVMMDNodeInContext(context, md_args, 2);
2223 LLVMSetMetadata(value, ctx->range_md_kind, range_md);
2224 }
2225
2226 LLVMValueRef
2227 ac_get_thread_id(struct ac_llvm_context *ctx)
2228 {
2229 LLVMValueRef tid;
2230
2231 LLVMValueRef tid_args[2];
2232 tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
2233 tid_args[1] = ctx->i32_0;
2234 tid_args[1] = ac_build_intrinsic(ctx,
2235 "llvm.amdgcn.mbcnt.lo", ctx->i32,
2236 tid_args, 2, AC_FUNC_ATTR_READNONE);
2237
2238 tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
2239 ctx->i32, tid_args,
2240 2, AC_FUNC_ATTR_READNONE);
2241 set_range_metadata(ctx, tid, 0, 64);
2242 return tid;
2243 }
2244
2245 /*
2246 * AMD GCN implements derivatives using the local data store (LDS)
2247 * All writes to the LDS happen in all executing threads at
2248 * the same time. TID is the Thread ID for the current
2249 * thread and is a value between 0 and 63, representing
2250 * the thread's position in the wavefront.
2251 *
2252 * For the pixel shader threads are grouped into quads of four pixels.
2253 * The TIDs of the pixels of a quad are:
2254 *
2255 * +------+------+
2256 * |4n + 0|4n + 1|
2257 * +------+------+
2258 * |4n + 2|4n + 3|
2259 * +------+------+
2260 *
2261 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
2262 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
2263 * the current pixel's column, and masking with 0xfffffffe yields the TID
2264 * of the left pixel of the current pixel's row.
2265 *
2266 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
2267 * adding 2 yields the TID of the pixel below the top pixel.
2268 */
2269 LLVMValueRef
2270 ac_build_ddxy(struct ac_llvm_context *ctx,
2271 uint32_t mask,
2272 int idx,
2273 LLVMValueRef val)
2274 {
2275 unsigned tl_lanes[4], trbl_lanes[4];
2276 char name[32], type[8];
2277 LLVMValueRef tl, trbl;
2278 LLVMTypeRef result_type;
2279 LLVMValueRef result;
2280
2281 result_type = ac_to_float_type(ctx, LLVMTypeOf(val));
2282
2283 if (result_type == ctx->f16)
2284 val = LLVMBuildZExt(ctx->builder, val, ctx->i32, "");
2285
2286 for (unsigned i = 0; i < 4; ++i) {
2287 tl_lanes[i] = i & mask;
2288 trbl_lanes[i] = (i & mask) + idx;
2289 }
2290
2291 tl = ac_build_quad_swizzle(ctx, val,
2292 tl_lanes[0], tl_lanes[1],
2293 tl_lanes[2], tl_lanes[3]);
2294 trbl = ac_build_quad_swizzle(ctx, val,
2295 trbl_lanes[0], trbl_lanes[1],
2296 trbl_lanes[2], trbl_lanes[3]);
2297
2298 if (result_type == ctx->f16) {
2299 tl = LLVMBuildTrunc(ctx->builder, tl, ctx->i16, "");
2300 trbl = LLVMBuildTrunc(ctx->builder, trbl, ctx->i16, "");
2301 }
2302
2303 tl = LLVMBuildBitCast(ctx->builder, tl, result_type, "");
2304 trbl = LLVMBuildBitCast(ctx->builder, trbl, result_type, "");
2305 result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
2306
2307 ac_build_type_name_for_intr(result_type, type, sizeof(type));
2308 snprintf(name, sizeof(name), "llvm.amdgcn.wqm.%s", type);
2309
2310 return ac_build_intrinsic(ctx, name, result_type, &result, 1, 0);
2311 }
2312
2313 void
2314 ac_build_sendmsg(struct ac_llvm_context *ctx,
2315 uint32_t msg,
2316 LLVMValueRef wave_id)
2317 {
2318 LLVMValueRef args[2];
2319 args[0] = LLVMConstInt(ctx->i32, msg, false);
2320 args[1] = wave_id;
2321 ac_build_intrinsic(ctx, "llvm.amdgcn.s.sendmsg", ctx->voidt, args, 2, 0);
2322 }
2323
2324 LLVMValueRef
2325 ac_build_imsb(struct ac_llvm_context *ctx,
2326 LLVMValueRef arg,
2327 LLVMTypeRef dst_type)
2328 {
2329 LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.amdgcn.sffbh.i32",
2330 dst_type, &arg, 1,
2331 AC_FUNC_ATTR_READNONE);
2332
2333 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
2334 * the index from LSB. Invert it by doing "31 - msb". */
2335 msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
2336 msb, "");
2337
2338 LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
2339 LLVMValueRef cond = LLVMBuildOr(ctx->builder,
2340 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
2341 arg, ctx->i32_0, ""),
2342 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
2343 arg, all_ones, ""), "");
2344
2345 return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
2346 }
2347
2348 LLVMValueRef
2349 ac_build_umsb(struct ac_llvm_context *ctx,
2350 LLVMValueRef arg,
2351 LLVMTypeRef dst_type)
2352 {
2353 const char *intrin_name;
2354 LLVMTypeRef type;
2355 LLVMValueRef highest_bit;
2356 LLVMValueRef zero;
2357 unsigned bitsize;
2358
2359 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(arg));
2360 switch (bitsize) {
2361 case 64:
2362 intrin_name = "llvm.ctlz.i64";
2363 type = ctx->i64;
2364 highest_bit = LLVMConstInt(ctx->i64, 63, false);
2365 zero = ctx->i64_0;
2366 break;
2367 case 32:
2368 intrin_name = "llvm.ctlz.i32";
2369 type = ctx->i32;
2370 highest_bit = LLVMConstInt(ctx->i32, 31, false);
2371 zero = ctx->i32_0;
2372 break;
2373 case 16:
2374 intrin_name = "llvm.ctlz.i16";
2375 type = ctx->i16;
2376 highest_bit = LLVMConstInt(ctx->i16, 15, false);
2377 zero = ctx->i16_0;
2378 break;
2379 case 8:
2380 intrin_name = "llvm.ctlz.i8";
2381 type = ctx->i8;
2382 highest_bit = LLVMConstInt(ctx->i8, 7, false);
2383 zero = ctx->i8_0;
2384 break;
2385 default:
2386 unreachable(!"invalid bitsize");
2387 break;
2388 }
2389
2390 LLVMValueRef params[2] = {
2391 arg,
2392 ctx->i1true,
2393 };
2394
2395 LLVMValueRef msb = ac_build_intrinsic(ctx, intrin_name, type,
2396 params, 2,
2397 AC_FUNC_ATTR_READNONE);
2398
2399 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
2400 * the index from LSB. Invert it by doing "31 - msb". */
2401 msb = LLVMBuildSub(ctx->builder, highest_bit, msb, "");
2402
2403 if (bitsize == 64) {
2404 msb = LLVMBuildTrunc(ctx->builder, msb, ctx->i32, "");
2405 } else if (bitsize < 32) {
2406 msb = LLVMBuildSExt(ctx->builder, msb, ctx->i32, "");
2407 }
2408
2409 /* check for zero */
2410 return LLVMBuildSelect(ctx->builder,
2411 LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg, zero, ""),
2412 LLVMConstInt(ctx->i32, -1, true), msb, "");
2413 }
2414
2415 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
2416 LLVMValueRef b)
2417 {
2418 char name[64];
2419 snprintf(name, sizeof(name), "llvm.minnum.f%d", ac_get_elem_bits(ctx, LLVMTypeOf(a)));
2420 LLVMValueRef args[2] = {a, b};
2421 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
2422 AC_FUNC_ATTR_READNONE);
2423 }
2424
2425 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
2426 LLVMValueRef b)
2427 {
2428 char name[64];
2429 snprintf(name, sizeof(name), "llvm.maxnum.f%d", ac_get_elem_bits(ctx, LLVMTypeOf(a)));
2430 LLVMValueRef args[2] = {a, b};
2431 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
2432 AC_FUNC_ATTR_READNONE);
2433 }
2434
2435 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
2436 LLVMValueRef b)
2437 {
2438 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSLE, a, b, "");
2439 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2440 }
2441
2442 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
2443 LLVMValueRef b)
2444 {
2445 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, a, b, "");
2446 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2447 }
2448
2449 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a,
2450 LLVMValueRef b)
2451 {
2452 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntULE, a, b, "");
2453 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2454 }
2455
2456 LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a,
2457 LLVMValueRef b)
2458 {
2459 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, a, b, "");
2460 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2461 }
2462
2463 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value)
2464 {
2465 LLVMTypeRef t = LLVMTypeOf(value);
2466 return ac_build_fmin(ctx, ac_build_fmax(ctx, value, LLVMConstReal(t, 0.0)),
2467 LLVMConstReal(t, 1.0));
2468 }
2469
2470 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a)
2471 {
2472 LLVMValueRef args[9];
2473
2474 args[0] = LLVMConstInt(ctx->i32, a->target, 0);
2475 args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
2476
2477 if (a->compr) {
2478 LLVMTypeRef i16 = LLVMInt16TypeInContext(ctx->context);
2479 LLVMTypeRef v2i16 = LLVMVectorType(i16, 2);
2480
2481 args[2] = LLVMBuildBitCast(ctx->builder, a->out[0],
2482 v2i16, "");
2483 args[3] = LLVMBuildBitCast(ctx->builder, a->out[1],
2484 v2i16, "");
2485 args[4] = LLVMConstInt(ctx->i1, a->done, 0);
2486 args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
2487
2488 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16",
2489 ctx->voidt, args, 6, 0);
2490 } else {
2491 args[2] = a->out[0];
2492 args[3] = a->out[1];
2493 args[4] = a->out[2];
2494 args[5] = a->out[3];
2495 args[6] = LLVMConstInt(ctx->i1, a->done, 0);
2496 args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
2497
2498 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.f32",
2499 ctx->voidt, args, 8, 0);
2500 }
2501 }
2502
2503 void ac_build_export_null(struct ac_llvm_context *ctx)
2504 {
2505 struct ac_export_args args;
2506
2507 args.enabled_channels = 0x0; /* enabled channels */
2508 args.valid_mask = 1; /* whether the EXEC mask is valid */
2509 args.done = 1; /* DONE bit */
2510 args.target = V_008DFC_SQ_EXP_NULL;
2511 args.compr = 0; /* COMPR flag (0 = 32-bit export) */
2512 args.out[0] = LLVMGetUndef(ctx->f32); /* R */
2513 args.out[1] = LLVMGetUndef(ctx->f32); /* G */
2514 args.out[2] = LLVMGetUndef(ctx->f32); /* B */
2515 args.out[3] = LLVMGetUndef(ctx->f32); /* A */
2516
2517 ac_build_export(ctx, &args);
2518 }
2519
2520 static unsigned ac_num_coords(enum ac_image_dim dim)
2521 {
2522 switch (dim) {
2523 case ac_image_1d:
2524 return 1;
2525 case ac_image_2d:
2526 case ac_image_1darray:
2527 return 2;
2528 case ac_image_3d:
2529 case ac_image_cube:
2530 case ac_image_2darray:
2531 case ac_image_2dmsaa:
2532 return 3;
2533 case ac_image_2darraymsaa:
2534 return 4;
2535 default:
2536 unreachable("ac_num_coords: bad dim");
2537 }
2538 }
2539
2540 static unsigned ac_num_derivs(enum ac_image_dim dim)
2541 {
2542 switch (dim) {
2543 case ac_image_1d:
2544 case ac_image_1darray:
2545 return 2;
2546 case ac_image_2d:
2547 case ac_image_2darray:
2548 case ac_image_cube:
2549 return 4;
2550 case ac_image_3d:
2551 return 6;
2552 case ac_image_2dmsaa:
2553 case ac_image_2darraymsaa:
2554 default:
2555 unreachable("derivatives not supported");
2556 }
2557 }
2558
2559 static const char *get_atomic_name(enum ac_atomic_op op)
2560 {
2561 switch (op) {
2562 case ac_atomic_swap: return "swap";
2563 case ac_atomic_add: return "add";
2564 case ac_atomic_sub: return "sub";
2565 case ac_atomic_smin: return "smin";
2566 case ac_atomic_umin: return "umin";
2567 case ac_atomic_smax: return "smax";
2568 case ac_atomic_umax: return "umax";
2569 case ac_atomic_and: return "and";
2570 case ac_atomic_or: return "or";
2571 case ac_atomic_xor: return "xor";
2572 }
2573 unreachable("bad atomic op");
2574 }
2575
2576 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
2577 struct ac_image_args *a)
2578 {
2579 const char *overload[3] = { "", "", "" };
2580 unsigned num_overloads = 0;
2581 LLVMValueRef args[18];
2582 unsigned num_args = 0;
2583 enum ac_image_dim dim = a->dim;
2584
2585 assert(!a->lod || a->lod == ctx->i32_0 || a->lod == ctx->f32_0 ||
2586 !a->level_zero);
2587 assert((a->opcode != ac_image_get_resinfo && a->opcode != ac_image_load_mip &&
2588 a->opcode != ac_image_store_mip) ||
2589 a->lod);
2590 assert(a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
2591 (!a->compare && !a->offset));
2592 assert((a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
2593 a->opcode == ac_image_get_lod) ||
2594 !a->bias);
2595 assert((a->bias ? 1 : 0) +
2596 (a->lod ? 1 : 0) +
2597 (a->level_zero ? 1 : 0) +
2598 (a->derivs[0] ? 1 : 0) <= 1);
2599
2600 if (a->opcode == ac_image_get_lod) {
2601 switch (dim) {
2602 case ac_image_1darray:
2603 dim = ac_image_1d;
2604 break;
2605 case ac_image_2darray:
2606 case ac_image_cube:
2607 dim = ac_image_2d;
2608 break;
2609 default:
2610 break;
2611 }
2612 }
2613
2614 bool sample = a->opcode == ac_image_sample ||
2615 a->opcode == ac_image_gather4 ||
2616 a->opcode == ac_image_get_lod;
2617 bool atomic = a->opcode == ac_image_atomic ||
2618 a->opcode == ac_image_atomic_cmpswap;
2619 LLVMTypeRef coord_type = sample ? ctx->f32 : ctx->i32;
2620
2621 if (atomic || a->opcode == ac_image_store || a->opcode == ac_image_store_mip) {
2622 args[num_args++] = a->data[0];
2623 if (a->opcode == ac_image_atomic_cmpswap)
2624 args[num_args++] = a->data[1];
2625 }
2626
2627 if (!atomic)
2628 args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, false);
2629
2630 if (a->offset)
2631 args[num_args++] = ac_to_integer(ctx, a->offset);
2632 if (a->bias) {
2633 args[num_args++] = ac_to_float(ctx, a->bias);
2634 overload[num_overloads++] = ".f32";
2635 }
2636 if (a->compare)
2637 args[num_args++] = ac_to_float(ctx, a->compare);
2638 if (a->derivs[0]) {
2639 unsigned count = ac_num_derivs(dim);
2640 for (unsigned i = 0; i < count; ++i)
2641 args[num_args++] = ac_to_float(ctx, a->derivs[i]);
2642 overload[num_overloads++] = ".f32";
2643 }
2644 unsigned num_coords =
2645 a->opcode != ac_image_get_resinfo ? ac_num_coords(dim) : 0;
2646 for (unsigned i = 0; i < num_coords; ++i)
2647 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->coords[i], coord_type, "");
2648 if (a->lod)
2649 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->lod, coord_type, "");
2650 overload[num_overloads++] = sample ? ".f32" : ".i32";
2651
2652 args[num_args++] = a->resource;
2653 if (sample) {
2654 args[num_args++] = a->sampler;
2655 args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, false);
2656 }
2657
2658 args[num_args++] = ctx->i32_0; /* texfailctrl */
2659 args[num_args++] = LLVMConstInt(ctx->i32, a->cache_policy, false);
2660
2661 const char *name;
2662 const char *atomic_subop = "";
2663 switch (a->opcode) {
2664 case ac_image_sample: name = "sample"; break;
2665 case ac_image_gather4: name = "gather4"; break;
2666 case ac_image_load: name = "load"; break;
2667 case ac_image_load_mip: name = "load.mip"; break;
2668 case ac_image_store: name = "store"; break;
2669 case ac_image_store_mip: name = "store.mip"; break;
2670 case ac_image_atomic:
2671 name = "atomic.";
2672 atomic_subop = get_atomic_name(a->atomic);
2673 break;
2674 case ac_image_atomic_cmpswap:
2675 name = "atomic.";
2676 atomic_subop = "cmpswap";
2677 break;
2678 case ac_image_get_lod: name = "getlod"; break;
2679 case ac_image_get_resinfo: name = "getresinfo"; break;
2680 default: unreachable("invalid image opcode");
2681 }
2682
2683 const char *dimname;
2684 switch (dim) {
2685 case ac_image_1d: dimname = "1d"; break;
2686 case ac_image_2d: dimname = "2d"; break;
2687 case ac_image_3d: dimname = "3d"; break;
2688 case ac_image_cube: dimname = "cube"; break;
2689 case ac_image_1darray: dimname = "1darray"; break;
2690 case ac_image_2darray: dimname = "2darray"; break;
2691 case ac_image_2dmsaa: dimname = "2dmsaa"; break;
2692 case ac_image_2darraymsaa: dimname = "2darraymsaa"; break;
2693 default: unreachable("invalid dim");
2694 }
2695
2696 bool lod_suffix =
2697 a->lod && (a->opcode == ac_image_sample || a->opcode == ac_image_gather4);
2698 char intr_name[96];
2699 snprintf(intr_name, sizeof(intr_name),
2700 "llvm.amdgcn.image.%s%s" /* base name */
2701 "%s%s%s" /* sample/gather modifiers */
2702 ".%s.%s%s%s%s", /* dimension and type overloads */
2703 name, atomic_subop,
2704 a->compare ? ".c" : "",
2705 a->bias ? ".b" :
2706 lod_suffix ? ".l" :
2707 a->derivs[0] ? ".d" :
2708 a->level_zero ? ".lz" : "",
2709 a->offset ? ".o" : "",
2710 dimname,
2711 atomic ? "i32" : "v4f32",
2712 overload[0], overload[1], overload[2]);
2713
2714 LLVMTypeRef retty;
2715 if (atomic)
2716 retty = ctx->i32;
2717 else if (a->opcode == ac_image_store || a->opcode == ac_image_store_mip)
2718 retty = ctx->voidt;
2719 else
2720 retty = ctx->v4f32;
2721
2722 LLVMValueRef result =
2723 ac_build_intrinsic(ctx, intr_name, retty, args, num_args,
2724 a->attributes);
2725 if (!sample && retty == ctx->v4f32) {
2726 result = LLVMBuildBitCast(ctx->builder, result,
2727 ctx->v4i32, "");
2728 }
2729 return result;
2730 }
2731
2732 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
2733 LLVMValueRef args[2])
2734 {
2735 LLVMTypeRef v2f16 =
2736 LLVMVectorType(LLVMHalfTypeInContext(ctx->context), 2);
2737
2738 return ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz", v2f16,
2739 args, 2, AC_FUNC_ATTR_READNONE);
2740 }
2741
2742 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
2743 LLVMValueRef args[2])
2744 {
2745 LLVMValueRef res =
2746 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16",
2747 ctx->v2i16, args, 2,
2748 AC_FUNC_ATTR_READNONE);
2749 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2750 }
2751
2752 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
2753 LLVMValueRef args[2])
2754 {
2755 LLVMValueRef res =
2756 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16",
2757 ctx->v2i16, args, 2,
2758 AC_FUNC_ATTR_READNONE);
2759 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2760 }
2761
2762 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2763 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
2764 LLVMValueRef args[2], unsigned bits, bool hi)
2765 {
2766 assert(bits == 8 || bits == 10 || bits == 16);
2767
2768 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2769 bits == 8 ? 127 : bits == 10 ? 511 : 32767, 0);
2770 LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
2771 bits == 8 ? -128 : bits == 10 ? -512 : -32768, 0);
2772 LLVMValueRef max_alpha =
2773 bits != 10 ? max_rgb : ctx->i32_1;
2774 LLVMValueRef min_alpha =
2775 bits != 10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
2776
2777 /* Clamp. */
2778 if (bits != 16) {
2779 for (int i = 0; i < 2; i++) {
2780 bool alpha = hi && i == 1;
2781 args[i] = ac_build_imin(ctx, args[i],
2782 alpha ? max_alpha : max_rgb);
2783 args[i] = ac_build_imax(ctx, args[i],
2784 alpha ? min_alpha : min_rgb);
2785 }
2786 }
2787
2788 LLVMValueRef res =
2789 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16",
2790 ctx->v2i16, args, 2,
2791 AC_FUNC_ATTR_READNONE);
2792 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2793 }
2794
2795 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2796 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
2797 LLVMValueRef args[2], unsigned bits, bool hi)
2798 {
2799 assert(bits == 8 || bits == 10 || bits == 16);
2800
2801 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2802 bits == 8 ? 255 : bits == 10 ? 1023 : 65535, 0);
2803 LLVMValueRef max_alpha =
2804 bits != 10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
2805
2806 /* Clamp. */
2807 if (bits != 16) {
2808 for (int i = 0; i < 2; i++) {
2809 bool alpha = hi && i == 1;
2810 args[i] = ac_build_umin(ctx, args[i],
2811 alpha ? max_alpha : max_rgb);
2812 }
2813 }
2814
2815 LLVMValueRef res =
2816 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16",
2817 ctx->v2i16, args, 2,
2818 AC_FUNC_ATTR_READNONE);
2819 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2820 }
2821
2822 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1)
2823 {
2824 return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1,
2825 &i1, 1, AC_FUNC_ATTR_READNONE);
2826 }
2827
2828 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1)
2829 {
2830 ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt,
2831 &i1, 1, 0);
2832 }
2833
2834 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
2835 LLVMValueRef offset, LLVMValueRef width,
2836 bool is_signed)
2837 {
2838 LLVMValueRef args[] = {
2839 input,
2840 offset,
2841 width,
2842 };
2843
2844 LLVMValueRef result = ac_build_intrinsic(ctx,
2845 is_signed ? "llvm.amdgcn.sbfe.i32" :
2846 "llvm.amdgcn.ubfe.i32",
2847 ctx->i32, args, 3,
2848 AC_FUNC_ATTR_READNONE);
2849
2850 if (HAVE_LLVM < 0x0800) {
2851 /* FIXME: LLVM 7+ returns incorrect result when count is 0.
2852 * https://bugs.freedesktop.org/show_bug.cgi?id=107276
2853 */
2854 LLVMValueRef zero = ctx->i32_0;
2855 LLVMValueRef icond = LLVMBuildICmp(ctx->builder, LLVMIntEQ, width, zero, "");
2856 result = LLVMBuildSelect(ctx->builder, icond, zero, result, "");
2857 }
2858
2859 return result;
2860 }
2861
2862 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2863 LLVMValueRef s1, LLVMValueRef s2)
2864 {
2865 return LLVMBuildAdd(ctx->builder,
2866 LLVMBuildMul(ctx->builder, s0, s1, ""), s2, "");
2867 }
2868
2869 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2870 LLVMValueRef s1, LLVMValueRef s2)
2871 {
2872 return LLVMBuildFAdd(ctx->builder,
2873 LLVMBuildFMul(ctx->builder, s0, s1, ""), s2, "");
2874 }
2875
2876 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned simm16)
2877 {
2878 LLVMValueRef args[1] = {
2879 LLVMConstInt(ctx->i32, simm16, false),
2880 };
2881 ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt",
2882 ctx->voidt, args, 1, 0);
2883 }
2884
2885 LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
2886 LLVMValueRef src1, LLVMValueRef src2,
2887 unsigned bitsize)
2888 {
2889 LLVMTypeRef type;
2890 char *intr;
2891
2892 if (bitsize == 16) {
2893 intr = "llvm.amdgcn.fmed3.f16";
2894 type = ctx->f16;
2895 } else if (bitsize == 32) {
2896 intr = "llvm.amdgcn.fmed3.f32";
2897 type = ctx->f32;
2898 } else {
2899 intr = "llvm.amdgcn.fmed3.f64";
2900 type = ctx->f64;
2901 }
2902
2903 LLVMValueRef params[] = {
2904 src0,
2905 src1,
2906 src2,
2907 };
2908 return ac_build_intrinsic(ctx, intr, type, params, 3,
2909 AC_FUNC_ATTR_READNONE);
2910 }
2911
2912 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
2913 unsigned bitsize)
2914 {
2915 LLVMTypeRef type;
2916 char *intr;
2917
2918 if (bitsize == 16) {
2919 intr = "llvm.amdgcn.fract.f16";
2920 type = ctx->f16;
2921 } else if (bitsize == 32) {
2922 intr = "llvm.amdgcn.fract.f32";
2923 type = ctx->f32;
2924 } else {
2925 intr = "llvm.amdgcn.fract.f64";
2926 type = ctx->f64;
2927 }
2928
2929 LLVMValueRef params[] = {
2930 src0,
2931 };
2932 return ac_build_intrinsic(ctx, intr, type, params, 1,
2933 AC_FUNC_ATTR_READNONE);
2934 }
2935
2936 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2937 unsigned bitsize)
2938 {
2939 LLVMTypeRef type = LLVMIntTypeInContext(ctx->context, bitsize);
2940 LLVMValueRef zero = LLVMConstInt(type, 0, false);
2941 LLVMValueRef one = LLVMConstInt(type, 1, false);
2942
2943 LLVMValueRef cmp, val;
2944 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, src0, zero, "");
2945 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2946 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGE, val, zero, "");
2947 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstInt(type, -1, true), "");
2948 return val;
2949 }
2950
2951 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2952 unsigned bitsize)
2953 {
2954 LLVMValueRef cmp, val, zero, one;
2955 LLVMTypeRef type;
2956
2957 if (bitsize == 16) {
2958 type = ctx->f16;
2959 zero = ctx->f16_0;
2960 one = ctx->f16_1;
2961 } else if (bitsize == 32) {
2962 type = ctx->f32;
2963 zero = ctx->f32_0;
2964 one = ctx->f32_1;
2965 } else {
2966 type = ctx->f64;
2967 zero = ctx->f64_0;
2968 one = ctx->f64_1;
2969 }
2970
2971 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src0, zero, "");
2972 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2973 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGE, val, zero, "");
2974 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstReal(type, -1.0), "");
2975 return val;
2976 }
2977
2978 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0)
2979 {
2980 LLVMValueRef result;
2981 unsigned bitsize;
2982
2983 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
2984
2985 switch (bitsize) {
2986 case 64:
2987 result = ac_build_intrinsic(ctx, "llvm.ctpop.i64", ctx->i64,
2988 (LLVMValueRef []) { src0 }, 1,
2989 AC_FUNC_ATTR_READNONE);
2990
2991 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
2992 break;
2993 case 32:
2994 result = ac_build_intrinsic(ctx, "llvm.ctpop.i32", ctx->i32,
2995 (LLVMValueRef []) { src0 }, 1,
2996 AC_FUNC_ATTR_READNONE);
2997 break;
2998 case 16:
2999 result = ac_build_intrinsic(ctx, "llvm.ctpop.i16", ctx->i16,
3000 (LLVMValueRef []) { src0 }, 1,
3001 AC_FUNC_ATTR_READNONE);
3002
3003 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
3004 break;
3005 case 8:
3006 result = ac_build_intrinsic(ctx, "llvm.ctpop.i8", ctx->i8,
3007 (LLVMValueRef []) { src0 }, 1,
3008 AC_FUNC_ATTR_READNONE);
3009
3010 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
3011 break;
3012 default:
3013 unreachable(!"invalid bitsize");
3014 break;
3015 }
3016
3017 return result;
3018 }
3019
3020 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
3021 LLVMValueRef src0)
3022 {
3023 LLVMValueRef result;
3024 unsigned bitsize;
3025
3026 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
3027
3028 switch (bitsize) {
3029 case 64:
3030 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i64", ctx->i64,
3031 (LLVMValueRef []) { src0 }, 1,
3032 AC_FUNC_ATTR_READNONE);
3033
3034 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
3035 break;
3036 case 32:
3037 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i32", ctx->i32,
3038 (LLVMValueRef []) { src0 }, 1,
3039 AC_FUNC_ATTR_READNONE);
3040 break;
3041 case 16:
3042 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i16", ctx->i16,
3043 (LLVMValueRef []) { src0 }, 1,
3044 AC_FUNC_ATTR_READNONE);
3045
3046 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
3047 break;
3048 case 8:
3049 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i8", ctx->i8,
3050 (LLVMValueRef []) { src0 }, 1,
3051 AC_FUNC_ATTR_READNONE);
3052
3053 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
3054 break;
3055 default:
3056 unreachable(!"invalid bitsize");
3057 break;
3058 }
3059
3060 return result;
3061 }
3062
3063 #define AC_EXP_TARGET 0
3064 #define AC_EXP_ENABLED_CHANNELS 1
3065 #define AC_EXP_OUT0 2
3066
3067 enum ac_ir_type {
3068 AC_IR_UNDEF,
3069 AC_IR_CONST,
3070 AC_IR_VALUE,
3071 };
3072
3073 struct ac_vs_exp_chan
3074 {
3075 LLVMValueRef value;
3076 float const_float;
3077 enum ac_ir_type type;
3078 };
3079
3080 struct ac_vs_exp_inst {
3081 unsigned offset;
3082 LLVMValueRef inst;
3083 struct ac_vs_exp_chan chan[4];
3084 };
3085
3086 struct ac_vs_exports {
3087 unsigned num;
3088 struct ac_vs_exp_inst exp[VARYING_SLOT_MAX];
3089 };
3090
3091 /* Return true if the PARAM export has been eliminated. */
3092 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset,
3093 uint32_t num_outputs,
3094 struct ac_vs_exp_inst *exp)
3095 {
3096 unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
3097 bool is_zero[4] = {}, is_one[4] = {};
3098
3099 for (i = 0; i < 4; i++) {
3100 /* It's a constant expression. Undef outputs are eliminated too. */
3101 if (exp->chan[i].type == AC_IR_UNDEF) {
3102 is_zero[i] = true;
3103 is_one[i] = true;
3104 } else if (exp->chan[i].type == AC_IR_CONST) {
3105 if (exp->chan[i].const_float == 0)
3106 is_zero[i] = true;
3107 else if (exp->chan[i].const_float == 1)
3108 is_one[i] = true;
3109 else
3110 return false; /* other constant */
3111 } else
3112 return false;
3113 }
3114
3115 /* Only certain combinations of 0 and 1 can be eliminated. */
3116 if (is_zero[0] && is_zero[1] && is_zero[2])
3117 default_val = is_zero[3] ? 0 : 1;
3118 else if (is_one[0] && is_one[1] && is_one[2])
3119 default_val = is_zero[3] ? 2 : 3;
3120 else
3121 return false;
3122
3123 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
3124 LLVMInstructionEraseFromParent(exp->inst);
3125
3126 /* Change OFFSET to DEFAULT_VAL. */
3127 for (i = 0; i < num_outputs; i++) {
3128 if (vs_output_param_offset[i] == exp->offset) {
3129 vs_output_param_offset[i] =
3130 AC_EXP_PARAM_DEFAULT_VAL_0000 + default_val;
3131 break;
3132 }
3133 }
3134 return true;
3135 }
3136
3137 static bool ac_eliminate_duplicated_output(struct ac_llvm_context *ctx,
3138 uint8_t *vs_output_param_offset,
3139 uint32_t num_outputs,
3140 struct ac_vs_exports *processed,
3141 struct ac_vs_exp_inst *exp)
3142 {
3143 unsigned p, copy_back_channels = 0;
3144
3145 /* See if the output is already in the list of processed outputs.
3146 * The LLVMValueRef comparison relies on SSA.
3147 */
3148 for (p = 0; p < processed->num; p++) {
3149 bool different = false;
3150
3151 for (unsigned j = 0; j < 4; j++) {
3152 struct ac_vs_exp_chan *c1 = &processed->exp[p].chan[j];
3153 struct ac_vs_exp_chan *c2 = &exp->chan[j];
3154
3155 /* Treat undef as a match. */
3156 if (c2->type == AC_IR_UNDEF)
3157 continue;
3158
3159 /* If c1 is undef but c2 isn't, we can copy c2 to c1
3160 * and consider the instruction duplicated.
3161 */
3162 if (c1->type == AC_IR_UNDEF) {
3163 copy_back_channels |= 1 << j;
3164 continue;
3165 }
3166
3167 /* Test whether the channels are not equal. */
3168 if (c1->type != c2->type ||
3169 (c1->type == AC_IR_CONST &&
3170 c1->const_float != c2->const_float) ||
3171 (c1->type == AC_IR_VALUE &&
3172 c1->value != c2->value)) {
3173 different = true;
3174 break;
3175 }
3176 }
3177 if (!different)
3178 break;
3179
3180 copy_back_channels = 0;
3181 }
3182 if (p == processed->num)
3183 return false;
3184
3185 /* If a match was found, but the matching export has undef where the new
3186 * one has a normal value, copy the normal value to the undef channel.
3187 */
3188 struct ac_vs_exp_inst *match = &processed->exp[p];
3189
3190 /* Get current enabled channels mask. */
3191 LLVMValueRef arg = LLVMGetOperand(match->inst, AC_EXP_ENABLED_CHANNELS);
3192 unsigned enabled_channels = LLVMConstIntGetZExtValue(arg);
3193
3194 while (copy_back_channels) {
3195 unsigned chan = u_bit_scan(&copy_back_channels);
3196
3197 assert(match->chan[chan].type == AC_IR_UNDEF);
3198 LLVMSetOperand(match->inst, AC_EXP_OUT0 + chan,
3199 exp->chan[chan].value);
3200 match->chan[chan] = exp->chan[chan];
3201
3202 /* Update number of enabled channels because the original mask
3203 * is not always 0xf.
3204 */
3205 enabled_channels |= (1 << chan);
3206 LLVMSetOperand(match->inst, AC_EXP_ENABLED_CHANNELS,
3207 LLVMConstInt(ctx->i32, enabled_channels, 0));
3208 }
3209
3210 /* The PARAM export is duplicated. Kill it. */
3211 LLVMInstructionEraseFromParent(exp->inst);
3212
3213 /* Change OFFSET to the matching export. */
3214 for (unsigned i = 0; i < num_outputs; i++) {
3215 if (vs_output_param_offset[i] == exp->offset) {
3216 vs_output_param_offset[i] = match->offset;
3217 break;
3218 }
3219 }
3220 return true;
3221 }
3222
3223 void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
3224 LLVMValueRef main_fn,
3225 uint8_t *vs_output_param_offset,
3226 uint32_t num_outputs,
3227 uint8_t *num_param_exports)
3228 {
3229 LLVMBasicBlockRef bb;
3230 bool removed_any = false;
3231 struct ac_vs_exports exports;
3232
3233 exports.num = 0;
3234
3235 /* Process all LLVM instructions. */
3236 bb = LLVMGetFirstBasicBlock(main_fn);
3237 while (bb) {
3238 LLVMValueRef inst = LLVMGetFirstInstruction(bb);
3239
3240 while (inst) {
3241 LLVMValueRef cur = inst;
3242 inst = LLVMGetNextInstruction(inst);
3243 struct ac_vs_exp_inst exp;
3244
3245 if (LLVMGetInstructionOpcode(cur) != LLVMCall)
3246 continue;
3247
3248 LLVMValueRef callee = ac_llvm_get_called_value(cur);
3249
3250 if (!ac_llvm_is_function(callee))
3251 continue;
3252
3253 const char *name = LLVMGetValueName(callee);
3254 unsigned num_args = LLVMCountParams(callee);
3255
3256 /* Check if this is an export instruction. */
3257 if ((num_args != 9 && num_args != 8) ||
3258 (strcmp(name, "llvm.SI.export") &&
3259 strcmp(name, "llvm.amdgcn.exp.f32")))
3260 continue;
3261
3262 LLVMValueRef arg = LLVMGetOperand(cur, AC_EXP_TARGET);
3263 unsigned target = LLVMConstIntGetZExtValue(arg);
3264
3265 if (target < V_008DFC_SQ_EXP_PARAM)
3266 continue;
3267
3268 target -= V_008DFC_SQ_EXP_PARAM;
3269
3270 /* Parse the instruction. */
3271 memset(&exp, 0, sizeof(exp));
3272 exp.offset = target;
3273 exp.inst = cur;
3274
3275 for (unsigned i = 0; i < 4; i++) {
3276 LLVMValueRef v = LLVMGetOperand(cur, AC_EXP_OUT0 + i);
3277
3278 exp.chan[i].value = v;
3279
3280 if (LLVMIsUndef(v)) {
3281 exp.chan[i].type = AC_IR_UNDEF;
3282 } else if (LLVMIsAConstantFP(v)) {
3283 LLVMBool loses_info;
3284 exp.chan[i].type = AC_IR_CONST;
3285 exp.chan[i].const_float =
3286 LLVMConstRealGetDouble(v, &loses_info);
3287 } else {
3288 exp.chan[i].type = AC_IR_VALUE;
3289 }
3290 }
3291
3292 /* Eliminate constant and duplicated PARAM exports. */
3293 if (ac_eliminate_const_output(vs_output_param_offset,
3294 num_outputs, &exp) ||
3295 ac_eliminate_duplicated_output(ctx,
3296 vs_output_param_offset,
3297 num_outputs, &exports,
3298 &exp)) {
3299 removed_any = true;
3300 } else {
3301 exports.exp[exports.num++] = exp;
3302 }
3303 }
3304 bb = LLVMGetNextBasicBlock(bb);
3305 }
3306
3307 /* Remove holes in export memory due to removed PARAM exports.
3308 * This is done by renumbering all PARAM exports.
3309 */
3310 if (removed_any) {
3311 uint8_t old_offset[VARYING_SLOT_MAX];
3312 unsigned out, i;
3313
3314 /* Make a copy of the offsets. We need the old version while
3315 * we are modifying some of them. */
3316 memcpy(old_offset, vs_output_param_offset,
3317 sizeof(old_offset));
3318
3319 for (i = 0; i < exports.num; i++) {
3320 unsigned offset = exports.exp[i].offset;
3321
3322 /* Update vs_output_param_offset. Multiple outputs can
3323 * have the same offset.
3324 */
3325 for (out = 0; out < num_outputs; out++) {
3326 if (old_offset[out] == offset)
3327 vs_output_param_offset[out] = i;
3328 }
3329
3330 /* Change the PARAM offset in the instruction. */
3331 LLVMSetOperand(exports.exp[i].inst, AC_EXP_TARGET,
3332 LLVMConstInt(ctx->i32,
3333 V_008DFC_SQ_EXP_PARAM + i, 0));
3334 }
3335 *num_param_exports = exports.num;
3336 }
3337 }
3338
3339 void ac_init_exec_full_mask(struct ac_llvm_context *ctx)
3340 {
3341 LLVMValueRef full_mask = LLVMConstInt(ctx->i64, ~0ull, 0);
3342 ac_build_intrinsic(ctx,
3343 "llvm.amdgcn.init.exec", ctx->voidt,
3344 &full_mask, 1, AC_FUNC_ATTR_CONVERGENT);
3345 }
3346
3347 void ac_declare_lds_as_pointer(struct ac_llvm_context *ctx)
3348 {
3349 unsigned lds_size = ctx->chip_class >= GFX7 ? 65536 : 32768;
3350 ctx->lds = LLVMBuildIntToPtr(ctx->builder, ctx->i32_0,
3351 LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), AC_ADDR_SPACE_LDS),
3352 "lds");
3353 }
3354
3355 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
3356 LLVMValueRef dw_addr)
3357 {
3358 return LLVMBuildLoad(ctx->builder, ac_build_gep0(ctx, ctx->lds, dw_addr), "");
3359 }
3360
3361 void ac_lds_store(struct ac_llvm_context *ctx,
3362 LLVMValueRef dw_addr,
3363 LLVMValueRef value)
3364 {
3365 value = ac_to_integer(ctx, value);
3366 ac_build_indexed_store(ctx, ctx->lds,
3367 dw_addr, value);
3368 }
3369
3370 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
3371 LLVMTypeRef dst_type,
3372 LLVMValueRef src0)
3373 {
3374 unsigned src0_bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
3375 const char *intrin_name;
3376 LLVMTypeRef type;
3377 LLVMValueRef zero;
3378
3379 switch (src0_bitsize) {
3380 case 64:
3381 intrin_name = "llvm.cttz.i64";
3382 type = ctx->i64;
3383 zero = ctx->i64_0;
3384 break;
3385 case 32:
3386 intrin_name = "llvm.cttz.i32";
3387 type = ctx->i32;
3388 zero = ctx->i32_0;
3389 break;
3390 case 16:
3391 intrin_name = "llvm.cttz.i16";
3392 type = ctx->i16;
3393 zero = ctx->i16_0;
3394 break;
3395 case 8:
3396 intrin_name = "llvm.cttz.i8";
3397 type = ctx->i8;
3398 zero = ctx->i8_0;
3399 break;
3400 default:
3401 unreachable(!"invalid bitsize");
3402 }
3403
3404 LLVMValueRef params[2] = {
3405 src0,
3406
3407 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
3408 * add special code to check for x=0. The reason is that
3409 * the LLVM behavior for x=0 is different from what we
3410 * need here. However, LLVM also assumes that ffs(x) is
3411 * in [0, 31], but GLSL expects that ffs(0) = -1, so
3412 * a conditional assignment to handle 0 is still required.
3413 *
3414 * The hardware already implements the correct behavior.
3415 */
3416 ctx->i1true,
3417 };
3418
3419 LLVMValueRef lsb = ac_build_intrinsic(ctx, intrin_name, type,
3420 params, 2,
3421 AC_FUNC_ATTR_READNONE);
3422
3423 if (src0_bitsize == 64) {
3424 lsb = LLVMBuildTrunc(ctx->builder, lsb, ctx->i32, "");
3425 } else if (src0_bitsize < 32) {
3426 lsb = LLVMBuildSExt(ctx->builder, lsb, ctx->i32, "");
3427 }
3428
3429 /* TODO: We need an intrinsic to skip this conditional. */
3430 /* Check for zero: */
3431 return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder,
3432 LLVMIntEQ, src0,
3433 zero, ""),
3434 LLVMConstInt(ctx->i32, -1, 0), lsb, "");
3435 }
3436
3437 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type)
3438 {
3439 return LLVMPointerType(elem_type, AC_ADDR_SPACE_CONST);
3440 }
3441
3442 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type)
3443 {
3444 return LLVMPointerType(elem_type, AC_ADDR_SPACE_CONST_32BIT);
3445 }
3446
3447 static struct ac_llvm_flow *
3448 get_current_flow(struct ac_llvm_context *ctx)
3449 {
3450 if (ctx->flow_depth > 0)
3451 return &ctx->flow[ctx->flow_depth - 1];
3452 return NULL;
3453 }
3454
3455 static struct ac_llvm_flow *
3456 get_innermost_loop(struct ac_llvm_context *ctx)
3457 {
3458 for (unsigned i = ctx->flow_depth; i > 0; --i) {
3459 if (ctx->flow[i - 1].loop_entry_block)
3460 return &ctx->flow[i - 1];
3461 }
3462 return NULL;
3463 }
3464
3465 static struct ac_llvm_flow *
3466 push_flow(struct ac_llvm_context *ctx)
3467 {
3468 struct ac_llvm_flow *flow;
3469
3470 if (ctx->flow_depth >= ctx->flow_depth_max) {
3471 unsigned new_max = MAX2(ctx->flow_depth << 1,
3472 AC_LLVM_INITIAL_CF_DEPTH);
3473
3474 ctx->flow = realloc(ctx->flow, new_max * sizeof(*ctx->flow));
3475 ctx->flow_depth_max = new_max;
3476 }
3477
3478 flow = &ctx->flow[ctx->flow_depth];
3479 ctx->flow_depth++;
3480
3481 flow->next_block = NULL;
3482 flow->loop_entry_block = NULL;
3483 return flow;
3484 }
3485
3486 static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base,
3487 int label_id)
3488 {
3489 char buf[32];
3490 snprintf(buf, sizeof(buf), "%s%d", base, label_id);
3491 LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
3492 }
3493
3494 /* Append a basic block at the level of the parent flow.
3495 */
3496 static LLVMBasicBlockRef append_basic_block(struct ac_llvm_context *ctx,
3497 const char *name)
3498 {
3499 assert(ctx->flow_depth >= 1);
3500
3501 if (ctx->flow_depth >= 2) {
3502 struct ac_llvm_flow *flow = &ctx->flow[ctx->flow_depth - 2];
3503
3504 return LLVMInsertBasicBlockInContext(ctx->context,
3505 flow->next_block, name);
3506 }
3507
3508 LLVMValueRef main_fn =
3509 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->builder));
3510 return LLVMAppendBasicBlockInContext(ctx->context, main_fn, name);
3511 }
3512
3513 /* Emit a branch to the given default target for the current block if
3514 * applicable -- that is, if the current block does not already contain a
3515 * branch from a break or continue.
3516 */
3517 static void emit_default_branch(LLVMBuilderRef builder,
3518 LLVMBasicBlockRef target)
3519 {
3520 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
3521 LLVMBuildBr(builder, target);
3522 }
3523
3524 void ac_build_bgnloop(struct ac_llvm_context *ctx, int label_id)
3525 {
3526 struct ac_llvm_flow *flow = push_flow(ctx);
3527 flow->loop_entry_block = append_basic_block(ctx, "LOOP");
3528 flow->next_block = append_basic_block(ctx, "ENDLOOP");
3529 set_basicblock_name(flow->loop_entry_block, "loop", label_id);
3530 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
3531 LLVMPositionBuilderAtEnd(ctx->builder, flow->loop_entry_block);
3532 }
3533
3534 void ac_build_break(struct ac_llvm_context *ctx)
3535 {
3536 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
3537 LLVMBuildBr(ctx->builder, flow->next_block);
3538 }
3539
3540 void ac_build_continue(struct ac_llvm_context *ctx)
3541 {
3542 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
3543 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
3544 }
3545
3546 void ac_build_else(struct ac_llvm_context *ctx, int label_id)
3547 {
3548 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
3549 LLVMBasicBlockRef endif_block;
3550
3551 assert(!current_branch->loop_entry_block);
3552
3553 endif_block = append_basic_block(ctx, "ENDIF");
3554 emit_default_branch(ctx->builder, endif_block);
3555
3556 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
3557 set_basicblock_name(current_branch->next_block, "else", label_id);
3558
3559 current_branch->next_block = endif_block;
3560 }
3561
3562 void ac_build_endif(struct ac_llvm_context *ctx, int label_id)
3563 {
3564 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
3565
3566 assert(!current_branch->loop_entry_block);
3567
3568 emit_default_branch(ctx->builder, current_branch->next_block);
3569 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
3570 set_basicblock_name(current_branch->next_block, "endif", label_id);
3571
3572 ctx->flow_depth--;
3573 }
3574
3575 void ac_build_endloop(struct ac_llvm_context *ctx, int label_id)
3576 {
3577 struct ac_llvm_flow *current_loop = get_current_flow(ctx);
3578
3579 assert(current_loop->loop_entry_block);
3580
3581 emit_default_branch(ctx->builder, current_loop->loop_entry_block);
3582
3583 LLVMPositionBuilderAtEnd(ctx->builder, current_loop->next_block);
3584 set_basicblock_name(current_loop->next_block, "endloop", label_id);
3585 ctx->flow_depth--;
3586 }
3587
3588 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id)
3589 {
3590 struct ac_llvm_flow *flow = push_flow(ctx);
3591 LLVMBasicBlockRef if_block;
3592
3593 if_block = append_basic_block(ctx, "IF");
3594 flow->next_block = append_basic_block(ctx, "ELSE");
3595 set_basicblock_name(if_block, "if", label_id);
3596 LLVMBuildCondBr(ctx->builder, cond, if_block, flow->next_block);
3597 LLVMPositionBuilderAtEnd(ctx->builder, if_block);
3598 }
3599
3600 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
3601 int label_id)
3602 {
3603 LLVMValueRef cond = LLVMBuildFCmp(ctx->builder, LLVMRealUNE,
3604 value, ctx->f32_0, "");
3605 ac_build_ifcc(ctx, cond, label_id);
3606 }
3607
3608 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
3609 int label_id)
3610 {
3611 LLVMValueRef cond = LLVMBuildICmp(ctx->builder, LLVMIntNE,
3612 ac_to_integer(ctx, value),
3613 ctx->i32_0, "");
3614 ac_build_ifcc(ctx, cond, label_id);
3615 }
3616
3617 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
3618 const char *name)
3619 {
3620 LLVMBuilderRef builder = ac->builder;
3621 LLVMBasicBlockRef current_block = LLVMGetInsertBlock(builder);
3622 LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
3623 LLVMBasicBlockRef first_block = LLVMGetEntryBasicBlock(function);
3624 LLVMValueRef first_instr = LLVMGetFirstInstruction(first_block);
3625 LLVMBuilderRef first_builder = LLVMCreateBuilderInContext(ac->context);
3626 LLVMValueRef res;
3627
3628 if (first_instr) {
3629 LLVMPositionBuilderBefore(first_builder, first_instr);
3630 } else {
3631 LLVMPositionBuilderAtEnd(first_builder, first_block);
3632 }
3633
3634 res = LLVMBuildAlloca(first_builder, type, name);
3635 LLVMDisposeBuilder(first_builder);
3636 return res;
3637 }
3638
3639 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac,
3640 LLVMTypeRef type, const char *name)
3641 {
3642 LLVMValueRef ptr = ac_build_alloca_undef(ac, type, name);
3643 LLVMBuildStore(ac->builder, LLVMConstNull(type), ptr);
3644 return ptr;
3645 }
3646
3647 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
3648 LLVMTypeRef type)
3649 {
3650 int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
3651 return LLVMBuildBitCast(ctx->builder, ptr,
3652 LLVMPointerType(type, addr_space), "");
3653 }
3654
3655 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
3656 unsigned count)
3657 {
3658 unsigned num_components = ac_get_llvm_num_components(value);
3659 if (count == num_components)
3660 return value;
3661
3662 LLVMValueRef masks[MAX2(count, 2)];
3663 masks[0] = ctx->i32_0;
3664 masks[1] = ctx->i32_1;
3665 for (unsigned i = 2; i < count; i++)
3666 masks[i] = LLVMConstInt(ctx->i32, i, false);
3667
3668 if (count == 1)
3669 return LLVMBuildExtractElement(ctx->builder, value, masks[0],
3670 "");
3671
3672 LLVMValueRef swizzle = LLVMConstVector(masks, count);
3673 return LLVMBuildShuffleVector(ctx->builder, value, value, swizzle, "");
3674 }
3675
3676 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
3677 unsigned rshift, unsigned bitwidth)
3678 {
3679 LLVMValueRef value = param;
3680 if (rshift)
3681 value = LLVMBuildLShr(ctx->builder, value,
3682 LLVMConstInt(ctx->i32, rshift, false), "");
3683
3684 if (rshift + bitwidth < 32) {
3685 unsigned mask = (1 << bitwidth) - 1;
3686 value = LLVMBuildAnd(ctx->builder, value,
3687 LLVMConstInt(ctx->i32, mask, false), "");
3688 }
3689 return value;
3690 }
3691
3692 /* Adjust the sample index according to FMASK.
3693 *
3694 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3695 * which is the identity mapping. Each nibble says which physical sample
3696 * should be fetched to get that sample.
3697 *
3698 * For example, 0x11111100 means there are only 2 samples stored and
3699 * the second sample covers 3/4 of the pixel. When reading samples 0
3700 * and 1, return physical sample 0 (determined by the first two 0s
3701 * in FMASK), otherwise return physical sample 1.
3702 *
3703 * The sample index should be adjusted as follows:
3704 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
3705 */
3706 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
3707 LLVMValueRef *addr, bool is_array_tex)
3708 {
3709 struct ac_image_args fmask_load = {};
3710 fmask_load.opcode = ac_image_load;
3711 fmask_load.resource = fmask;
3712 fmask_load.dmask = 0xf;
3713 fmask_load.dim = is_array_tex ? ac_image_2darray : ac_image_2d;
3714 fmask_load.attributes = AC_FUNC_ATTR_READNONE;
3715
3716 fmask_load.coords[0] = addr[0];
3717 fmask_load.coords[1] = addr[1];
3718 if (is_array_tex)
3719 fmask_load.coords[2] = addr[2];
3720
3721 LLVMValueRef fmask_value = ac_build_image_opcode(ac, &fmask_load);
3722 fmask_value = LLVMBuildExtractElement(ac->builder, fmask_value,
3723 ac->i32_0, "");
3724
3725 /* Apply the formula. */
3726 unsigned sample_chan = is_array_tex ? 3 : 2;
3727 LLVMValueRef final_sample;
3728 final_sample = LLVMBuildMul(ac->builder, addr[sample_chan],
3729 LLVMConstInt(ac->i32, 4, 0), "");
3730 final_sample = LLVMBuildLShr(ac->builder, fmask_value, final_sample, "");
3731 /* Mask the sample index by 0x7, because 0x8 means an unknown value
3732 * with EQAA, so those will map to 0. */
3733 final_sample = LLVMBuildAnd(ac->builder, final_sample,
3734 LLVMConstInt(ac->i32, 0x7, 0), "");
3735
3736 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3737 * resource descriptor is 0 (invalid).
3738 */
3739 LLVMValueRef tmp;
3740 tmp = LLVMBuildBitCast(ac->builder, fmask, ac->v8i32, "");
3741 tmp = LLVMBuildExtractElement(ac->builder, tmp, ac->i32_1, "");
3742 tmp = LLVMBuildICmp(ac->builder, LLVMIntNE, tmp, ac->i32_0, "");
3743
3744 /* Replace the MSAA sample index. */
3745 addr[sample_chan] = LLVMBuildSelect(ac->builder, tmp, final_sample,
3746 addr[sample_chan], "");
3747 }
3748
3749 static LLVMValueRef
3750 _ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
3751 {
3752 ac_build_optimization_barrier(ctx, &src);
3753 return ac_build_intrinsic(ctx,
3754 lane == NULL ? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
3755 LLVMTypeOf(src), (LLVMValueRef []) {
3756 src, lane },
3757 lane == NULL ? 1 : 2,
3758 AC_FUNC_ATTR_READNONE |
3759 AC_FUNC_ATTR_CONVERGENT);
3760 }
3761
3762 /**
3763 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
3764 * @param ctx
3765 * @param src
3766 * @param lane - id of the lane or NULL for the first active lane
3767 * @return value of the lane
3768 */
3769 LLVMValueRef
3770 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
3771 {
3772 LLVMTypeRef src_type = LLVMTypeOf(src);
3773 src = ac_to_integer(ctx, src);
3774 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3775 LLVMValueRef ret;
3776
3777 if (bits == 32) {
3778 ret = _ac_build_readlane(ctx, src, lane);
3779 } else {
3780 assert(bits % 32 == 0);
3781 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3782 LLVMValueRef src_vector =
3783 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3784 ret = LLVMGetUndef(vec_type);
3785 for (unsigned i = 0; i < bits / 32; i++) {
3786 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3787 LLVMConstInt(ctx->i32, i, 0), "");
3788 LLVMValueRef ret_comp = _ac_build_readlane(ctx, src, lane);
3789 ret = LLVMBuildInsertElement(ctx->builder, ret, ret_comp,
3790 LLVMConstInt(ctx->i32, i, 0), "");
3791 }
3792 }
3793 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3794 }
3795
3796 LLVMValueRef
3797 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane)
3798 {
3799 /* TODO: Use the actual instruction when LLVM adds an intrinsic for it.
3800 */
3801 LLVMValueRef pred = LLVMBuildICmp(ctx->builder, LLVMIntEQ, lane,
3802 ac_get_thread_id(ctx), "");
3803 return LLVMBuildSelect(ctx->builder, pred, value, src, "");
3804 }
3805
3806 LLVMValueRef
3807 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask)
3808 {
3809 LLVMValueRef mask_vec = LLVMBuildBitCast(ctx->builder, mask,
3810 LLVMVectorType(ctx->i32, 2),
3811 "");
3812 LLVMValueRef mask_lo = LLVMBuildExtractElement(ctx->builder, mask_vec,
3813 ctx->i32_0, "");
3814 LLVMValueRef mask_hi = LLVMBuildExtractElement(ctx->builder, mask_vec,
3815 ctx->i32_1, "");
3816 LLVMValueRef val =
3817 ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
3818 (LLVMValueRef []) { mask_lo, ctx->i32_0 },
3819 2, AC_FUNC_ATTR_READNONE);
3820 val = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi", ctx->i32,
3821 (LLVMValueRef []) { mask_hi, val },
3822 2, AC_FUNC_ATTR_READNONE);
3823 return val;
3824 }
3825
3826 enum dpp_ctrl {
3827 _dpp_quad_perm = 0x000,
3828 _dpp_row_sl = 0x100,
3829 _dpp_row_sr = 0x110,
3830 _dpp_row_rr = 0x120,
3831 dpp_wf_sl1 = 0x130,
3832 dpp_wf_rl1 = 0x134,
3833 dpp_wf_sr1 = 0x138,
3834 dpp_wf_rr1 = 0x13C,
3835 dpp_row_mirror = 0x140,
3836 dpp_row_half_mirror = 0x141,
3837 dpp_row_bcast15 = 0x142,
3838 dpp_row_bcast31 = 0x143
3839 };
3840
3841 static inline enum dpp_ctrl
3842 dpp_quad_perm(unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
3843 {
3844 assert(lane0 < 4 && lane1 < 4 && lane2 < 4 && lane3 < 4);
3845 return _dpp_quad_perm | lane0 | (lane1 << 2) | (lane2 << 4) | (lane3 << 6);
3846 }
3847
3848 static inline enum dpp_ctrl
3849 dpp_row_sl(unsigned amount)
3850 {
3851 assert(amount > 0 && amount < 16);
3852 return _dpp_row_sl | amount;
3853 }
3854
3855 static inline enum dpp_ctrl
3856 dpp_row_sr(unsigned amount)
3857 {
3858 assert(amount > 0 && amount < 16);
3859 return _dpp_row_sr | amount;
3860 }
3861
3862 static LLVMValueRef
3863 _ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3864 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3865 bool bound_ctrl)
3866 {
3867 return ac_build_intrinsic(ctx, "llvm.amdgcn.update.dpp.i32",
3868 LLVMTypeOf(old),
3869 (LLVMValueRef[]) {
3870 old, src,
3871 LLVMConstInt(ctx->i32, dpp_ctrl, 0),
3872 LLVMConstInt(ctx->i32, row_mask, 0),
3873 LLVMConstInt(ctx->i32, bank_mask, 0),
3874 LLVMConstInt(ctx->i1, bound_ctrl, 0) },
3875 6, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3876 }
3877
3878 static LLVMValueRef
3879 ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3880 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3881 bool bound_ctrl)
3882 {
3883 LLVMTypeRef src_type = LLVMTypeOf(src);
3884 src = ac_to_integer(ctx, src);
3885 old = ac_to_integer(ctx, old);
3886 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3887 LLVMValueRef ret;
3888 if (bits == 32) {
3889 ret = _ac_build_dpp(ctx, old, src, dpp_ctrl, row_mask,
3890 bank_mask, bound_ctrl);
3891 } else {
3892 assert(bits % 32 == 0);
3893 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3894 LLVMValueRef src_vector =
3895 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3896 LLVMValueRef old_vector =
3897 LLVMBuildBitCast(ctx->builder, old, vec_type, "");
3898 ret = LLVMGetUndef(vec_type);
3899 for (unsigned i = 0; i < bits / 32; i++) {
3900 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3901 LLVMConstInt(ctx->i32, i,
3902 0), "");
3903 old = LLVMBuildExtractElement(ctx->builder, old_vector,
3904 LLVMConstInt(ctx->i32, i,
3905 0), "");
3906 LLVMValueRef ret_comp = _ac_build_dpp(ctx, old, src,
3907 dpp_ctrl,
3908 row_mask,
3909 bank_mask,
3910 bound_ctrl);
3911 ret = LLVMBuildInsertElement(ctx->builder, ret,
3912 ret_comp,
3913 LLVMConstInt(ctx->i32, i,
3914 0), "");
3915 }
3916 }
3917 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3918 }
3919
3920 static LLVMValueRef
3921 _ac_build_permlane16(struct ac_llvm_context *ctx, LLVMValueRef src, uint64_t sel,
3922 bool exchange_rows, bool bound_ctrl)
3923 {
3924 LLVMValueRef args[6] = {
3925 src,
3926 src,
3927 LLVMConstInt(ctx->i32, sel, false),
3928 LLVMConstInt(ctx->i32, sel >> 32, false),
3929 ctx->i1true, /* fi */
3930 bound_ctrl ? ctx->i1true : ctx->i1false,
3931 };
3932 return ac_build_intrinsic(ctx, exchange_rows ? "llvm.amdgcn.permlanex16"
3933 : "llvm.amdgcn.permlane16",
3934 ctx->i32, args, 6,
3935 AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3936 }
3937
3938 static LLVMValueRef
3939 ac_build_permlane16(struct ac_llvm_context *ctx, LLVMValueRef src, uint64_t sel,
3940 bool exchange_rows, bool bound_ctrl)
3941 {
3942 LLVMTypeRef src_type = LLVMTypeOf(src);
3943 src = ac_to_integer(ctx, src);
3944 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3945 LLVMValueRef ret;
3946 if (bits == 32) {
3947 ret = _ac_build_permlane16(ctx, src, sel, exchange_rows,
3948 bound_ctrl);
3949 } else {
3950 assert(bits % 32 == 0);
3951 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3952 LLVMValueRef src_vector =
3953 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3954 ret = LLVMGetUndef(vec_type);
3955 for (unsigned i = 0; i < bits / 32; i++) {
3956 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3957 LLVMConstInt(ctx->i32, i,
3958 0), "");
3959 LLVMValueRef ret_comp =
3960 _ac_build_permlane16(ctx, src, sel,
3961 exchange_rows,
3962 bound_ctrl);
3963 ret = LLVMBuildInsertElement(ctx->builder, ret,
3964 ret_comp,
3965 LLVMConstInt(ctx->i32, i,
3966 0), "");
3967 }
3968 }
3969 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3970 }
3971
3972 static inline unsigned
3973 ds_pattern_bitmode(unsigned and_mask, unsigned or_mask, unsigned xor_mask)
3974 {
3975 assert(and_mask < 32 && or_mask < 32 && xor_mask < 32);
3976 return and_mask | (or_mask << 5) | (xor_mask << 10);
3977 }
3978
3979 static LLVMValueRef
3980 _ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
3981 {
3982 return ac_build_intrinsic(ctx, "llvm.amdgcn.ds.swizzle",
3983 LLVMTypeOf(src), (LLVMValueRef []) {
3984 src, LLVMConstInt(ctx->i32, mask, 0) },
3985 2, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3986 }
3987
3988 LLVMValueRef
3989 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
3990 {
3991 LLVMTypeRef src_type = LLVMTypeOf(src);
3992 src = ac_to_integer(ctx, src);
3993 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3994 LLVMValueRef ret;
3995 if (bits == 32) {
3996 ret = _ac_build_ds_swizzle(ctx, src, mask);
3997 } else {
3998 assert(bits % 32 == 0);
3999 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
4000 LLVMValueRef src_vector =
4001 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
4002 ret = LLVMGetUndef(vec_type);
4003 for (unsigned i = 0; i < bits / 32; i++) {
4004 src = LLVMBuildExtractElement(ctx->builder, src_vector,
4005 LLVMConstInt(ctx->i32, i,
4006 0), "");
4007 LLVMValueRef ret_comp = _ac_build_ds_swizzle(ctx, src,
4008 mask);
4009 ret = LLVMBuildInsertElement(ctx->builder, ret,
4010 ret_comp,
4011 LLVMConstInt(ctx->i32, i,
4012 0), "");
4013 }
4014 }
4015 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
4016 }
4017
4018 static LLVMValueRef
4019 ac_build_wwm(struct ac_llvm_context *ctx, LLVMValueRef src)
4020 {
4021 char name[32], type[8];
4022 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
4023 snprintf(name, sizeof(name), "llvm.amdgcn.wwm.%s", type);
4024 return ac_build_intrinsic(ctx, name, LLVMTypeOf(src),
4025 (LLVMValueRef []) { src }, 1,
4026 AC_FUNC_ATTR_READNONE);
4027 }
4028
4029 static LLVMValueRef
4030 ac_build_set_inactive(struct ac_llvm_context *ctx, LLVMValueRef src,
4031 LLVMValueRef inactive)
4032 {
4033 char name[33], type[8];
4034 LLVMTypeRef src_type = LLVMTypeOf(src);
4035 src = ac_to_integer(ctx, src);
4036 inactive = ac_to_integer(ctx, inactive);
4037 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
4038 snprintf(name, sizeof(name), "llvm.amdgcn.set.inactive.%s", type);
4039 LLVMValueRef ret =
4040 ac_build_intrinsic(ctx, name,
4041 LLVMTypeOf(src), (LLVMValueRef []) {
4042 src, inactive }, 2,
4043 AC_FUNC_ATTR_READNONE |
4044 AC_FUNC_ATTR_CONVERGENT);
4045 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
4046 }
4047
4048 static LLVMValueRef
4049 get_reduction_identity(struct ac_llvm_context *ctx, nir_op op, unsigned type_size)
4050 {
4051 if (type_size == 4) {
4052 switch (op) {
4053 case nir_op_iadd: return ctx->i32_0;
4054 case nir_op_fadd: return ctx->f32_0;
4055 case nir_op_imul: return ctx->i32_1;
4056 case nir_op_fmul: return ctx->f32_1;
4057 case nir_op_imin: return LLVMConstInt(ctx->i32, INT32_MAX, 0);
4058 case nir_op_umin: return LLVMConstInt(ctx->i32, UINT32_MAX, 0);
4059 case nir_op_fmin: return LLVMConstReal(ctx->f32, INFINITY);
4060 case nir_op_imax: return LLVMConstInt(ctx->i32, INT32_MIN, 0);
4061 case nir_op_umax: return ctx->i32_0;
4062 case nir_op_fmax: return LLVMConstReal(ctx->f32, -INFINITY);
4063 case nir_op_iand: return LLVMConstInt(ctx->i32, -1, 0);
4064 case nir_op_ior: return ctx->i32_0;
4065 case nir_op_ixor: return ctx->i32_0;
4066 default:
4067 unreachable("bad reduction intrinsic");
4068 }
4069 } else { /* type_size == 64bit */
4070 switch (op) {
4071 case nir_op_iadd: return ctx->i64_0;
4072 case nir_op_fadd: return ctx->f64_0;
4073 case nir_op_imul: return ctx->i64_1;
4074 case nir_op_fmul: return ctx->f64_1;
4075 case nir_op_imin: return LLVMConstInt(ctx->i64, INT64_MAX, 0);
4076 case nir_op_umin: return LLVMConstInt(ctx->i64, UINT64_MAX, 0);
4077 case nir_op_fmin: return LLVMConstReal(ctx->f64, INFINITY);
4078 case nir_op_imax: return LLVMConstInt(ctx->i64, INT64_MIN, 0);
4079 case nir_op_umax: return ctx->i64_0;
4080 case nir_op_fmax: return LLVMConstReal(ctx->f64, -INFINITY);
4081 case nir_op_iand: return LLVMConstInt(ctx->i64, -1, 0);
4082 case nir_op_ior: return ctx->i64_0;
4083 case nir_op_ixor: return ctx->i64_0;
4084 default:
4085 unreachable("bad reduction intrinsic");
4086 }
4087 }
4088 }
4089
4090 static LLVMValueRef
4091 ac_build_alu_op(struct ac_llvm_context *ctx, LLVMValueRef lhs, LLVMValueRef rhs, nir_op op)
4092 {
4093 bool _64bit = ac_get_type_size(LLVMTypeOf(lhs)) == 8;
4094 switch (op) {
4095 case nir_op_iadd: return LLVMBuildAdd(ctx->builder, lhs, rhs, "");
4096 case nir_op_fadd: return LLVMBuildFAdd(ctx->builder, lhs, rhs, "");
4097 case nir_op_imul: return LLVMBuildMul(ctx->builder, lhs, rhs, "");
4098 case nir_op_fmul: return LLVMBuildFMul(ctx->builder, lhs, rhs, "");
4099 case nir_op_imin: return LLVMBuildSelect(ctx->builder,
4100 LLVMBuildICmp(ctx->builder, LLVMIntSLT, lhs, rhs, ""),
4101 lhs, rhs, "");
4102 case nir_op_umin: return LLVMBuildSelect(ctx->builder,
4103 LLVMBuildICmp(ctx->builder, LLVMIntULT, lhs, rhs, ""),
4104 lhs, rhs, "");
4105 case nir_op_fmin: return ac_build_intrinsic(ctx,
4106 _64bit ? "llvm.minnum.f64" : "llvm.minnum.f32",
4107 _64bit ? ctx->f64 : ctx->f32,
4108 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
4109 case nir_op_imax: return LLVMBuildSelect(ctx->builder,
4110 LLVMBuildICmp(ctx->builder, LLVMIntSGT, lhs, rhs, ""),
4111 lhs, rhs, "");
4112 case nir_op_umax: return LLVMBuildSelect(ctx->builder,
4113 LLVMBuildICmp(ctx->builder, LLVMIntUGT, lhs, rhs, ""),
4114 lhs, rhs, "");
4115 case nir_op_fmax: return ac_build_intrinsic(ctx,
4116 _64bit ? "llvm.maxnum.f64" : "llvm.maxnum.f32",
4117 _64bit ? ctx->f64 : ctx->f32,
4118 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
4119 case nir_op_iand: return LLVMBuildAnd(ctx->builder, lhs, rhs, "");
4120 case nir_op_ior: return LLVMBuildOr(ctx->builder, lhs, rhs, "");
4121 case nir_op_ixor: return LLVMBuildXor(ctx->builder, lhs, rhs, "");
4122 default:
4123 unreachable("bad reduction intrinsic");
4124 }
4125 }
4126
4127 /**
4128 * \param maxprefix specifies that the result only needs to be correct for a
4129 * prefix of this many threads
4130 *
4131 * TODO: add inclusive and excluse scan functions for GFX6.
4132 */
4133 static LLVMValueRef
4134 ac_build_scan(struct ac_llvm_context *ctx, nir_op op, LLVMValueRef src, LLVMValueRef identity,
4135 unsigned maxprefix, bool inclusive)
4136 {
4137 LLVMValueRef result, tmp;
4138
4139 if (ctx->chip_class >= GFX10) {
4140 result = inclusive ? src : identity;
4141 } else {
4142 if (inclusive)
4143 result = src;
4144 else
4145 result = ac_build_dpp(ctx, identity, src, dpp_wf_sr1, 0xf, 0xf, false);
4146 }
4147 if (maxprefix <= 1)
4148 return result;
4149 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
4150 result = ac_build_alu_op(ctx, result, tmp, op);
4151 if (maxprefix <= 2)
4152 return result;
4153 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(2), 0xf, 0xf, false);
4154 result = ac_build_alu_op(ctx, result, tmp, op);
4155 if (maxprefix <= 3)
4156 return result;
4157 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(3), 0xf, 0xf, false);
4158 result = ac_build_alu_op(ctx, result, tmp, op);
4159 if (maxprefix <= 4)
4160 return result;
4161 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(4), 0xf, 0xe, false);
4162 result = ac_build_alu_op(ctx, result, tmp, op);
4163 if (maxprefix <= 8)
4164 return result;
4165 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(8), 0xf, 0xc, false);
4166 result = ac_build_alu_op(ctx, result, tmp, op);
4167 if (maxprefix <= 16)
4168 return result;
4169
4170 if (ctx->chip_class >= GFX10) {
4171 /* dpp_row_bcast{15,31} are not supported on gfx10. */
4172 LLVMBuilderRef builder = ctx->builder;
4173 LLVMValueRef tid = ac_get_thread_id(ctx);
4174 LLVMValueRef cc;
4175 /* TODO-GFX10: Can we get better code-gen by putting this into
4176 * a branch so that LLVM generates EXEC mask manipulations? */
4177 if (inclusive)
4178 tmp = result;
4179 else
4180 tmp = ac_build_alu_op(ctx, result, src, op);
4181 tmp = ac_build_permlane16(ctx, tmp, ~(uint64_t)0, true, false);
4182 tmp = ac_build_alu_op(ctx, result, tmp, op);
4183 cc = LLVMBuildAnd(builder, tid, LLVMConstInt(ctx->i32, 16, false), "");
4184 cc = LLVMBuildICmp(builder, LLVMIntNE, cc, ctx->i32_0, "");
4185 result = LLVMBuildSelect(builder, cc, tmp, result, "");
4186 if (maxprefix <= 32)
4187 return result;
4188
4189 if (inclusive)
4190 tmp = result;
4191 else
4192 tmp = ac_build_alu_op(ctx, result, src, op);
4193 tmp = ac_build_readlane(ctx, tmp, LLVMConstInt(ctx->i32, 31, false));
4194 tmp = ac_build_alu_op(ctx, result, tmp, op);
4195 cc = LLVMBuildICmp(builder, LLVMIntUGE, tid,
4196 LLVMConstInt(ctx->i32, 32, false), "");
4197 result = LLVMBuildSelect(builder, cc, tmp, result, "");
4198 return result;
4199 }
4200
4201 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
4202 result = ac_build_alu_op(ctx, result, tmp, op);
4203 if (maxprefix <= 32)
4204 return result;
4205 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
4206 result = ac_build_alu_op(ctx, result, tmp, op);
4207 return result;
4208 }
4209
4210 LLVMValueRef
4211 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
4212 {
4213 LLVMValueRef result;
4214
4215 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
4216 LLVMBuilderRef builder = ctx->builder;
4217 src = LLVMBuildZExt(builder, src, ctx->i32, "");
4218 result = ac_build_ballot(ctx, src);
4219 result = ac_build_mbcnt(ctx, result);
4220 result = LLVMBuildAdd(builder, result, src, "");
4221 return result;
4222 }
4223
4224 ac_build_optimization_barrier(ctx, &src);
4225
4226 LLVMValueRef identity =
4227 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
4228 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
4229 LLVMTypeOf(identity), "");
4230 result = ac_build_scan(ctx, op, result, identity, 64, true);
4231
4232 return ac_build_wwm(ctx, result);
4233 }
4234
4235 LLVMValueRef
4236 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
4237 {
4238 LLVMValueRef result;
4239
4240 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
4241 LLVMBuilderRef builder = ctx->builder;
4242 src = LLVMBuildZExt(builder, src, ctx->i32, "");
4243 result = ac_build_ballot(ctx, src);
4244 result = ac_build_mbcnt(ctx, result);
4245 return result;
4246 }
4247
4248 ac_build_optimization_barrier(ctx, &src);
4249
4250 LLVMValueRef identity =
4251 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
4252 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
4253 LLVMTypeOf(identity), "");
4254 result = ac_build_scan(ctx, op, result, identity, 64, false);
4255
4256 return ac_build_wwm(ctx, result);
4257 }
4258
4259 LLVMValueRef
4260 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size)
4261 {
4262 if (cluster_size == 1) return src;
4263 ac_build_optimization_barrier(ctx, &src);
4264 LLVMValueRef result, swap;
4265 LLVMValueRef identity = get_reduction_identity(ctx, op,
4266 ac_get_type_size(LLVMTypeOf(src)));
4267 result = LLVMBuildBitCast(ctx->builder,
4268 ac_build_set_inactive(ctx, src, identity),
4269 LLVMTypeOf(identity), "");
4270 swap = ac_build_quad_swizzle(ctx, result, 1, 0, 3, 2);
4271 result = ac_build_alu_op(ctx, result, swap, op);
4272 if (cluster_size == 2) return ac_build_wwm(ctx, result);
4273
4274 swap = ac_build_quad_swizzle(ctx, result, 2, 3, 0, 1);
4275 result = ac_build_alu_op(ctx, result, swap, op);
4276 if (cluster_size == 4) return ac_build_wwm(ctx, result);
4277
4278 if (ctx->chip_class >= GFX8)
4279 swap = ac_build_dpp(ctx, identity, result, dpp_row_half_mirror, 0xf, 0xf, false);
4280 else
4281 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x04));
4282 result = ac_build_alu_op(ctx, result, swap, op);
4283 if (cluster_size == 8) return ac_build_wwm(ctx, result);
4284
4285 if (ctx->chip_class >= GFX8)
4286 swap = ac_build_dpp(ctx, identity, result, dpp_row_mirror, 0xf, 0xf, false);
4287 else
4288 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x08));
4289 result = ac_build_alu_op(ctx, result, swap, op);
4290 if (cluster_size == 16) return ac_build_wwm(ctx, result);
4291
4292 if (ctx->chip_class >= GFX10)
4293 swap = ac_build_permlane16(ctx, result, 0, true, false);
4294 else if (ctx->chip_class >= GFX8 && cluster_size != 32)
4295 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
4296 else
4297 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x10));
4298 result = ac_build_alu_op(ctx, result, swap, op);
4299 if (cluster_size == 32) return ac_build_wwm(ctx, result);
4300
4301 if (ctx->chip_class >= GFX8) {
4302 if (ctx->chip_class >= GFX10)
4303 swap = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, false));
4304 else
4305 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
4306 result = ac_build_alu_op(ctx, result, swap, op);
4307 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 63, 0));
4308 return ac_build_wwm(ctx, result);
4309 } else {
4310 swap = ac_build_readlane(ctx, result, ctx->i32_0);
4311 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 32, 0));
4312 result = ac_build_alu_op(ctx, result, swap, op);
4313 return ac_build_wwm(ctx, result);
4314 }
4315 }
4316
4317 /**
4318 * "Top half" of a scan that reduces per-wave values across an entire
4319 * workgroup.
4320 *
4321 * The source value must be present in the highest lane of the wave, and the
4322 * highest lane must be live.
4323 */
4324 void
4325 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4326 {
4327 if (ws->maxwaves <= 1)
4328 return;
4329
4330 const LLVMValueRef i32_63 = LLVMConstInt(ctx->i32, 63, false);
4331 LLVMBuilderRef builder = ctx->builder;
4332 LLVMValueRef tid = ac_get_thread_id(ctx);
4333 LLVMValueRef tmp;
4334
4335 tmp = LLVMBuildICmp(builder, LLVMIntEQ, tid, i32_63, "");
4336 ac_build_ifcc(ctx, tmp, 1000);
4337 LLVMBuildStore(builder, ws->src, LLVMBuildGEP(builder, ws->scratch, &ws->waveidx, 1, ""));
4338 ac_build_endif(ctx, 1000);
4339 }
4340
4341 /**
4342 * "Bottom half" of a scan that reduces per-wave values across an entire
4343 * workgroup.
4344 *
4345 * The caller must place a barrier between the top and bottom halves.
4346 */
4347 void
4348 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4349 {
4350 const LLVMTypeRef type = LLVMTypeOf(ws->src);
4351 const LLVMValueRef identity =
4352 get_reduction_identity(ctx, ws->op, ac_get_type_size(type));
4353
4354 if (ws->maxwaves <= 1) {
4355 ws->result_reduce = ws->src;
4356 ws->result_inclusive = ws->src;
4357 ws->result_exclusive = identity;
4358 return;
4359 }
4360 assert(ws->maxwaves <= 32);
4361
4362 LLVMBuilderRef builder = ctx->builder;
4363 LLVMValueRef tid = ac_get_thread_id(ctx);
4364 LLVMBasicBlockRef bbs[2];
4365 LLVMValueRef phivalues_scan[2];
4366 LLVMValueRef tmp, tmp2;
4367
4368 bbs[0] = LLVMGetInsertBlock(builder);
4369 phivalues_scan[0] = LLVMGetUndef(type);
4370
4371 if (ws->enable_reduce)
4372 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->numwaves, "");
4373 else if (ws->enable_inclusive)
4374 tmp = LLVMBuildICmp(builder, LLVMIntULE, tid, ws->waveidx, "");
4375 else
4376 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->waveidx, "");
4377 ac_build_ifcc(ctx, tmp, 1001);
4378 {
4379 tmp = LLVMBuildLoad(builder, LLVMBuildGEP(builder, ws->scratch, &tid, 1, ""), "");
4380
4381 ac_build_optimization_barrier(ctx, &tmp);
4382
4383 bbs[1] = LLVMGetInsertBlock(builder);
4384 phivalues_scan[1] = ac_build_scan(ctx, ws->op, tmp, identity, ws->maxwaves, true);
4385 }
4386 ac_build_endif(ctx, 1001);
4387
4388 const LLVMValueRef scan = ac_build_phi(ctx, type, 2, phivalues_scan, bbs);
4389
4390 if (ws->enable_reduce) {
4391 tmp = LLVMBuildSub(builder, ws->numwaves, ctx->i32_1, "");
4392 ws->result_reduce = ac_build_readlane(ctx, scan, tmp);
4393 }
4394 if (ws->enable_inclusive)
4395 ws->result_inclusive = ac_build_readlane(ctx, scan, ws->waveidx);
4396 if (ws->enable_exclusive) {
4397 tmp = LLVMBuildSub(builder, ws->waveidx, ctx->i32_1, "");
4398 tmp = ac_build_readlane(ctx, scan, tmp);
4399 tmp2 = LLVMBuildICmp(builder, LLVMIntEQ, ws->waveidx, ctx->i32_0, "");
4400 ws->result_exclusive = LLVMBuildSelect(builder, tmp2, identity, tmp, "");
4401 }
4402 }
4403
4404 /**
4405 * Inclusive scan of a per-wave value across an entire workgroup.
4406 *
4407 * This implies an s_barrier instruction.
4408 *
4409 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
4410 * of the workgroup are live. (This requirement cannot easily be relaxed in a
4411 * useful manner because of the barrier in the algorithm.)
4412 */
4413 void
4414 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4415 {
4416 ac_build_wg_wavescan_top(ctx, ws);
4417 ac_build_s_barrier(ctx);
4418 ac_build_wg_wavescan_bottom(ctx, ws);
4419 }
4420
4421 /**
4422 * "Top half" of a scan that reduces per-thread values across an entire
4423 * workgroup.
4424 *
4425 * All lanes must be active when this code runs.
4426 */
4427 void
4428 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4429 {
4430 if (ws->enable_exclusive) {
4431 ws->extra = ac_build_exclusive_scan(ctx, ws->src, ws->op);
4432 if (LLVMTypeOf(ws->src) == ctx->i1 && ws->op == nir_op_iadd)
4433 ws->src = LLVMBuildZExt(ctx->builder, ws->src, ctx->i32, "");
4434 ws->src = ac_build_alu_op(ctx, ws->extra, ws->src, ws->op);
4435 } else {
4436 ws->src = ac_build_inclusive_scan(ctx, ws->src, ws->op);
4437 }
4438
4439 bool enable_inclusive = ws->enable_inclusive;
4440 bool enable_exclusive = ws->enable_exclusive;
4441 ws->enable_inclusive = false;
4442 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
4443 ac_build_wg_wavescan_top(ctx, ws);
4444 ws->enable_inclusive = enable_inclusive;
4445 ws->enable_exclusive = enable_exclusive;
4446 }
4447
4448 /**
4449 * "Bottom half" of a scan that reduces per-thread values across an entire
4450 * workgroup.
4451 *
4452 * The caller must place a barrier between the top and bottom halves.
4453 */
4454 void
4455 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4456 {
4457 bool enable_inclusive = ws->enable_inclusive;
4458 bool enable_exclusive = ws->enable_exclusive;
4459 ws->enable_inclusive = false;
4460 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
4461 ac_build_wg_wavescan_bottom(ctx, ws);
4462 ws->enable_inclusive = enable_inclusive;
4463 ws->enable_exclusive = enable_exclusive;
4464
4465 /* ws->result_reduce is already the correct value */
4466 if (ws->enable_inclusive)
4467 ws->result_inclusive = ac_build_alu_op(ctx, ws->result_inclusive, ws->src, ws->op);
4468 if (ws->enable_exclusive)
4469 ws->result_exclusive = ac_build_alu_op(ctx, ws->result_exclusive, ws->extra, ws->op);
4470 }
4471
4472 /**
4473 * A scan that reduces per-thread values across an entire workgroup.
4474 *
4475 * The caller must ensure that all lanes are active when this code runs
4476 * (WWM is insufficient!), because there is an implied barrier.
4477 */
4478 void
4479 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4480 {
4481 ac_build_wg_scan_top(ctx, ws);
4482 ac_build_s_barrier(ctx);
4483 ac_build_wg_scan_bottom(ctx, ws);
4484 }
4485
4486 LLVMValueRef
4487 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
4488 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
4489 {
4490 unsigned mask = dpp_quad_perm(lane0, lane1, lane2, lane3);
4491 if (ctx->chip_class >= GFX8) {
4492 return ac_build_dpp(ctx, src, src, mask, 0xf, 0xf, false);
4493 } else {
4494 return ac_build_ds_swizzle(ctx, src, (1 << 15) | mask);
4495 }
4496 }
4497
4498 LLVMValueRef
4499 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index)
4500 {
4501 index = LLVMBuildMul(ctx->builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
4502 return ac_build_intrinsic(ctx,
4503 "llvm.amdgcn.ds.bpermute", ctx->i32,
4504 (LLVMValueRef []) {index, src}, 2,
4505 AC_FUNC_ATTR_READNONE |
4506 AC_FUNC_ATTR_CONVERGENT);
4507 }
4508
4509 LLVMValueRef
4510 ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
4511 unsigned bitsize)
4512 {
4513 LLVMTypeRef type;
4514 char *intr;
4515
4516 if (bitsize == 16) {
4517 intr = "llvm.amdgcn.frexp.exp.i16.f16";
4518 type = ctx->i16;
4519 } else if (bitsize == 32) {
4520 intr = "llvm.amdgcn.frexp.exp.i32.f32";
4521 type = ctx->i32;
4522 } else {
4523 intr = "llvm.amdgcn.frexp.exp.i32.f64";
4524 type = ctx->i32;
4525 }
4526
4527 LLVMValueRef params[] = {
4528 src0,
4529 };
4530 return ac_build_intrinsic(ctx, intr, type, params, 1,
4531 AC_FUNC_ATTR_READNONE);
4532 }
4533 LLVMValueRef
4534 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
4535 unsigned bitsize)
4536 {
4537 LLVMTypeRef type;
4538 char *intr;
4539
4540 if (bitsize == 16) {
4541 intr = "llvm.amdgcn.frexp.mant.f16";
4542 type = ctx->f16;
4543 } else if (bitsize == 32) {
4544 intr = "llvm.amdgcn.frexp.mant.f32";
4545 type = ctx->f32;
4546 } else {
4547 intr = "llvm.amdgcn.frexp.mant.f64";
4548 type = ctx->f64;
4549 }
4550
4551 LLVMValueRef params[] = {
4552 src0,
4553 };
4554 return ac_build_intrinsic(ctx, intr, type, params, 1,
4555 AC_FUNC_ATTR_READNONE);
4556 }
4557
4558 /*
4559 * this takes an I,J coordinate pair,
4560 * and works out the X and Y derivatives.
4561 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
4562 */
4563 LLVMValueRef
4564 ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij)
4565 {
4566 LLVMValueRef result[4], a;
4567 unsigned i;
4568
4569 for (i = 0; i < 2; i++) {
4570 a = LLVMBuildExtractElement(ctx->builder, interp_ij,
4571 LLVMConstInt(ctx->i32, i, false), "");
4572 result[i] = ac_build_ddxy(ctx, AC_TID_MASK_TOP_LEFT, 1, a);
4573 result[2+i] = ac_build_ddxy(ctx, AC_TID_MASK_TOP_LEFT, 2, a);
4574 }
4575 return ac_build_gather_values(ctx, result, 4);
4576 }
4577
4578 LLVMValueRef
4579 ac_build_load_helper_invocation(struct ac_llvm_context *ctx)
4580 {
4581 LLVMValueRef result = ac_build_intrinsic(ctx, "llvm.amdgcn.ps.live",
4582 ctx->i1, NULL, 0,
4583 AC_FUNC_ATTR_READNONE);
4584 result = LLVMBuildNot(ctx->builder, result, "");
4585 return LLVMBuildSExt(ctx->builder, result, ctx->i32, "");
4586 }
4587
4588 LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
4589 LLVMValueRef *args, unsigned num_args)
4590 {
4591 LLVMValueRef ret = LLVMBuildCall(ctx->builder, func, args, num_args, "");
4592 LLVMSetInstructionCallConv(ret, LLVMGetFunctionCallConv(func));
4593 return ret;
4594 }