projects / mesa.git / blob
? search:


a666ecd1ea18805f971ac4a2e301700fa0f1b09f
[mesa.git] / src / amd / common / ac_llvm_build.c
1 /*
2 * Copyright 2014 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the
6 * "Software"), to deal in the Software without restriction, including
7 * without limitation the rights to use, copy, modify, merge, publish,
8 * distribute, sub license, and/or sell copies of the Software, and to
9 * permit persons to whom the Software is furnished to do so, subject to
10 * the following conditions:
11 *
12 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
13 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
14 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
15 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
16 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
17 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
18 * USE OR OTHER DEALINGS IN THE SOFTWARE.
19 *
20 * The above copyright notice and this permission notice (including the
21 * next paragraph) shall be included in all copies or substantial portions
22 * of the Software.
23 *
24 */
25 /* based on pieces from si_pipe.c and radeon_llvm_emit.c */
26 #include "ac_llvm_build.h"
27
28 #include <llvm-c/Core.h>
29
30 #include "c11/threads.h"
31
32 #include <assert.h>
33 #include <stdio.h>
34
35 #include "ac_llvm_util.h"
36 #include "ac_exp_param.h"
37 #include "util/bitscan.h"
38 #include "util/macros.h"
39 #include "util/u_atomic.h"
40 #include "util/u_math.h"
41 #include "sid.h"
42
43 #include "shader_enums.h"
44
45 #define AC_LLVM_INITIAL_CF_DEPTH 4
46
47 /* Data for if/else/endif and bgnloop/endloop control flow structures.
48 */
49 struct ac_llvm_flow {
50 /* Loop exit or next part of if/else/endif. */
51 LLVMBasicBlockRef next_block;
52 LLVMBasicBlockRef loop_entry_block;
53 };
54
55 /* Initialize module-independent parts of the context.
56 *
57 * The caller is responsible for initializing ctx::module and ctx::builder.
58 */
59 void
60 ac_llvm_context_init(struct ac_llvm_context *ctx,
61 struct ac_llvm_compiler *compiler,
62 enum chip_class chip_class, enum radeon_family family,
63 enum ac_float_mode float_mode, unsigned wave_size)
64 {
65 LLVMValueRef args[1];
66
67 ctx->context = LLVMContextCreate();
68
69 ctx->chip_class = chip_class;
70 ctx->family = family;
71 ctx->wave_size = wave_size;
72 ctx->module = ac_create_module(wave_size == 32 ? compiler->tm_wave32
73 : compiler->tm,
74 ctx->context);
75 ctx->builder = ac_create_builder(ctx->context, float_mode);
76
77 ctx->voidt = LLVMVoidTypeInContext(ctx->context);
78 ctx->i1 = LLVMInt1TypeInContext(ctx->context);
79 ctx->i8 = LLVMInt8TypeInContext(ctx->context);
80 ctx->i16 = LLVMIntTypeInContext(ctx->context, 16);
81 ctx->i32 = LLVMIntTypeInContext(ctx->context, 32);
82 ctx->i64 = LLVMIntTypeInContext(ctx->context, 64);
83 ctx->intptr = ctx->i32;
84 ctx->f16 = LLVMHalfTypeInContext(ctx->context);
85 ctx->f32 = LLVMFloatTypeInContext(ctx->context);
86 ctx->f64 = LLVMDoubleTypeInContext(ctx->context);
87 ctx->v2i16 = LLVMVectorType(ctx->i16, 2);
88 ctx->v2i32 = LLVMVectorType(ctx->i32, 2);
89 ctx->v3i32 = LLVMVectorType(ctx->i32, 3);
90 ctx->v4i32 = LLVMVectorType(ctx->i32, 4);
91 ctx->v2f32 = LLVMVectorType(ctx->f32, 2);
92 ctx->v3f32 = LLVMVectorType(ctx->f32, 3);
93 ctx->v4f32 = LLVMVectorType(ctx->f32, 4);
94 ctx->v8i32 = LLVMVectorType(ctx->i32, 8);
95 ctx->iN_wavemask = LLVMIntTypeInContext(ctx->context, ctx->wave_size);
96
97 ctx->i8_0 = LLVMConstInt(ctx->i8, 0, false);
98 ctx->i8_1 = LLVMConstInt(ctx->i8, 1, false);
99 ctx->i16_0 = LLVMConstInt(ctx->i16, 0, false);
100 ctx->i16_1 = LLVMConstInt(ctx->i16, 1, false);
101 ctx->i32_0 = LLVMConstInt(ctx->i32, 0, false);
102 ctx->i32_1 = LLVMConstInt(ctx->i32, 1, false);
103 ctx->i64_0 = LLVMConstInt(ctx->i64, 0, false);
104 ctx->i64_1 = LLVMConstInt(ctx->i64, 1, false);
105 ctx->f16_0 = LLVMConstReal(ctx->f16, 0.0);
106 ctx->f16_1 = LLVMConstReal(ctx->f16, 1.0);
107 ctx->f32_0 = LLVMConstReal(ctx->f32, 0.0);
108 ctx->f32_1 = LLVMConstReal(ctx->f32, 1.0);
109 ctx->f64_0 = LLVMConstReal(ctx->f64, 0.0);
110 ctx->f64_1 = LLVMConstReal(ctx->f64, 1.0);
111
112 ctx->i1false = LLVMConstInt(ctx->i1, 0, false);
113 ctx->i1true = LLVMConstInt(ctx->i1, 1, false);
114
115 ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context,
116 "range", 5);
117
118 ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context,
119 "invariant.load", 14);
120
121 ctx->fpmath_md_kind = LLVMGetMDKindIDInContext(ctx->context, "fpmath", 6);
122
123 args[0] = LLVMConstReal(ctx->f32, 2.5);
124 ctx->fpmath_md_2p5_ulp = LLVMMDNodeInContext(ctx->context, args, 1);
125
126 ctx->uniform_md_kind = LLVMGetMDKindIDInContext(ctx->context,
127 "amdgpu.uniform", 14);
128
129 ctx->empty_md = LLVMMDNodeInContext(ctx->context, NULL, 0);
130 ctx->flow = calloc(1, sizeof(*ctx->flow));
131 }
132
133 void
134 ac_llvm_context_dispose(struct ac_llvm_context *ctx)
135 {
136 free(ctx->flow->stack);
137 free(ctx->flow);
138 ctx->flow = NULL;
139 }
140
141 int
142 ac_get_llvm_num_components(LLVMValueRef value)
143 {
144 LLVMTypeRef type = LLVMTypeOf(value);
145 unsigned num_components = LLVMGetTypeKind(type) == LLVMVectorTypeKind
146 ? LLVMGetVectorSize(type)
147 : 1;
148 return num_components;
149 }
150
151 LLVMValueRef
152 ac_llvm_extract_elem(struct ac_llvm_context *ac,
153 LLVMValueRef value,
154 int index)
155 {
156 if (LLVMGetTypeKind(LLVMTypeOf(value)) != LLVMVectorTypeKind) {
157 assert(index == 0);
158 return value;
159 }
160
161 return LLVMBuildExtractElement(ac->builder, value,
162 LLVMConstInt(ac->i32, index, false), "");
163 }
164
165 int
166 ac_get_elem_bits(struct ac_llvm_context *ctx, LLVMTypeRef type)
167 {
168 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind)
169 type = LLVMGetElementType(type);
170
171 if (LLVMGetTypeKind(type) == LLVMIntegerTypeKind)
172 return LLVMGetIntTypeWidth(type);
173
174 if (type == ctx->f16)
175 return 16;
176 if (type == ctx->f32)
177 return 32;
178 if (type == ctx->f64)
179 return 64;
180
181 unreachable("Unhandled type kind in get_elem_bits");
182 }
183
184 unsigned
185 ac_get_type_size(LLVMTypeRef type)
186 {
187 LLVMTypeKind kind = LLVMGetTypeKind(type);
188
189 switch (kind) {
190 case LLVMIntegerTypeKind:
191 return LLVMGetIntTypeWidth(type) / 8;
192 case LLVMHalfTypeKind:
193 return 2;
194 case LLVMFloatTypeKind:
195 return 4;
196 case LLVMDoubleTypeKind:
197 return 8;
198 case LLVMPointerTypeKind:
199 if (LLVMGetPointerAddressSpace(type) == AC_ADDR_SPACE_CONST_32BIT)
200 return 4;
201 return 8;
202 case LLVMVectorTypeKind:
203 return LLVMGetVectorSize(type) *
204 ac_get_type_size(LLVMGetElementType(type));
205 case LLVMArrayTypeKind:
206 return LLVMGetArrayLength(type) *
207 ac_get_type_size(LLVMGetElementType(type));
208 default:
209 assert(0);
210 return 0;
211 }
212 }
213
214 static LLVMTypeRef to_integer_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
215 {
216 if (t == ctx->i8)
217 return ctx->i8;
218 else if (t == ctx->f16 || t == ctx->i16)
219 return ctx->i16;
220 else if (t == ctx->f32 || t == ctx->i32)
221 return ctx->i32;
222 else if (t == ctx->f64 || t == ctx->i64)
223 return ctx->i64;
224 else
225 unreachable("Unhandled integer size");
226 }
227
228 LLVMTypeRef
229 ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
230 {
231 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
232 LLVMTypeRef elem_type = LLVMGetElementType(t);
233 return LLVMVectorType(to_integer_type_scalar(ctx, elem_type),
234 LLVMGetVectorSize(t));
235 }
236 if (LLVMGetTypeKind(t) == LLVMPointerTypeKind) {
237 switch (LLVMGetPointerAddressSpace(t)) {
238 case AC_ADDR_SPACE_GLOBAL:
239 return ctx->i64;
240 case AC_ADDR_SPACE_LDS:
241 return ctx->i32;
242 default:
243 unreachable("unhandled address space");
244 }
245 }
246 return to_integer_type_scalar(ctx, t);
247 }
248
249 LLVMValueRef
250 ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v)
251 {
252 LLVMTypeRef type = LLVMTypeOf(v);
253 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind) {
254 return LLVMBuildPtrToInt(ctx->builder, v, ac_to_integer_type(ctx, type), "");
255 }
256 return LLVMBuildBitCast(ctx->builder, v, ac_to_integer_type(ctx, type), "");
257 }
258
259 LLVMValueRef
260 ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v)
261 {
262 LLVMTypeRef type = LLVMTypeOf(v);
263 if (LLVMGetTypeKind(type) == LLVMPointerTypeKind)
264 return v;
265 return ac_to_integer(ctx, v);
266 }
267
268 static LLVMTypeRef to_float_type_scalar(struct ac_llvm_context *ctx, LLVMTypeRef t)
269 {
270 if (t == ctx->i8)
271 return ctx->i8;
272 else if (t == ctx->i16 || t == ctx->f16)
273 return ctx->f16;
274 else if (t == ctx->i32 || t == ctx->f32)
275 return ctx->f32;
276 else if (t == ctx->i64 || t == ctx->f64)
277 return ctx->f64;
278 else
279 unreachable("Unhandled float size");
280 }
281
282 LLVMTypeRef
283 ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t)
284 {
285 if (LLVMGetTypeKind(t) == LLVMVectorTypeKind) {
286 LLVMTypeRef elem_type = LLVMGetElementType(t);
287 return LLVMVectorType(to_float_type_scalar(ctx, elem_type),
288 LLVMGetVectorSize(t));
289 }
290 return to_float_type_scalar(ctx, t);
291 }
292
293 LLVMValueRef
294 ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v)
295 {
296 LLVMTypeRef type = LLVMTypeOf(v);
297 return LLVMBuildBitCast(ctx->builder, v, ac_to_float_type(ctx, type), "");
298 }
299
300
301 LLVMValueRef
302 ac_build_intrinsic(struct ac_llvm_context *ctx, const char *name,
303 LLVMTypeRef return_type, LLVMValueRef *params,
304 unsigned param_count, unsigned attrib_mask)
305 {
306 LLVMValueRef function, call;
307 bool set_callsite_attrs = !(attrib_mask & AC_FUNC_ATTR_LEGACY);
308
309 function = LLVMGetNamedFunction(ctx->module, name);
310 if (!function) {
311 LLVMTypeRef param_types[32], function_type;
312 unsigned i;
313
314 assert(param_count <= 32);
315
316 for (i = 0; i < param_count; ++i) {
317 assert(params[i]);
318 param_types[i] = LLVMTypeOf(params[i]);
319 }
320 function_type =
321 LLVMFunctionType(return_type, param_types, param_count, 0);
322 function = LLVMAddFunction(ctx->module, name, function_type);
323
324 LLVMSetFunctionCallConv(function, LLVMCCallConv);
325 LLVMSetLinkage(function, LLVMExternalLinkage);
326
327 if (!set_callsite_attrs)
328 ac_add_func_attributes(ctx->context, function, attrib_mask);
329 }
330
331 call = LLVMBuildCall(ctx->builder, function, params, param_count, "");
332 if (set_callsite_attrs)
333 ac_add_func_attributes(ctx->context, call, attrib_mask);
334 return call;
335 }
336
337 /**
338 * Given the i32 or vNi32 \p type, generate the textual name (e.g. for use with
339 * intrinsic names).
340 */
341 void ac_build_type_name_for_intr(LLVMTypeRef type, char *buf, unsigned bufsize)
342 {
343 LLVMTypeRef elem_type = type;
344
345 assert(bufsize >= 8);
346
347 if (LLVMGetTypeKind(type) == LLVMVectorTypeKind) {
348 int ret = snprintf(buf, bufsize, "v%u",
349 LLVMGetVectorSize(type));
350 if (ret < 0) {
351 char *type_name = LLVMPrintTypeToString(type);
352 fprintf(stderr, "Error building type name for: %s\n",
353 type_name);
354 LLVMDisposeMessage(type_name);
355 return;
356 }
357 elem_type = LLVMGetElementType(type);
358 buf += ret;
359 bufsize -= ret;
360 }
361 switch (LLVMGetTypeKind(elem_type)) {
362 default: break;
363 case LLVMIntegerTypeKind:
364 snprintf(buf, bufsize, "i%d", LLVMGetIntTypeWidth(elem_type));
365 break;
366 case LLVMHalfTypeKind:
367 snprintf(buf, bufsize, "f16");
368 break;
369 case LLVMFloatTypeKind:
370 snprintf(buf, bufsize, "f32");
371 break;
372 case LLVMDoubleTypeKind:
373 snprintf(buf, bufsize, "f64");
374 break;
375 }
376 }
377
378 /**
379 * Helper function that builds an LLVM IR PHI node and immediately adds
380 * incoming edges.
381 */
382 LLVMValueRef
383 ac_build_phi(struct ac_llvm_context *ctx, LLVMTypeRef type,
384 unsigned count_incoming, LLVMValueRef *values,
385 LLVMBasicBlockRef *blocks)
386 {
387 LLVMValueRef phi = LLVMBuildPhi(ctx->builder, type, "");
388 LLVMAddIncoming(phi, values, blocks, count_incoming);
389 return phi;
390 }
391
392 void ac_build_s_barrier(struct ac_llvm_context *ctx)
393 {
394 ac_build_intrinsic(ctx, "llvm.amdgcn.s.barrier", ctx->voidt, NULL,
395 0, AC_FUNC_ATTR_CONVERGENT);
396 }
397
398 /* Prevent optimizations (at least of memory accesses) across the current
399 * point in the program by emitting empty inline assembly that is marked as
400 * having side effects.
401 *
402 * Optionally, a value can be passed through the inline assembly to prevent
403 * LLVM from hoisting calls to ReadNone functions.
404 */
405 void
406 ac_build_optimization_barrier(struct ac_llvm_context *ctx,
407 LLVMValueRef *pvgpr)
408 {
409 static int counter = 0;
410
411 LLVMBuilderRef builder = ctx->builder;
412 char code[16];
413
414 snprintf(code, sizeof(code), "; %d", p_atomic_inc_return(&counter));
415
416 if (!pvgpr) {
417 LLVMTypeRef ftype = LLVMFunctionType(ctx->voidt, NULL, 0, false);
418 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "", true, false);
419 LLVMBuildCall(builder, inlineasm, NULL, 0, "");
420 } else {
421 LLVMTypeRef ftype = LLVMFunctionType(ctx->i32, &ctx->i32, 1, false);
422 LLVMValueRef inlineasm = LLVMConstInlineAsm(ftype, code, "=v,0", true, false);
423 LLVMValueRef vgpr = *pvgpr;
424 LLVMTypeRef vgpr_type = LLVMTypeOf(vgpr);
425 unsigned vgpr_size = ac_get_type_size(vgpr_type);
426 LLVMValueRef vgpr0;
427
428 assert(vgpr_size % 4 == 0);
429
430 vgpr = LLVMBuildBitCast(builder, vgpr, LLVMVectorType(ctx->i32, vgpr_size / 4), "");
431 vgpr0 = LLVMBuildExtractElement(builder, vgpr, ctx->i32_0, "");
432 vgpr0 = LLVMBuildCall(builder, inlineasm, &vgpr0, 1, "");
433 vgpr = LLVMBuildInsertElement(builder, vgpr, vgpr0, ctx->i32_0, "");
434 vgpr = LLVMBuildBitCast(builder, vgpr, vgpr_type, "");
435
436 *pvgpr = vgpr;
437 }
438 }
439
440 LLVMValueRef
441 ac_build_shader_clock(struct ac_llvm_context *ctx)
442 {
443 const char *intr = HAVE_LLVM >= 0x0900 && ctx->chip_class >= GFX8 ?
444 "llvm.amdgcn.s.memrealtime" : "llvm.readcyclecounter";
445 LLVMValueRef tmp = ac_build_intrinsic(ctx, intr, ctx->i64, NULL, 0, 0);
446 return LLVMBuildBitCast(ctx->builder, tmp, ctx->v2i32, "");
447 }
448
449 LLVMValueRef
450 ac_build_ballot(struct ac_llvm_context *ctx,
451 LLVMValueRef value)
452 {
453 const char *name;
454
455 if (HAVE_LLVM >= 0x900) {
456 if (ctx->wave_size == 64)
457 name = "llvm.amdgcn.icmp.i64.i32";
458 else
459 name = "llvm.amdgcn.icmp.i32.i32";
460 } else {
461 name = "llvm.amdgcn.icmp.i32";
462 }
463 LLVMValueRef args[3] = {
464 value,
465 ctx->i32_0,
466 LLVMConstInt(ctx->i32, LLVMIntNE, 0)
467 };
468
469 /* We currently have no other way to prevent LLVM from lifting the icmp
470 * calls to a dominating basic block.
471 */
472 ac_build_optimization_barrier(ctx, &args[0]);
473
474 args[0] = ac_to_integer(ctx, args[0]);
475
476 return ac_build_intrinsic(ctx, name, ctx->iN_wavemask, args, 3,
477 AC_FUNC_ATTR_NOUNWIND |
478 AC_FUNC_ATTR_READNONE |
479 AC_FUNC_ATTR_CONVERGENT);
480 }
481
482 LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
483 LLVMValueRef value)
484 {
485 const char *name = HAVE_LLVM >= 0x900 ? "llvm.amdgcn.icmp.i64.i1" : "llvm.amdgcn.icmp.i1";
486 LLVMValueRef args[3] = {
487 value,
488 ctx->i1false,
489 LLVMConstInt(ctx->i32, LLVMIntNE, 0),
490 };
491
492 assert(HAVE_LLVM >= 0x0800);
493 return ac_build_intrinsic(ctx, name, ctx->i64, args, 3,
494 AC_FUNC_ATTR_NOUNWIND |
495 AC_FUNC_ATTR_READNONE |
496 AC_FUNC_ATTR_CONVERGENT);
497 }
498
499 LLVMValueRef
500 ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value)
501 {
502 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
503 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
504 return LLVMBuildICmp(ctx->builder, LLVMIntEQ, vote_set, active_set, "");
505 }
506
507 LLVMValueRef
508 ac_build_vote_any(struct ac_llvm_context *ctx, LLVMValueRef value)
509 {
510 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
511 return LLVMBuildICmp(ctx->builder, LLVMIntNE, vote_set,
512 LLVMConstInt(ctx->iN_wavemask, 0, 0), "");
513 }
514
515 LLVMValueRef
516 ac_build_vote_eq(struct ac_llvm_context *ctx, LLVMValueRef value)
517 {
518 LLVMValueRef active_set = ac_build_ballot(ctx, ctx->i32_1);
519 LLVMValueRef vote_set = ac_build_ballot(ctx, value);
520
521 LLVMValueRef all = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
522 vote_set, active_set, "");
523 LLVMValueRef none = LLVMBuildICmp(ctx->builder, LLVMIntEQ,
524 vote_set,
525 LLVMConstInt(ctx->iN_wavemask, 0, 0), "");
526 return LLVMBuildOr(ctx->builder, all, none, "");
527 }
528
529 LLVMValueRef
530 ac_build_varying_gather_values(struct ac_llvm_context *ctx, LLVMValueRef *values,
531 unsigned value_count, unsigned component)
532 {
533 LLVMValueRef vec = NULL;
534
535 if (value_count == 1) {
536 return values[component];
537 } else if (!value_count)
538 unreachable("value_count is 0");
539
540 for (unsigned i = component; i < value_count + component; i++) {
541 LLVMValueRef value = values[i];
542
543 if (i == component)
544 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
545 LLVMValueRef index = LLVMConstInt(ctx->i32, i - component, false);
546 vec = LLVMBuildInsertElement(ctx->builder, vec, value, index, "");
547 }
548 return vec;
549 }
550
551 LLVMValueRef
552 ac_build_gather_values_extended(struct ac_llvm_context *ctx,
553 LLVMValueRef *values,
554 unsigned value_count,
555 unsigned value_stride,
556 bool load,
557 bool always_vector)
558 {
559 LLVMBuilderRef builder = ctx->builder;
560 LLVMValueRef vec = NULL;
561 unsigned i;
562
563 if (value_count == 1 && !always_vector) {
564 if (load)
565 return LLVMBuildLoad(builder, values[0], "");
566 return values[0];
567 } else if (!value_count)
568 unreachable("value_count is 0");
569
570 for (i = 0; i < value_count; i++) {
571 LLVMValueRef value = values[i * value_stride];
572 if (load)
573 value = LLVMBuildLoad(builder, value, "");
574
575 if (!i)
576 vec = LLVMGetUndef( LLVMVectorType(LLVMTypeOf(value), value_count));
577 LLVMValueRef index = LLVMConstInt(ctx->i32, i, false);
578 vec = LLVMBuildInsertElement(builder, vec, value, index, "");
579 }
580 return vec;
581 }
582
583 LLVMValueRef
584 ac_build_gather_values(struct ac_llvm_context *ctx,
585 LLVMValueRef *values,
586 unsigned value_count)
587 {
588 return ac_build_gather_values_extended(ctx, values, value_count, 1, false, false);
589 }
590
591 /* Expand a scalar or vector to <dst_channels x type> by filling the remaining
592 * channels with undef. Extract at most src_channels components from the input.
593 */
594 static LLVMValueRef
595 ac_build_expand(struct ac_llvm_context *ctx,
596 LLVMValueRef value,
597 unsigned src_channels,
598 unsigned dst_channels)
599 {
600 LLVMTypeRef elemtype;
601 LLVMValueRef chan[dst_channels];
602
603 if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMVectorTypeKind) {
604 unsigned vec_size = LLVMGetVectorSize(LLVMTypeOf(value));
605
606 if (src_channels == dst_channels && vec_size == dst_channels)
607 return value;
608
609 src_channels = MIN2(src_channels, vec_size);
610
611 for (unsigned i = 0; i < src_channels; i++)
612 chan[i] = ac_llvm_extract_elem(ctx, value, i);
613
614 elemtype = LLVMGetElementType(LLVMTypeOf(value));
615 } else {
616 if (src_channels) {
617 assert(src_channels == 1);
618 chan[0] = value;
619 }
620 elemtype = LLVMTypeOf(value);
621 }
622
623 for (unsigned i = src_channels; i < dst_channels; i++)
624 chan[i] = LLVMGetUndef(elemtype);
625
626 return ac_build_gather_values(ctx, chan, dst_channels);
627 }
628
629 /* Expand a scalar or vector to <4 x type> by filling the remaining channels
630 * with undef. Extract at most num_channels components from the input.
631 */
632 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
633 LLVMValueRef value,
634 unsigned num_channels)
635 {
636 return ac_build_expand(ctx, value, num_channels, 4);
637 }
638
639 LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value)
640 {
641 unsigned type_size = ac_get_type_size(LLVMTypeOf(value));
642 const char *name;
643
644 if (type_size == 2)
645 name = "llvm.rint.f16";
646 else if (type_size == 4)
647 name = "llvm.rint.f32";
648 else
649 name = "llvm.rint.f64";
650
651 return ac_build_intrinsic(ctx, name, LLVMTypeOf(value), &value, 1,
652 AC_FUNC_ATTR_READNONE);
653 }
654
655 LLVMValueRef
656 ac_build_fdiv(struct ac_llvm_context *ctx,
657 LLVMValueRef num,
658 LLVMValueRef den)
659 {
660 /* If we do (num / den), LLVM >= 7.0 does:
661 * return num * v_rcp_f32(den * (fabs(den) > 0x1.0p+96f ? 0x1.0p-32f : 1.0f));
662 *
663 * If we do (num * (1 / den)), LLVM does:
664 * return num * v_rcp_f32(den);
665 */
666 LLVMValueRef one = LLVMConstReal(LLVMTypeOf(num), 1.0);
667 LLVMValueRef rcp = LLVMBuildFDiv(ctx->builder, one, den, "");
668 LLVMValueRef ret = LLVMBuildFMul(ctx->builder, num, rcp, "");
669
670 /* Use v_rcp_f32 instead of precise division. */
671 if (!LLVMIsConstant(ret))
672 LLVMSetMetadata(ret, ctx->fpmath_md_kind, ctx->fpmath_md_2p5_ulp);
673 return ret;
674 }
675
676 /* See fast_idiv_by_const.h. */
677 /* Set: increment = util_fast_udiv_info::increment ? multiplier : 0; */
678 LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
679 LLVMValueRef num,
680 LLVMValueRef multiplier,
681 LLVMValueRef pre_shift,
682 LLVMValueRef post_shift,
683 LLVMValueRef increment)
684 {
685 LLVMBuilderRef builder = ctx->builder;
686
687 num = LLVMBuildLShr(builder, num, pre_shift, "");
688 num = LLVMBuildMul(builder,
689 LLVMBuildZExt(builder, num, ctx->i64, ""),
690 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
691 num = LLVMBuildAdd(builder, num,
692 LLVMBuildZExt(builder, increment, ctx->i64, ""), "");
693 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
694 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
695 return LLVMBuildLShr(builder, num, post_shift, "");
696 }
697
698 /* See fast_idiv_by_const.h. */
699 /* If num != UINT_MAX, this more efficient version can be used. */
700 /* Set: increment = util_fast_udiv_info::increment; */
701 LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
702 LLVMValueRef num,
703 LLVMValueRef multiplier,
704 LLVMValueRef pre_shift,
705 LLVMValueRef post_shift,
706 LLVMValueRef increment)
707 {
708 LLVMBuilderRef builder = ctx->builder;
709
710 num = LLVMBuildLShr(builder, num, pre_shift, "");
711 num = LLVMBuildNUWAdd(builder, num, increment, "");
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 /* See fast_idiv_by_const.h. */
721 /* Both operands must fit in 31 bits and the divisor must not be 1. */
722 LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
723 LLVMValueRef num,
724 LLVMValueRef multiplier,
725 LLVMValueRef post_shift)
726 {
727 LLVMBuilderRef builder = ctx->builder;
728
729 num = LLVMBuildMul(builder,
730 LLVMBuildZExt(builder, num, ctx->i64, ""),
731 LLVMBuildZExt(builder, multiplier, ctx->i64, ""), "");
732 num = LLVMBuildLShr(builder, num, LLVMConstInt(ctx->i64, 32, 0), "");
733 num = LLVMBuildTrunc(builder, num, ctx->i32, "");
734 return LLVMBuildLShr(builder, num, post_shift, "");
735 }
736
737 /* Coordinates for cube map selection. sc, tc, and ma are as in Table 8.27
738 * of the OpenGL 4.5 (Compatibility Profile) specification, except ma is
739 * already multiplied by two. id is the cube face number.
740 */
741 struct cube_selection_coords {
742 LLVMValueRef stc[2];
743 LLVMValueRef ma;
744 LLVMValueRef id;
745 };
746
747 static void
748 build_cube_intrinsic(struct ac_llvm_context *ctx,
749 LLVMValueRef in[3],
750 struct cube_selection_coords *out)
751 {
752 LLVMTypeRef f32 = ctx->f32;
753
754 out->stc[1] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubetc",
755 f32, in, 3, AC_FUNC_ATTR_READNONE);
756 out->stc[0] = ac_build_intrinsic(ctx, "llvm.amdgcn.cubesc",
757 f32, in, 3, AC_FUNC_ATTR_READNONE);
758 out->ma = ac_build_intrinsic(ctx, "llvm.amdgcn.cubema",
759 f32, in, 3, AC_FUNC_ATTR_READNONE);
760 out->id = ac_build_intrinsic(ctx, "llvm.amdgcn.cubeid",
761 f32, in, 3, AC_FUNC_ATTR_READNONE);
762 }
763
764 /**
765 * Build a manual selection sequence for cube face sc/tc coordinates and
766 * major axis vector (multiplied by 2 for consistency) for the given
767 * vec3 \p coords, for the face implied by \p selcoords.
768 *
769 * For the major axis, we always adjust the sign to be in the direction of
770 * selcoords.ma; i.e., a positive out_ma means that coords is pointed towards
771 * the selcoords major axis.
772 */
773 static void build_cube_select(struct ac_llvm_context *ctx,
774 const struct cube_selection_coords *selcoords,
775 const LLVMValueRef *coords,
776 LLVMValueRef *out_st,
777 LLVMValueRef *out_ma)
778 {
779 LLVMBuilderRef builder = ctx->builder;
780 LLVMTypeRef f32 = LLVMTypeOf(coords[0]);
781 LLVMValueRef is_ma_positive;
782 LLVMValueRef sgn_ma;
783 LLVMValueRef is_ma_z, is_not_ma_z;
784 LLVMValueRef is_ma_y;
785 LLVMValueRef is_ma_x;
786 LLVMValueRef sgn;
787 LLVMValueRef tmp;
788
789 is_ma_positive = LLVMBuildFCmp(builder, LLVMRealUGE,
790 selcoords->ma, LLVMConstReal(f32, 0.0), "");
791 sgn_ma = LLVMBuildSelect(builder, is_ma_positive,
792 LLVMConstReal(f32, 1.0), LLVMConstReal(f32, -1.0), "");
793
794 is_ma_z = LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 4.0), "");
795 is_not_ma_z = LLVMBuildNot(builder, is_ma_z, "");
796 is_ma_y = LLVMBuildAnd(builder, is_not_ma_z,
797 LLVMBuildFCmp(builder, LLVMRealUGE, selcoords->id, LLVMConstReal(f32, 2.0), ""), "");
798 is_ma_x = LLVMBuildAnd(builder, is_not_ma_z, LLVMBuildNot(builder, is_ma_y, ""), "");
799
800 /* Select sc */
801 tmp = LLVMBuildSelect(builder, is_ma_x, coords[2], coords[0], "");
802 sgn = LLVMBuildSelect(builder, is_ma_y, LLVMConstReal(f32, 1.0),
803 LLVMBuildSelect(builder, is_ma_z, sgn_ma,
804 LLVMBuildFNeg(builder, sgn_ma, ""), ""), "");
805 out_st[0] = LLVMBuildFMul(builder, tmp, sgn, "");
806
807 /* Select tc */
808 tmp = LLVMBuildSelect(builder, is_ma_y, coords[2], coords[1], "");
809 sgn = LLVMBuildSelect(builder, is_ma_y, sgn_ma,
810 LLVMConstReal(f32, -1.0), "");
811 out_st[1] = LLVMBuildFMul(builder, tmp, sgn, "");
812
813 /* Select ma */
814 tmp = LLVMBuildSelect(builder, is_ma_z, coords[2],
815 LLVMBuildSelect(builder, is_ma_y, coords[1], coords[0], ""), "");
816 tmp = ac_build_intrinsic(ctx, "llvm.fabs.f32",
817 ctx->f32, &tmp, 1, AC_FUNC_ATTR_READNONE);
818 *out_ma = LLVMBuildFMul(builder, tmp, LLVMConstReal(f32, 2.0), "");
819 }
820
821 void
822 ac_prepare_cube_coords(struct ac_llvm_context *ctx,
823 bool is_deriv, bool is_array, bool is_lod,
824 LLVMValueRef *coords_arg,
825 LLVMValueRef *derivs_arg)
826 {
827
828 LLVMBuilderRef builder = ctx->builder;
829 struct cube_selection_coords selcoords;
830 LLVMValueRef coords[3];
831 LLVMValueRef invma;
832
833 if (is_array && !is_lod) {
834 LLVMValueRef tmp = ac_build_round(ctx, coords_arg[3]);
835
836 /* Section 8.9 (Texture Functions) of the GLSL 4.50 spec says:
837 *
838 * "For Array forms, the array layer used will be
839 *
840 * max(0, min(d−1, floor(layer+0.5)))
841 *
842 * where d is the depth of the texture array and layer
843 * comes from the component indicated in the tables below.
844 * Workaroudn for an issue where the layer is taken from a
845 * helper invocation which happens to fall on a different
846 * layer due to extrapolation."
847 *
848 * GFX8 and earlier attempt to implement this in hardware by
849 * clamping the value of coords[2] = (8 * layer) + face.
850 * Unfortunately, this means that the we end up with the wrong
851 * face when clamping occurs.
852 *
853 * Clamp the layer earlier to work around the issue.
854 */
855 if (ctx->chip_class <= GFX8) {
856 LLVMValueRef ge0;
857 ge0 = LLVMBuildFCmp(builder, LLVMRealOGE, tmp, ctx->f32_0, "");
858 tmp = LLVMBuildSelect(builder, ge0, tmp, ctx->f32_0, "");
859 }
860
861 coords_arg[3] = tmp;
862 }
863
864 build_cube_intrinsic(ctx, coords_arg, &selcoords);
865
866 invma = ac_build_intrinsic(ctx, "llvm.fabs.f32",
867 ctx->f32, &selcoords.ma, 1, AC_FUNC_ATTR_READNONE);
868 invma = ac_build_fdiv(ctx, LLVMConstReal(ctx->f32, 1.0), invma);
869
870 for (int i = 0; i < 2; ++i)
871 coords[i] = LLVMBuildFMul(builder, selcoords.stc[i], invma, "");
872
873 coords[2] = selcoords.id;
874
875 if (is_deriv && derivs_arg) {
876 LLVMValueRef derivs[4];
877 int axis;
878
879 /* Convert cube derivatives to 2D derivatives. */
880 for (axis = 0; axis < 2; axis++) {
881 LLVMValueRef deriv_st[2];
882 LLVMValueRef deriv_ma;
883
884 /* Transform the derivative alongside the texture
885 * coordinate. Mathematically, the correct formula is
886 * as follows. Assume we're projecting onto the +Z face
887 * and denote by dx/dh the derivative of the (original)
888 * X texture coordinate with respect to horizontal
889 * window coordinates. The projection onto the +Z face
890 * plane is:
891 *
892 * f(x,z) = x/z
893 *
894 * Then df/dh = df/dx * dx/dh + df/dz * dz/dh
895 * = 1/z * dx/dh - x/z * 1/z * dz/dh.
896 *
897 * This motivatives the implementation below.
898 *
899 * Whether this actually gives the expected results for
900 * apps that might feed in derivatives obtained via
901 * finite differences is anyone's guess. The OpenGL spec
902 * seems awfully quiet about how textureGrad for cube
903 * maps should be handled.
904 */
905 build_cube_select(ctx, &selcoords, &derivs_arg[axis * 3],
906 deriv_st, &deriv_ma);
907
908 deriv_ma = LLVMBuildFMul(builder, deriv_ma, invma, "");
909
910 for (int i = 0; i < 2; ++i)
911 derivs[axis * 2 + i] =
912 LLVMBuildFSub(builder,
913 LLVMBuildFMul(builder, deriv_st[i], invma, ""),
914 LLVMBuildFMul(builder, deriv_ma, coords[i], ""), "");
915 }
916
917 memcpy(derivs_arg, derivs, sizeof(derivs));
918 }
919
920 /* Shift the texture coordinate. This must be applied after the
921 * derivative calculation.
922 */
923 for (int i = 0; i < 2; ++i)
924 coords[i] = LLVMBuildFAdd(builder, coords[i], LLVMConstReal(ctx->f32, 1.5), "");
925
926 if (is_array) {
927 /* for cube arrays coord.z = coord.w(array_index) * 8 + face */
928 /* coords_arg.w component - array_index for cube arrays */
929 coords[2] = ac_build_fmad(ctx, coords_arg[3], LLVMConstReal(ctx->f32, 8.0), coords[2]);
930 }
931
932 memcpy(coords_arg, coords, sizeof(coords));
933 }
934
935
936 LLVMValueRef
937 ac_build_fs_interp(struct ac_llvm_context *ctx,
938 LLVMValueRef llvm_chan,
939 LLVMValueRef attr_number,
940 LLVMValueRef params,
941 LLVMValueRef i,
942 LLVMValueRef j)
943 {
944 LLVMValueRef args[5];
945 LLVMValueRef p1;
946
947 args[0] = i;
948 args[1] = llvm_chan;
949 args[2] = attr_number;
950 args[3] = params;
951
952 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1",
953 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
954
955 args[0] = p1;
956 args[1] = j;
957 args[2] = llvm_chan;
958 args[3] = attr_number;
959 args[4] = params;
960
961 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2",
962 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
963 }
964
965 LLVMValueRef
966 ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
967 LLVMValueRef llvm_chan,
968 LLVMValueRef attr_number,
969 LLVMValueRef params,
970 LLVMValueRef i,
971 LLVMValueRef j)
972 {
973 LLVMValueRef args[6];
974 LLVMValueRef p1;
975
976 args[0] = i;
977 args[1] = llvm_chan;
978 args[2] = attr_number;
979 args[3] = ctx->i1false;
980 args[4] = params;
981
982 p1 = ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p1.f16",
983 ctx->f32, args, 5, AC_FUNC_ATTR_READNONE);
984
985 args[0] = p1;
986 args[1] = j;
987 args[2] = llvm_chan;
988 args[3] = attr_number;
989 args[4] = ctx->i1false;
990 args[5] = params;
991
992 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.p2.f16",
993 ctx->f16, args, 6, AC_FUNC_ATTR_READNONE);
994 }
995
996 LLVMValueRef
997 ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
998 LLVMValueRef parameter,
999 LLVMValueRef llvm_chan,
1000 LLVMValueRef attr_number,
1001 LLVMValueRef params)
1002 {
1003 LLVMValueRef args[4];
1004
1005 args[0] = parameter;
1006 args[1] = llvm_chan;
1007 args[2] = attr_number;
1008 args[3] = params;
1009
1010 return ac_build_intrinsic(ctx, "llvm.amdgcn.interp.mov",
1011 ctx->f32, args, 4, AC_FUNC_ATTR_READNONE);
1012 }
1013
1014 LLVMValueRef
1015 ac_build_gep_ptr(struct ac_llvm_context *ctx,
1016 LLVMValueRef base_ptr,
1017 LLVMValueRef index)
1018 {
1019 return LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
1020 }
1021
1022 LLVMValueRef
1023 ac_build_gep0(struct ac_llvm_context *ctx,
1024 LLVMValueRef base_ptr,
1025 LLVMValueRef index)
1026 {
1027 LLVMValueRef indices[2] = {
1028 ctx->i32_0,
1029 index,
1030 };
1031 return LLVMBuildGEP(ctx->builder, base_ptr, indices, 2, "");
1032 }
1033
1034 LLVMValueRef ac_build_pointer_add(struct ac_llvm_context *ctx, LLVMValueRef ptr,
1035 LLVMValueRef index)
1036 {
1037 return LLVMBuildPointerCast(ctx->builder,
1038 LLVMBuildGEP(ctx->builder, ptr, &index, 1, ""),
1039 LLVMTypeOf(ptr), "");
1040 }
1041
1042 void
1043 ac_build_indexed_store(struct ac_llvm_context *ctx,
1044 LLVMValueRef base_ptr, LLVMValueRef index,
1045 LLVMValueRef value)
1046 {
1047 LLVMBuildStore(ctx->builder, value,
1048 ac_build_gep0(ctx, base_ptr, index));
1049 }
1050
1051 /**
1052 * Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
1053 * It's equivalent to doing a load from &base_ptr[index].
1054 *
1055 * \param base_ptr Where the array starts.
1056 * \param index The element index into the array.
1057 * \param uniform Whether the base_ptr and index can be assumed to be
1058 * dynamically uniform (i.e. load to an SGPR)
1059 * \param invariant Whether the load is invariant (no other opcodes affect it)
1060 * \param no_unsigned_wraparound
1061 * For all possible re-associations and re-distributions of an expression
1062 * "base_ptr + index * elemsize" into "addr + offset" (excluding GEPs
1063 * without inbounds in base_ptr), this parameter is true if "addr + offset"
1064 * does not result in an unsigned integer wraparound. This is used for
1065 * optimal code generation of 32-bit pointer arithmetic.
1066 *
1067 * For example, a 32-bit immediate offset that causes a 32-bit unsigned
1068 * integer wraparound can't be an imm offset in s_load_dword, because
1069 * the instruction performs "addr + offset" in 64 bits.
1070 *
1071 * Expected usage for bindless textures by chaining GEPs:
1072 * // possible unsigned wraparound, don't use InBounds:
1073 * ptr1 = LLVMBuildGEP(base_ptr, index);
1074 * image = load(ptr1); // becomes "s_load ptr1, 0"
1075 *
1076 * ptr2 = LLVMBuildInBoundsGEP(ptr1, 32 / elemsize);
1077 * sampler = load(ptr2); // becomes "s_load ptr1, 32" thanks to InBounds
1078 */
1079 static LLVMValueRef
1080 ac_build_load_custom(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1081 LLVMValueRef index, bool uniform, bool invariant,
1082 bool no_unsigned_wraparound)
1083 {
1084 LLVMValueRef pointer, result;
1085
1086 if (no_unsigned_wraparound &&
1087 LLVMGetPointerAddressSpace(LLVMTypeOf(base_ptr)) == AC_ADDR_SPACE_CONST_32BIT)
1088 pointer = LLVMBuildInBoundsGEP(ctx->builder, base_ptr, &index, 1, "");
1089 else
1090 pointer = LLVMBuildGEP(ctx->builder, base_ptr, &index, 1, "");
1091
1092 if (uniform)
1093 LLVMSetMetadata(pointer, ctx->uniform_md_kind, ctx->empty_md);
1094 result = LLVMBuildLoad(ctx->builder, pointer, "");
1095 if (invariant)
1096 LLVMSetMetadata(result, ctx->invariant_load_md_kind, ctx->empty_md);
1097 return result;
1098 }
1099
1100 LLVMValueRef ac_build_load(struct ac_llvm_context *ctx, LLVMValueRef base_ptr,
1101 LLVMValueRef index)
1102 {
1103 return ac_build_load_custom(ctx, base_ptr, index, false, false, false);
1104 }
1105
1106 LLVMValueRef ac_build_load_invariant(struct ac_llvm_context *ctx,
1107 LLVMValueRef base_ptr, LLVMValueRef index)
1108 {
1109 return ac_build_load_custom(ctx, base_ptr, index, false, true, false);
1110 }
1111
1112 /* This assumes that there is no unsigned integer wraparound during the address
1113 * computation, excluding all GEPs within base_ptr. */
1114 LLVMValueRef ac_build_load_to_sgpr(struct ac_llvm_context *ctx,
1115 LLVMValueRef base_ptr, LLVMValueRef index)
1116 {
1117 return ac_build_load_custom(ctx, base_ptr, index, true, true, true);
1118 }
1119
1120 /* See ac_build_load_custom() documentation. */
1121 LLVMValueRef ac_build_load_to_sgpr_uint_wraparound(struct ac_llvm_context *ctx,
1122 LLVMValueRef base_ptr, LLVMValueRef index)
1123 {
1124 return ac_build_load_custom(ctx, base_ptr, index, true, true, false);
1125 }
1126
1127 static unsigned get_load_cache_policy(struct ac_llvm_context *ctx,
1128 unsigned cache_policy)
1129 {
1130 return cache_policy |
1131 (ctx->chip_class >= GFX10 && cache_policy & ac_glc ? ac_dlc : 0);
1132 }
1133
1134 static void
1135 ac_build_llvm7_buffer_store_common(struct ac_llvm_context *ctx,
1136 LLVMValueRef rsrc,
1137 LLVMValueRef data,
1138 LLVMValueRef vindex,
1139 LLVMValueRef voffset,
1140 unsigned num_channels,
1141 unsigned cache_policy,
1142 bool use_format)
1143 {
1144 LLVMValueRef args[] = {
1145 data,
1146 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
1147 vindex ? vindex : ctx->i32_0,
1148 voffset,
1149 LLVMConstInt(ctx->i1, !!(cache_policy & ac_glc), 0),
1150 LLVMConstInt(ctx->i1, !!(cache_policy & ac_slc), 0)
1151 };
1152 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1153
1154 const char *type_names[] = {"f32", "v2f32", "v4f32"};
1155 char name[256];
1156
1157 if (use_format) {
1158 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.store.format.%s",
1159 type_names[func]);
1160 } else {
1161 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.store.%s",
1162 type_names[func]);
1163 }
1164
1165 ac_build_intrinsic(ctx, name, ctx->voidt, args, ARRAY_SIZE(args),
1166 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
1167 }
1168
1169 static void
1170 ac_build_llvm8_buffer_store_common(struct ac_llvm_context *ctx,
1171 LLVMValueRef rsrc,
1172 LLVMValueRef data,
1173 LLVMValueRef vindex,
1174 LLVMValueRef voffset,
1175 LLVMValueRef soffset,
1176 unsigned num_channels,
1177 LLVMTypeRef return_channel_type,
1178 unsigned cache_policy,
1179 bool use_format,
1180 bool structurized)
1181 {
1182 LLVMValueRef args[6];
1183 int idx = 0;
1184 args[idx++] = data;
1185 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1186 if (structurized)
1187 args[idx++] = vindex ? vindex : ctx->i32_0;
1188 args[idx++] = voffset ? voffset : ctx->i32_0;
1189 args[idx++] = soffset ? soffset : ctx->i32_0;
1190 args[idx++] = LLVMConstInt(ctx->i32, cache_policy, 0);
1191 unsigned func = !ac_has_vec3_support(ctx->chip_class, use_format) && num_channels == 3 ? 4 : num_channels;
1192 const char *indexing_kind = structurized ? "struct" : "raw";
1193 char name[256], type_name[8];
1194
1195 LLVMTypeRef type = func > 1 ? LLVMVectorType(return_channel_type, func) : return_channel_type;
1196 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1197
1198 if (use_format) {
1199 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.format.%s",
1200 indexing_kind, type_name);
1201 } else {
1202 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.store.%s",
1203 indexing_kind, type_name);
1204 }
1205
1206 ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
1207 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
1208 }
1209
1210 void
1211 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
1212 LLVMValueRef rsrc,
1213 LLVMValueRef data,
1214 LLVMValueRef vindex,
1215 LLVMValueRef voffset,
1216 unsigned num_channels,
1217 unsigned cache_policy)
1218 {
1219 if (HAVE_LLVM >= 0x800) {
1220 ac_build_llvm8_buffer_store_common(ctx, rsrc, data, vindex,
1221 voffset, NULL, num_channels,
1222 ctx->f32, cache_policy,
1223 true, true);
1224 } else {
1225 ac_build_llvm7_buffer_store_common(ctx, rsrc, data, vindex, voffset,
1226 num_channels, cache_policy,
1227 true);
1228 }
1229 }
1230
1231 /* TBUFFER_STORE_FORMAT_{X,XY,XYZ,XYZW} <- the suffix is selected by num_channels=1..4.
1232 * The type of vdata must be one of i32 (num_channels=1), v2i32 (num_channels=2),
1233 * or v4i32 (num_channels=3,4).
1234 */
1235 void
1236 ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
1237 LLVMValueRef rsrc,
1238 LLVMValueRef vdata,
1239 unsigned num_channels,
1240 LLVMValueRef voffset,
1241 LLVMValueRef soffset,
1242 unsigned inst_offset,
1243 unsigned cache_policy,
1244 bool swizzle_enable_hint)
1245 {
1246 /* Split 3 channel stores, because only LLVM 9+ support 3-channel
1247 * intrinsics. */
1248 if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false)) {
1249 LLVMValueRef v[3], v01;
1250
1251 for (int i = 0; i < 3; i++) {
1252 v[i] = LLVMBuildExtractElement(ctx->builder, vdata,
1253 LLVMConstInt(ctx->i32, i, 0), "");
1254 }
1255 v01 = ac_build_gather_values(ctx, v, 2);
1256
1257 ac_build_buffer_store_dword(ctx, rsrc, v01, 2, voffset,
1258 soffset, inst_offset, cache_policy,
1259 swizzle_enable_hint);
1260 ac_build_buffer_store_dword(ctx, rsrc, v[2], 1, voffset,
1261 soffset, inst_offset + 8,
1262 cache_policy,
1263 swizzle_enable_hint);
1264 return;
1265 }
1266
1267 /* SWIZZLE_ENABLE requires that soffset isn't folded into voffset
1268 * (voffset is swizzled, but soffset isn't swizzled).
1269 * llvm.amdgcn.buffer.store doesn't have a separate soffset parameter.
1270 */
1271 if (!swizzle_enable_hint) {
1272 LLVMValueRef offset = soffset;
1273
1274 if (inst_offset)
1275 offset = LLVMBuildAdd(ctx->builder, offset,
1276 LLVMConstInt(ctx->i32, inst_offset, 0), "");
1277
1278 if (HAVE_LLVM >= 0x800) {
1279 ac_build_llvm8_buffer_store_common(ctx, rsrc,
1280 ac_to_float(ctx, vdata),
1281 ctx->i32_0,
1282 voffset, offset,
1283 num_channels,
1284 ctx->f32,
1285 cache_policy,
1286 false, false);
1287 } else {
1288 if (voffset)
1289 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1290
1291 ac_build_llvm7_buffer_store_common(ctx, rsrc,
1292 ac_to_float(ctx, vdata),
1293 ctx->i32_0, offset,
1294 num_channels, cache_policy,
1295 false);
1296 }
1297 return;
1298 }
1299
1300 static const unsigned dfmts[] = {
1301 V_008F0C_BUF_DATA_FORMAT_32,
1302 V_008F0C_BUF_DATA_FORMAT_32_32,
1303 V_008F0C_BUF_DATA_FORMAT_32_32_32,
1304 V_008F0C_BUF_DATA_FORMAT_32_32_32_32
1305 };
1306 unsigned dfmt = dfmts[num_channels - 1];
1307 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1308 LLVMValueRef immoffset = LLVMConstInt(ctx->i32, inst_offset, 0);
1309
1310 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
1311 immoffset, num_channels, dfmt, nfmt, cache_policy);
1312 }
1313
1314 static LLVMValueRef
1315 ac_build_llvm7_buffer_load_common(struct ac_llvm_context *ctx,
1316 LLVMValueRef rsrc,
1317 LLVMValueRef vindex,
1318 LLVMValueRef voffset,
1319 unsigned num_channels,
1320 unsigned cache_policy,
1321 bool can_speculate,
1322 bool use_format)
1323 {
1324 LLVMValueRef args[] = {
1325 LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, ""),
1326 vindex ? vindex : ctx->i32_0,
1327 voffset,
1328 LLVMConstInt(ctx->i1, !!(cache_policy & ac_glc), 0),
1329 LLVMConstInt(ctx->i1, !!(cache_policy & ac_slc), 0)
1330 };
1331 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1332
1333 LLVMTypeRef types[] = {ctx->f32, ctx->v2f32, ctx->v4f32};
1334 const char *type_names[] = {"f32", "v2f32", "v4f32"};
1335 char name[256];
1336
1337 if (use_format) {
1338 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.format.%s",
1339 type_names[func]);
1340 } else {
1341 snprintf(name, sizeof(name), "llvm.amdgcn.buffer.load.%s",
1342 type_names[func]);
1343 }
1344
1345 return ac_build_intrinsic(ctx, name, types[func], args,
1346 ARRAY_SIZE(args),
1347 ac_get_load_intr_attribs(can_speculate));
1348 }
1349
1350 static LLVMValueRef
1351 ac_build_llvm8_buffer_load_common(struct ac_llvm_context *ctx,
1352 LLVMValueRef rsrc,
1353 LLVMValueRef vindex,
1354 LLVMValueRef voffset,
1355 LLVMValueRef soffset,
1356 unsigned num_channels,
1357 LLVMTypeRef channel_type,
1358 unsigned cache_policy,
1359 bool can_speculate,
1360 bool use_format,
1361 bool structurized)
1362 {
1363 LLVMValueRef args[5];
1364 int idx = 0;
1365 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1366 if (structurized)
1367 args[idx++] = vindex ? vindex : ctx->i32_0;
1368 args[idx++] = voffset ? voffset : ctx->i32_0;
1369 args[idx++] = soffset ? soffset : ctx->i32_0;
1370 args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
1371 unsigned func = !ac_has_vec3_support(ctx->chip_class, use_format) && num_channels == 3 ? 4 : num_channels;
1372 const char *indexing_kind = structurized ? "struct" : "raw";
1373 char name[256], type_name[8];
1374
1375 LLVMTypeRef type = func > 1 ? LLVMVectorType(channel_type, func) : channel_type;
1376 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1377
1378 if (use_format) {
1379 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.format.%s",
1380 indexing_kind, type_name);
1381 } else {
1382 snprintf(name, sizeof(name), "llvm.amdgcn.%s.buffer.load.%s",
1383 indexing_kind, type_name);
1384 }
1385
1386 return ac_build_intrinsic(ctx, name, type, args, idx,
1387 ac_get_load_intr_attribs(can_speculate));
1388 }
1389
1390 LLVMValueRef
1391 ac_build_buffer_load(struct ac_llvm_context *ctx,
1392 LLVMValueRef rsrc,
1393 int num_channels,
1394 LLVMValueRef vindex,
1395 LLVMValueRef voffset,
1396 LLVMValueRef soffset,
1397 unsigned inst_offset,
1398 unsigned cache_policy,
1399 bool can_speculate,
1400 bool allow_smem)
1401 {
1402 LLVMValueRef offset = LLVMConstInt(ctx->i32, inst_offset, 0);
1403 if (voffset)
1404 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
1405 if (soffset)
1406 offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
1407
1408 if (allow_smem && !(cache_policy & ac_slc) &&
1409 (!(cache_policy & ac_glc) || (HAVE_LLVM >= 0x0800 && ctx->chip_class >= GFX8))) {
1410 assert(vindex == NULL);
1411
1412 LLVMValueRef result[8];
1413
1414 for (int i = 0; i < num_channels; i++) {
1415 if (i) {
1416 offset = LLVMBuildAdd(ctx->builder, offset,
1417 LLVMConstInt(ctx->i32, 4, 0), "");
1418 }
1419 const char *intrname =
1420 HAVE_LLVM >= 0x0800 ? "llvm.amdgcn.s.buffer.load.f32"
1421 : "llvm.SI.load.const.v4i32";
1422 unsigned num_args = HAVE_LLVM >= 0x0800 ? 3 : 2;
1423 LLVMValueRef args[3] = {
1424 rsrc,
1425 offset,
1426 LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0),
1427 };
1428 result[i] = ac_build_intrinsic(ctx, intrname,
1429 ctx->f32, args, num_args,
1430 AC_FUNC_ATTR_READNONE |
1431 (HAVE_LLVM < 0x0800 ? AC_FUNC_ATTR_LEGACY : 0));
1432 }
1433 if (num_channels == 1)
1434 return result[0];
1435
1436 if (num_channels == 3 && !ac_has_vec3_support(ctx->chip_class, false))
1437 result[num_channels++] = LLVMGetUndef(ctx->f32);
1438 return ac_build_gather_values(ctx, result, num_channels);
1439 }
1440
1441 if (HAVE_LLVM >= 0x0800) {
1442 return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex,
1443 offset, ctx->i32_0,
1444 num_channels, ctx->f32,
1445 cache_policy,
1446 can_speculate, false,
1447 false);
1448 }
1449
1450 return ac_build_llvm7_buffer_load_common(ctx, rsrc, vindex, offset,
1451 num_channels, cache_policy,
1452 can_speculate, false);
1453 }
1454
1455 LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
1456 LLVMValueRef rsrc,
1457 LLVMValueRef vindex,
1458 LLVMValueRef voffset,
1459 unsigned num_channels,
1460 unsigned cache_policy,
1461 bool can_speculate)
1462 {
1463 if (HAVE_LLVM >= 0x800) {
1464 return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
1465 num_channels, ctx->f32,
1466 cache_policy, can_speculate, true, true);
1467 }
1468 return ac_build_llvm7_buffer_load_common(ctx, rsrc, vindex, voffset,
1469 num_channels, cache_policy,
1470 can_speculate, true);
1471 }
1472
1473 LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
1474 LLVMValueRef rsrc,
1475 LLVMValueRef vindex,
1476 LLVMValueRef voffset,
1477 unsigned num_channels,
1478 unsigned cache_policy,
1479 bool can_speculate)
1480 {
1481 if (HAVE_LLVM >= 0x800) {
1482 return ac_build_llvm8_buffer_load_common(ctx, rsrc, vindex, voffset, ctx->i32_0,
1483 num_channels, ctx->f32,
1484 cache_policy, can_speculate, true, true);
1485 }
1486
1487 LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->builder, rsrc, LLVMConstInt(ctx->i32, 2, 0), "");
1488 LLVMValueRef stride = LLVMBuildExtractElement(ctx->builder, rsrc, ctx->i32_1, "");
1489 stride = LLVMBuildLShr(ctx->builder, stride, LLVMConstInt(ctx->i32, 16, 0), "");
1490
1491 LLVMValueRef new_elem_count = LLVMBuildSelect(ctx->builder,
1492 LLVMBuildICmp(ctx->builder, LLVMIntUGT, elem_count, stride, ""),
1493 elem_count, stride, "");
1494
1495 LLVMValueRef new_rsrc = LLVMBuildInsertElement(ctx->builder, rsrc, new_elem_count,
1496 LLVMConstInt(ctx->i32, 2, 0), "");
1497
1498 return ac_build_llvm7_buffer_load_common(ctx, new_rsrc, vindex, voffset,
1499 num_channels, cache_policy,
1500 can_speculate, true);
1501 }
1502
1503 /// Translate a (dfmt, nfmt) pair into a chip-appropriate combined format
1504 /// value for LLVM8+ tbuffer intrinsics.
1505 static unsigned
1506 ac_get_tbuffer_format(struct ac_llvm_context *ctx,
1507 unsigned dfmt, unsigned nfmt)
1508 {
1509 if (ctx->chip_class >= GFX10) {
1510 unsigned format;
1511 switch (dfmt) {
1512 default: unreachable("bad dfmt");
1513 case V_008F0C_BUF_DATA_FORMAT_INVALID: format = V_008F0C_IMG_FORMAT_INVALID; break;
1514 case V_008F0C_BUF_DATA_FORMAT_8: format = V_008F0C_IMG_FORMAT_8_UINT; break;
1515 case V_008F0C_BUF_DATA_FORMAT_8_8: format = V_008F0C_IMG_FORMAT_8_8_UINT; break;
1516 case V_008F0C_BUF_DATA_FORMAT_8_8_8_8: format = V_008F0C_IMG_FORMAT_8_8_8_8_UINT; break;
1517 case V_008F0C_BUF_DATA_FORMAT_16: format = V_008F0C_IMG_FORMAT_16_UINT; break;
1518 case V_008F0C_BUF_DATA_FORMAT_16_16: format = V_008F0C_IMG_FORMAT_16_16_UINT; break;
1519 case V_008F0C_BUF_DATA_FORMAT_16_16_16_16: format = V_008F0C_IMG_FORMAT_16_16_16_16_UINT; break;
1520 case V_008F0C_BUF_DATA_FORMAT_32: format = V_008F0C_IMG_FORMAT_32_UINT; break;
1521 case V_008F0C_BUF_DATA_FORMAT_32_32: format = V_008F0C_IMG_FORMAT_32_32_UINT; break;
1522 case V_008F0C_BUF_DATA_FORMAT_32_32_32: format = V_008F0C_IMG_FORMAT_32_32_32_UINT; break;
1523 case V_008F0C_BUF_DATA_FORMAT_32_32_32_32: format = V_008F0C_IMG_FORMAT_32_32_32_32_UINT; break;
1524 case V_008F0C_BUF_DATA_FORMAT_2_10_10_10: format = V_008F0C_IMG_FORMAT_2_10_10_10_UINT; break;
1525 }
1526
1527 // Use the regularity properties of the combined format enum.
1528 //
1529 // Note: float is incompatible with 8-bit data formats,
1530 // [us]{norm,scaled} are incomparible with 32-bit data formats.
1531 // [us]scaled are not writable.
1532 switch (nfmt) {
1533 case V_008F0C_BUF_NUM_FORMAT_UNORM: format -= 4; break;
1534 case V_008F0C_BUF_NUM_FORMAT_SNORM: format -= 3; break;
1535 case V_008F0C_BUF_NUM_FORMAT_USCALED: format -= 2; break;
1536 case V_008F0C_BUF_NUM_FORMAT_SSCALED: format -= 1; break;
1537 default: unreachable("bad nfmt");
1538 case V_008F0C_BUF_NUM_FORMAT_UINT: break;
1539 case V_008F0C_BUF_NUM_FORMAT_SINT: format += 1; break;
1540 case V_008F0C_BUF_NUM_FORMAT_FLOAT: format += 2; break;
1541 }
1542
1543 return format;
1544 } else {
1545 return dfmt | (nfmt << 4);
1546 }
1547 }
1548
1549 static LLVMValueRef
1550 ac_build_llvm8_tbuffer_load(struct ac_llvm_context *ctx,
1551 LLVMValueRef rsrc,
1552 LLVMValueRef vindex,
1553 LLVMValueRef voffset,
1554 LLVMValueRef soffset,
1555 unsigned num_channels,
1556 unsigned dfmt,
1557 unsigned nfmt,
1558 unsigned cache_policy,
1559 bool can_speculate,
1560 bool structurized)
1561 {
1562 LLVMValueRef args[6];
1563 int idx = 0;
1564 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
1565 if (structurized)
1566 args[idx++] = vindex ? vindex : ctx->i32_0;
1567 args[idx++] = voffset ? voffset : ctx->i32_0;
1568 args[idx++] = soffset ? soffset : ctx->i32_0;
1569 args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx, dfmt, nfmt), 0);
1570 args[idx++] = LLVMConstInt(ctx->i32, get_load_cache_policy(ctx, cache_policy), 0);
1571 unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
1572 const char *indexing_kind = structurized ? "struct" : "raw";
1573 char name[256], type_name[8];
1574
1575 LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
1576 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
1577
1578 snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.load.%s",
1579 indexing_kind, type_name);
1580
1581 return ac_build_intrinsic(ctx, name, type, args, idx,
1582 ac_get_load_intr_attribs(can_speculate));
1583 }
1584
1585 static LLVMValueRef
1586 ac_build_tbuffer_load(struct ac_llvm_context *ctx,
1587 LLVMValueRef rsrc,
1588 LLVMValueRef vindex,
1589 LLVMValueRef voffset,
1590 LLVMValueRef soffset,
1591 LLVMValueRef immoffset,
1592 unsigned num_channels,
1593 unsigned dfmt,
1594 unsigned nfmt,
1595 unsigned cache_policy,
1596 bool can_speculate,
1597 bool structurized) /* only matters for LLVM 8+ */
1598 {
1599 if (HAVE_LLVM >= 0x800) {
1600 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1601
1602 return ac_build_llvm8_tbuffer_load(ctx, rsrc, vindex, voffset,
1603 soffset, num_channels,
1604 dfmt, nfmt, cache_policy,
1605 can_speculate, structurized);
1606 }
1607
1608 LLVMValueRef args[] = {
1609 rsrc,
1610 vindex ? vindex : ctx->i32_0,
1611 voffset,
1612 soffset,
1613 immoffset,
1614 LLVMConstInt(ctx->i32, dfmt, false),
1615 LLVMConstInt(ctx->i32, nfmt, false),
1616 LLVMConstInt(ctx->i1, !!(cache_policy & ac_glc), false),
1617 LLVMConstInt(ctx->i1, !!(cache_policy & ac_slc), false),
1618 };
1619 unsigned func = CLAMP(num_channels, 1, 3) - 1;
1620 LLVMTypeRef types[] = {ctx->i32, ctx->v2i32, ctx->v4i32};
1621 const char *type_names[] = {"i32", "v2i32", "v4i32"};
1622 char name[256];
1623
1624 snprintf(name, sizeof(name), "llvm.amdgcn.tbuffer.load.%s",
1625 type_names[func]);
1626
1627 return ac_build_intrinsic(ctx, name, types[func], args, 9,
1628 ac_get_load_intr_attribs(can_speculate));
1629 }
1630
1631 LLVMValueRef
1632 ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
1633 LLVMValueRef rsrc,
1634 LLVMValueRef vindex,
1635 LLVMValueRef voffset,
1636 LLVMValueRef soffset,
1637 LLVMValueRef immoffset,
1638 unsigned num_channels,
1639 unsigned dfmt,
1640 unsigned nfmt,
1641 unsigned cache_policy,
1642 bool can_speculate)
1643 {
1644 return ac_build_tbuffer_load(ctx, rsrc, vindex, voffset, soffset,
1645 immoffset, num_channels, dfmt, nfmt,
1646 cache_policy, can_speculate, true);
1647 }
1648
1649 LLVMValueRef
1650 ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
1651 LLVMValueRef rsrc,
1652 LLVMValueRef voffset,
1653 LLVMValueRef soffset,
1654 LLVMValueRef immoffset,
1655 unsigned num_channels,
1656 unsigned dfmt,
1657 unsigned nfmt,
1658 unsigned cache_policy,
1659 bool can_speculate)
1660 {
1661 return ac_build_tbuffer_load(ctx, rsrc, NULL, voffset, soffset,
1662 immoffset, num_channels, dfmt, nfmt,
1663 cache_policy, can_speculate, false);
1664 }
1665
1666 LLVMValueRef
1667 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
1668 LLVMValueRef rsrc,
1669 LLVMValueRef voffset,
1670 LLVMValueRef soffset,
1671 LLVMValueRef immoffset,
1672 unsigned cache_policy)
1673 {
1674 LLVMValueRef res;
1675
1676 if (HAVE_LLVM >= 0x900) {
1677 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1678
1679 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1680 res = ac_build_llvm8_buffer_load_common(ctx, rsrc, NULL,
1681 voffset, soffset,
1682 1, ctx->i16, cache_policy,
1683 false, false, false);
1684 } else {
1685 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
1686 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1687
1688 res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
1689 immoffset, 1, dfmt, nfmt, cache_policy,
1690 false);
1691
1692 res = LLVMBuildTrunc(ctx->builder, res, ctx->i16, "");
1693 }
1694
1695 return res;
1696 }
1697
1698 LLVMValueRef
1699 ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
1700 LLVMValueRef rsrc,
1701 LLVMValueRef voffset,
1702 LLVMValueRef soffset,
1703 LLVMValueRef immoffset,
1704 unsigned cache_policy)
1705 {
1706 LLVMValueRef res;
1707
1708 if (HAVE_LLVM >= 0x900) {
1709 voffset = LLVMBuildAdd(ctx->builder, voffset, immoffset, "");
1710
1711 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
1712 res = ac_build_llvm8_buffer_load_common(ctx, rsrc, NULL,
1713 voffset, soffset,
1714 1, ctx->i8, cache_policy,
1715 false, false, false);
1716 } else {
1717 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
1718 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
1719
1720 res = ac_build_raw_tbuffer_load(ctx, rsrc, voffset, soffset,
1721 immoffset, 1, dfmt, nfmt, cache_policy,
1722 false);
1723
1724 res = LLVMBuildTrunc(ctx->builder, res, ctx->i8, "");
1725 }
1726
1727 return res;
1728 }
1729
1730 /**
1731 * Convert an 11- or 10-bit unsigned floating point number to an f32.
1732 *
1733 * The input exponent is expected to be biased analogous to IEEE-754, i.e. by
1734 * 2^(exp_bits-1) - 1 (as defined in OpenGL and other graphics APIs).
1735 */
1736 static LLVMValueRef
1737 ac_ufN_to_float(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned exp_bits, unsigned mant_bits)
1738 {
1739 assert(LLVMTypeOf(src) == ctx->i32);
1740
1741 LLVMValueRef tmp;
1742 LLVMValueRef mantissa;
1743 mantissa = LLVMBuildAnd(ctx->builder, src, LLVMConstInt(ctx->i32, (1 << mant_bits) - 1, false), "");
1744
1745 /* Converting normal numbers is just a shift + correcting the exponent bias */
1746 unsigned normal_shift = 23 - mant_bits;
1747 unsigned bias_shift = 127 - ((1 << (exp_bits - 1)) - 1);
1748 LLVMValueRef shifted, normal;
1749
1750 shifted = LLVMBuildShl(ctx->builder, src, LLVMConstInt(ctx->i32, normal_shift, false), "");
1751 normal = LLVMBuildAdd(ctx->builder, shifted, LLVMConstInt(ctx->i32, bias_shift << 23, false), "");
1752
1753 /* Converting nan/inf numbers is the same, but with a different exponent update */
1754 LLVMValueRef naninf;
1755 naninf = LLVMBuildOr(ctx->builder, normal, LLVMConstInt(ctx->i32, 0xff << 23, false), "");
1756
1757 /* Converting denormals is the complex case: determine the leading zeros of the
1758 * mantissa to obtain the correct shift for the mantissa and exponent correction.
1759 */
1760 LLVMValueRef denormal;
1761 LLVMValueRef params[2] = {
1762 mantissa,
1763 ctx->i1true, /* result can be undef when arg is 0 */
1764 };
1765 LLVMValueRef ctlz = ac_build_intrinsic(ctx, "llvm.ctlz.i32", ctx->i32,
1766 params, 2, AC_FUNC_ATTR_READNONE);
1767
1768 /* Shift such that the leading 1 ends up as the LSB of the exponent field. */
1769 tmp = LLVMBuildSub(ctx->builder, ctlz, LLVMConstInt(ctx->i32, 8, false), "");
1770 denormal = LLVMBuildShl(ctx->builder, mantissa, tmp, "");
1771
1772 unsigned denormal_exp = bias_shift + (32 - mant_bits) - 1;
1773 tmp = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, denormal_exp, false), ctlz, "");
1774 tmp = LLVMBuildShl(ctx->builder, tmp, LLVMConstInt(ctx->i32, 23, false), "");
1775 denormal = LLVMBuildAdd(ctx->builder, denormal, tmp, "");
1776
1777 /* Select the final result. */
1778 LLVMValueRef result;
1779
1780 tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src,
1781 LLVMConstInt(ctx->i32, ((1 << exp_bits) - 1) << mant_bits, false), "");
1782 result = LLVMBuildSelect(ctx->builder, tmp, naninf, normal, "");
1783
1784 tmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, src,
1785 LLVMConstInt(ctx->i32, 1 << mant_bits, false), "");
1786 result = LLVMBuildSelect(ctx->builder, tmp, result, denormal, "");
1787
1788 tmp = LLVMBuildICmp(ctx->builder, LLVMIntNE, src, ctx->i32_0, "");
1789 result = LLVMBuildSelect(ctx->builder, tmp, result, ctx->i32_0, "");
1790
1791 return ac_to_float(ctx, result);
1792 }
1793
1794 /**
1795 * Generate a fully general open coded buffer format fetch with all required
1796 * fixups suitable for vertex fetch, using non-format buffer loads.
1797 *
1798 * Some combinations of argument values have special interpretations:
1799 * - size = 8 bytes, format = fixed indicates PIPE_FORMAT_R11G11B10_FLOAT
1800 * - size = 8 bytes, format != {float,fixed} indicates a 2_10_10_10 data format
1801 *
1802 * \param log_size log(size of channel in bytes)
1803 * \param num_channels number of channels (1 to 4)
1804 * \param format AC_FETCH_FORMAT_xxx value
1805 * \param reverse whether XYZ channels are reversed
1806 * \param known_aligned whether the source is known to be aligned to hardware's
1807 * effective element size for loading the given format
1808 * (note: this means dword alignment for 8_8_8_8, 16_16, etc.)
1809 * \param rsrc buffer resource descriptor
1810 * \return the resulting vector of floats or integers bitcast to <4 x i32>
1811 */
1812 LLVMValueRef
1813 ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
1814 unsigned log_size,
1815 unsigned num_channels,
1816 unsigned format,
1817 bool reverse,
1818 bool known_aligned,
1819 LLVMValueRef rsrc,
1820 LLVMValueRef vindex,
1821 LLVMValueRef voffset,
1822 LLVMValueRef soffset,
1823 unsigned cache_policy,
1824 bool can_speculate)
1825 {
1826 LLVMValueRef tmp;
1827 unsigned load_log_size = log_size;
1828 unsigned load_num_channels = num_channels;
1829 if (log_size == 3) {
1830 load_log_size = 2;
1831 if (format == AC_FETCH_FORMAT_FLOAT) {
1832 load_num_channels = 2 * num_channels;
1833 } else {
1834 load_num_channels = 1; /* 10_11_11 or 2_10_10_10 */
1835 }
1836 }
1837
1838 int log_recombine = 0;
1839 if (ctx->chip_class == GFX6 && !known_aligned) {
1840 /* Avoid alignment restrictions by loading one byte at a time. */
1841 load_num_channels <<= load_log_size;
1842 log_recombine = load_log_size;
1843 load_log_size = 0;
1844 } else if (load_num_channels == 2 || load_num_channels == 4) {
1845 log_recombine = -util_logbase2(load_num_channels);
1846 load_num_channels = 1;
1847 load_log_size += -log_recombine;
1848 }
1849
1850 assert(load_log_size >= 2 || HAVE_LLVM >= 0x0900);
1851
1852 LLVMValueRef loads[32]; /* up to 32 bytes */
1853 for (unsigned i = 0; i < load_num_channels; ++i) {
1854 tmp = LLVMBuildAdd(ctx->builder, soffset,
1855 LLVMConstInt(ctx->i32, i << load_log_size, false), "");
1856 if (HAVE_LLVM >= 0x0800) {
1857 LLVMTypeRef channel_type = load_log_size == 0 ? ctx->i8 :
1858 load_log_size == 1 ? ctx->i16 : ctx->i32;
1859 unsigned num_channels = 1 << (MAX2(load_log_size, 2) - 2);
1860 loads[i] = ac_build_llvm8_buffer_load_common(
1861 ctx, rsrc, vindex, voffset, tmp,
1862 num_channels, channel_type, cache_policy,
1863 can_speculate, false, true);
1864 } else {
1865 tmp = LLVMBuildAdd(ctx->builder, voffset, tmp, "");
1866 loads[i] = ac_build_llvm7_buffer_load_common(
1867 ctx, rsrc, vindex, tmp,
1868 1 << (load_log_size - 2), cache_policy, can_speculate, false);
1869 }
1870 if (load_log_size >= 2)
1871 loads[i] = ac_to_integer(ctx, loads[i]);
1872 }
1873
1874 if (log_recombine > 0) {
1875 /* Recombine bytes if necessary (GFX6 only) */
1876 LLVMTypeRef dst_type = log_recombine == 2 ? ctx->i32 : ctx->i16;
1877
1878 for (unsigned src = 0, dst = 0; src < load_num_channels; ++dst) {
1879 LLVMValueRef accum = NULL;
1880 for (unsigned i = 0; i < (1 << log_recombine); ++i, ++src) {
1881 tmp = LLVMBuildZExt(ctx->builder, loads[src], dst_type, "");
1882 if (i == 0) {
1883 accum = tmp;
1884 } else {
1885 tmp = LLVMBuildShl(ctx->builder, tmp,
1886 LLVMConstInt(dst_type, 8 * i, false), "");
1887 accum = LLVMBuildOr(ctx->builder, accum, tmp, "");
1888 }
1889 }
1890 loads[dst] = accum;
1891 }
1892 } else if (log_recombine < 0) {
1893 /* Split vectors of dwords */
1894 if (load_log_size > 2) {
1895 assert(load_num_channels == 1);
1896 LLVMValueRef loaded = loads[0];
1897 unsigned log_split = load_log_size - 2;
1898 log_recombine += log_split;
1899 load_num_channels = 1 << log_split;
1900 load_log_size = 2;
1901 for (unsigned i = 0; i < load_num_channels; ++i) {
1902 tmp = LLVMConstInt(ctx->i32, i, false);
1903 loads[i] = LLVMBuildExtractElement(ctx->builder, loaded, tmp, "");
1904 }
1905 }
1906
1907 /* Further split dwords and shorts if required */
1908 if (log_recombine < 0) {
1909 for (unsigned src = load_num_channels,
1910 dst = load_num_channels << -log_recombine;
1911 src > 0; --src) {
1912 unsigned dst_bits = 1 << (3 + load_log_size + log_recombine);
1913 LLVMTypeRef dst_type = LLVMIntTypeInContext(ctx->context, dst_bits);
1914 LLVMValueRef loaded = loads[src - 1];
1915 LLVMTypeRef loaded_type = LLVMTypeOf(loaded);
1916 for (unsigned i = 1 << -log_recombine; i > 0; --i, --dst) {
1917 tmp = LLVMConstInt(loaded_type, dst_bits * (i - 1), false);
1918 tmp = LLVMBuildLShr(ctx->builder, loaded, tmp, "");
1919 loads[dst - 1] = LLVMBuildTrunc(ctx->builder, tmp, dst_type, "");
1920 }
1921 }
1922 }
1923 }
1924
1925 if (log_size == 3) {
1926 if (format == AC_FETCH_FORMAT_FLOAT) {
1927 for (unsigned i = 0; i < num_channels; ++i) {
1928 tmp = ac_build_gather_values(ctx, &loads[2 * i], 2);
1929 loads[i] = LLVMBuildBitCast(ctx->builder, tmp, ctx->f64, "");
1930 }
1931 } else if (format == AC_FETCH_FORMAT_FIXED) {
1932 /* 10_11_11_FLOAT */
1933 LLVMValueRef data = loads[0];
1934 LLVMValueRef i32_2047 = LLVMConstInt(ctx->i32, 2047, false);
1935 LLVMValueRef r = LLVMBuildAnd(ctx->builder, data, i32_2047, "");
1936 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 11, false), "");
1937 LLVMValueRef g = LLVMBuildAnd(ctx->builder, tmp, i32_2047, "");
1938 LLVMValueRef b = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 22, false), "");
1939
1940 loads[0] = ac_to_integer(ctx, ac_ufN_to_float(ctx, r, 5, 6));
1941 loads[1] = ac_to_integer(ctx, ac_ufN_to_float(ctx, g, 5, 6));
1942 loads[2] = ac_to_integer(ctx, ac_ufN_to_float(ctx, b, 5, 5));
1943
1944 num_channels = 3;
1945 log_size = 2;
1946 format = AC_FETCH_FORMAT_FLOAT;
1947 } else {
1948 /* 2_10_10_10 data formats */
1949 LLVMValueRef data = loads[0];
1950 LLVMTypeRef i10 = LLVMIntTypeInContext(ctx->context, 10);
1951 LLVMTypeRef i2 = LLVMIntTypeInContext(ctx->context, 2);
1952 loads[0] = LLVMBuildTrunc(ctx->builder, data, i10, "");
1953 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 10, false), "");
1954 loads[1] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
1955 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 20, false), "");
1956 loads[2] = LLVMBuildTrunc(ctx->builder, tmp, i10, "");
1957 tmp = LLVMBuildLShr(ctx->builder, data, LLVMConstInt(ctx->i32, 30, false), "");
1958 loads[3] = LLVMBuildTrunc(ctx->builder, tmp, i2, "");
1959
1960 num_channels = 4;
1961 }
1962 }
1963
1964 if (format == AC_FETCH_FORMAT_FLOAT) {
1965 if (log_size != 2) {
1966 for (unsigned chan = 0; chan < num_channels; ++chan) {
1967 tmp = ac_to_float(ctx, loads[chan]);
1968 if (log_size == 3)
1969 tmp = LLVMBuildFPTrunc(ctx->builder, tmp, ctx->f32, "");
1970 else if (log_size == 1)
1971 tmp = LLVMBuildFPExt(ctx->builder, tmp, ctx->f32, "");
1972 loads[chan] = ac_to_integer(ctx, tmp);
1973 }
1974 }
1975 } else if (format == AC_FETCH_FORMAT_UINT) {
1976 if (log_size != 2) {
1977 for (unsigned chan = 0; chan < num_channels; ++chan)
1978 loads[chan] = LLVMBuildZExt(ctx->builder, loads[chan], ctx->i32, "");
1979 }
1980 } else if (format == AC_FETCH_FORMAT_SINT) {
1981 if (log_size != 2) {
1982 for (unsigned chan = 0; chan < num_channels; ++chan)
1983 loads[chan] = LLVMBuildSExt(ctx->builder, loads[chan], ctx->i32, "");
1984 }
1985 } else {
1986 bool unsign = format == AC_FETCH_FORMAT_UNORM ||
1987 format == AC_FETCH_FORMAT_USCALED ||
1988 format == AC_FETCH_FORMAT_UINT;
1989
1990 for (unsigned chan = 0; chan < num_channels; ++chan) {
1991 if (unsign) {
1992 tmp = LLVMBuildUIToFP(ctx->builder, loads[chan], ctx->f32, "");
1993 } else {
1994 tmp = LLVMBuildSIToFP(ctx->builder, loads[chan], ctx->f32, "");
1995 }
1996
1997 LLVMValueRef scale = NULL;
1998 if (format == AC_FETCH_FORMAT_FIXED) {
1999 assert(log_size == 2);
2000 scale = LLVMConstReal(ctx->f32, 1.0 / 0x10000);
2001 } else if (format == AC_FETCH_FORMAT_UNORM) {
2002 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
2003 scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << bits) - 1));
2004 } else if (format == AC_FETCH_FORMAT_SNORM) {
2005 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(loads[chan]));
2006 scale = LLVMConstReal(ctx->f32, 1.0 / (((uint64_t)1 << (bits - 1)) - 1));
2007 }
2008 if (scale)
2009 tmp = LLVMBuildFMul(ctx->builder, tmp, scale, "");
2010
2011 if (format == AC_FETCH_FORMAT_SNORM) {
2012 /* Clamp to [-1, 1] */
2013 LLVMValueRef neg_one = LLVMConstReal(ctx->f32, -1.0);
2014 LLVMValueRef clamp =
2015 LLVMBuildFCmp(ctx->builder, LLVMRealULT, tmp, neg_one, "");
2016 tmp = LLVMBuildSelect(ctx->builder, clamp, neg_one, tmp, "");
2017 }
2018
2019 loads[chan] = ac_to_integer(ctx, tmp);
2020 }
2021 }
2022
2023 while (num_channels < 4) {
2024 if (format == AC_FETCH_FORMAT_UINT || format == AC_FETCH_FORMAT_SINT) {
2025 loads[num_channels] = num_channels == 3 ? ctx->i32_1 : ctx->i32_0;
2026 } else {
2027 loads[num_channels] = ac_to_integer(ctx, num_channels == 3 ? ctx->f32_1 : ctx->f32_0);
2028 }
2029 num_channels++;
2030 }
2031
2032 if (reverse) {
2033 tmp = loads[0];
2034 loads[0] = loads[2];
2035 loads[2] = tmp;
2036 }
2037
2038 return ac_build_gather_values(ctx, loads, 4);
2039 }
2040
2041 static void
2042 ac_build_llvm8_tbuffer_store(struct ac_llvm_context *ctx,
2043 LLVMValueRef rsrc,
2044 LLVMValueRef vdata,
2045 LLVMValueRef vindex,
2046 LLVMValueRef voffset,
2047 LLVMValueRef soffset,
2048 unsigned num_channels,
2049 unsigned dfmt,
2050 unsigned nfmt,
2051 unsigned cache_policy,
2052 bool structurized)
2053 {
2054 LLVMValueRef args[7];
2055 int idx = 0;
2056 args[idx++] = vdata;
2057 args[idx++] = LLVMBuildBitCast(ctx->builder, rsrc, ctx->v4i32, "");
2058 if (structurized)
2059 args[idx++] = vindex ? vindex : ctx->i32_0;
2060 args[idx++] = voffset ? voffset : ctx->i32_0;
2061 args[idx++] = soffset ? soffset : ctx->i32_0;
2062 args[idx++] = LLVMConstInt(ctx->i32, ac_get_tbuffer_format(ctx, dfmt, nfmt), 0);
2063 args[idx++] = LLVMConstInt(ctx->i32, cache_policy, 0);
2064 unsigned func = !ac_has_vec3_support(ctx->chip_class, true) && num_channels == 3 ? 4 : num_channels;
2065 const char *indexing_kind = structurized ? "struct" : "raw";
2066 char name[256], type_name[8];
2067
2068 LLVMTypeRef type = func > 1 ? LLVMVectorType(ctx->i32, func) : ctx->i32;
2069 ac_build_type_name_for_intr(type, type_name, sizeof(type_name));
2070
2071 snprintf(name, sizeof(name), "llvm.amdgcn.%s.tbuffer.store.%s",
2072 indexing_kind, type_name);
2073
2074 ac_build_intrinsic(ctx, name, ctx->voidt, args, idx,
2075 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
2076 }
2077
2078 static void
2079 ac_build_tbuffer_store(struct ac_llvm_context *ctx,
2080 LLVMValueRef rsrc,
2081 LLVMValueRef vdata,
2082 LLVMValueRef vindex,
2083 LLVMValueRef voffset,
2084 LLVMValueRef soffset,
2085 LLVMValueRef immoffset,
2086 unsigned num_channels,
2087 unsigned dfmt,
2088 unsigned nfmt,
2089 unsigned cache_policy,
2090 bool structurized) /* only matters for LLVM 8+ */
2091 {
2092 if (HAVE_LLVM >= 0x800) {
2093 voffset = LLVMBuildAdd(ctx->builder,
2094 voffset ? voffset : ctx->i32_0,
2095 immoffset, "");
2096
2097 ac_build_llvm8_tbuffer_store(ctx, rsrc, vdata, vindex, voffset,
2098 soffset, num_channels, dfmt, nfmt,
2099 cache_policy, structurized);
2100 } else {
2101 LLVMValueRef params[] = {
2102 vdata,
2103 rsrc,
2104 vindex ? vindex : ctx->i32_0,
2105 voffset ? voffset : ctx->i32_0,
2106 soffset ? soffset : ctx->i32_0,
2107 immoffset,
2108 LLVMConstInt(ctx->i32, dfmt, false),
2109 LLVMConstInt(ctx->i32, nfmt, false),
2110 LLVMConstInt(ctx->i1, !!(cache_policy & ac_glc), false),
2111 LLVMConstInt(ctx->i1, !!(cache_policy & ac_slc), false),
2112 };
2113 unsigned func = CLAMP(num_channels, 1, 3) - 1;
2114 const char *type_names[] = {"i32", "v2i32", "v4i32"};
2115 char name[256];
2116
2117 snprintf(name, sizeof(name), "llvm.amdgcn.tbuffer.store.%s",
2118 type_names[func]);
2119
2120 ac_build_intrinsic(ctx, name, ctx->voidt, params, 10,
2121 AC_FUNC_ATTR_INACCESSIBLE_MEM_ONLY);
2122 }
2123 }
2124
2125 void
2126 ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
2127 LLVMValueRef rsrc,
2128 LLVMValueRef vdata,
2129 LLVMValueRef vindex,
2130 LLVMValueRef voffset,
2131 LLVMValueRef soffset,
2132 LLVMValueRef immoffset,
2133 unsigned num_channels,
2134 unsigned dfmt,
2135 unsigned nfmt,
2136 unsigned cache_policy)
2137 {
2138 ac_build_tbuffer_store(ctx, rsrc, vdata, vindex, voffset, soffset,
2139 immoffset, num_channels, dfmt, nfmt, cache_policy,
2140 true);
2141 }
2142
2143 void
2144 ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
2145 LLVMValueRef rsrc,
2146 LLVMValueRef vdata,
2147 LLVMValueRef voffset,
2148 LLVMValueRef soffset,
2149 LLVMValueRef immoffset,
2150 unsigned num_channels,
2151 unsigned dfmt,
2152 unsigned nfmt,
2153 unsigned cache_policy)
2154 {
2155 ac_build_tbuffer_store(ctx, rsrc, vdata, NULL, voffset, soffset,
2156 immoffset, num_channels, dfmt, nfmt, cache_policy,
2157 false);
2158 }
2159
2160 void
2161 ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
2162 LLVMValueRef rsrc,
2163 LLVMValueRef vdata,
2164 LLVMValueRef voffset,
2165 LLVMValueRef soffset,
2166 unsigned cache_policy)
2167 {
2168 vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i16, "");
2169
2170 if (HAVE_LLVM >= 0x900) {
2171 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
2172 ac_build_llvm8_buffer_store_common(ctx, rsrc, vdata, NULL,
2173 voffset, soffset, 1,
2174 ctx->i16, cache_policy,
2175 false, false);
2176 } else {
2177 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_16;
2178 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
2179
2180 vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
2181
2182 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
2183 ctx->i32_0, 1, dfmt, nfmt, cache_policy);
2184 }
2185 }
2186
2187 void
2188 ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
2189 LLVMValueRef rsrc,
2190 LLVMValueRef vdata,
2191 LLVMValueRef voffset,
2192 LLVMValueRef soffset,
2193 unsigned cache_policy)
2194 {
2195 vdata = LLVMBuildBitCast(ctx->builder, vdata, ctx->i8, "");
2196
2197 if (HAVE_LLVM >= 0x900) {
2198 /* LLVM 9+ supports i8/i16 with struct/raw intrinsics. */
2199 ac_build_llvm8_buffer_store_common(ctx, rsrc, vdata, NULL,
2200 voffset, soffset, 1,
2201 ctx->i8, cache_policy,
2202 false, false);
2203 } else {
2204 unsigned dfmt = V_008F0C_BUF_DATA_FORMAT_8;
2205 unsigned nfmt = V_008F0C_BUF_NUM_FORMAT_UINT;
2206
2207 vdata = LLVMBuildZExt(ctx->builder, vdata, ctx->i32, "");
2208
2209 ac_build_raw_tbuffer_store(ctx, rsrc, vdata, voffset, soffset,
2210 ctx->i32_0, 1, dfmt, nfmt, cache_policy);
2211 }
2212 }
2213 /**
2214 * Set range metadata on an instruction. This can only be used on load and
2215 * call instructions. If you know an instruction can only produce the values
2216 * 0, 1, 2, you would do set_range_metadata(value, 0, 3);
2217 * \p lo is the minimum value inclusive.
2218 * \p hi is the maximum value exclusive.
2219 */
2220 static void set_range_metadata(struct ac_llvm_context *ctx,
2221 LLVMValueRef value, unsigned lo, unsigned hi)
2222 {
2223 LLVMValueRef range_md, md_args[2];
2224 LLVMTypeRef type = LLVMTypeOf(value);
2225 LLVMContextRef context = LLVMGetTypeContext(type);
2226
2227 md_args[0] = LLVMConstInt(type, lo, false);
2228 md_args[1] = LLVMConstInt(type, hi, false);
2229 range_md = LLVMMDNodeInContext(context, md_args, 2);
2230 LLVMSetMetadata(value, ctx->range_md_kind, range_md);
2231 }
2232
2233 LLVMValueRef
2234 ac_get_thread_id(struct ac_llvm_context *ctx)
2235 {
2236 LLVMValueRef tid;
2237
2238 LLVMValueRef tid_args[2];
2239 tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false);
2240 tid_args[1] = ctx->i32_0;
2241 tid_args[1] = ac_build_intrinsic(ctx,
2242 "llvm.amdgcn.mbcnt.lo", ctx->i32,
2243 tid_args, 2, AC_FUNC_ATTR_READNONE);
2244
2245 if (ctx->wave_size == 32) {
2246 tid = tid_args[1];
2247 } else {
2248 tid = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi",
2249 ctx->i32, tid_args,
2250 2, AC_FUNC_ATTR_READNONE);
2251 }
2252 set_range_metadata(ctx, tid, 0, ctx->wave_size);
2253 return tid;
2254 }
2255
2256 /*
2257 * AMD GCN implements derivatives using the local data store (LDS)
2258 * All writes to the LDS happen in all executing threads at
2259 * the same time. TID is the Thread ID for the current
2260 * thread and is a value between 0 and 63, representing
2261 * the thread's position in the wavefront.
2262 *
2263 * For the pixel shader threads are grouped into quads of four pixels.
2264 * The TIDs of the pixels of a quad are:
2265 *
2266 * +------+------+
2267 * |4n + 0|4n + 1|
2268 * +------+------+
2269 * |4n + 2|4n + 3|
2270 * +------+------+
2271 *
2272 * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
2273 * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
2274 * the current pixel's column, and masking with 0xfffffffe yields the TID
2275 * of the left pixel of the current pixel's row.
2276 *
2277 * Adding 1 yields the TID of the pixel to the right of the left pixel, and
2278 * adding 2 yields the TID of the pixel below the top pixel.
2279 */
2280 LLVMValueRef
2281 ac_build_ddxy(struct ac_llvm_context *ctx,
2282 uint32_t mask,
2283 int idx,
2284 LLVMValueRef val)
2285 {
2286 unsigned tl_lanes[4], trbl_lanes[4];
2287 char name[32], type[8];
2288 LLVMValueRef tl, trbl;
2289 LLVMTypeRef result_type;
2290 LLVMValueRef result;
2291
2292 result_type = ac_to_float_type(ctx, LLVMTypeOf(val));
2293
2294 if (result_type == ctx->f16)
2295 val = LLVMBuildZExt(ctx->builder, val, ctx->i32, "");
2296
2297 for (unsigned i = 0; i < 4; ++i) {
2298 tl_lanes[i] = i & mask;
2299 trbl_lanes[i] = (i & mask) + idx;
2300 }
2301
2302 tl = ac_build_quad_swizzle(ctx, val,
2303 tl_lanes[0], tl_lanes[1],
2304 tl_lanes[2], tl_lanes[3]);
2305 trbl = ac_build_quad_swizzle(ctx, val,
2306 trbl_lanes[0], trbl_lanes[1],
2307 trbl_lanes[2], trbl_lanes[3]);
2308
2309 if (result_type == ctx->f16) {
2310 tl = LLVMBuildTrunc(ctx->builder, tl, ctx->i16, "");
2311 trbl = LLVMBuildTrunc(ctx->builder, trbl, ctx->i16, "");
2312 }
2313
2314 tl = LLVMBuildBitCast(ctx->builder, tl, result_type, "");
2315 trbl = LLVMBuildBitCast(ctx->builder, trbl, result_type, "");
2316 result = LLVMBuildFSub(ctx->builder, trbl, tl, "");
2317
2318 ac_build_type_name_for_intr(result_type, type, sizeof(type));
2319 snprintf(name, sizeof(name), "llvm.amdgcn.wqm.%s", type);
2320
2321 return ac_build_intrinsic(ctx, name, result_type, &result, 1, 0);
2322 }
2323
2324 void
2325 ac_build_sendmsg(struct ac_llvm_context *ctx,
2326 uint32_t msg,
2327 LLVMValueRef wave_id)
2328 {
2329 LLVMValueRef args[2];
2330 args[0] = LLVMConstInt(ctx->i32, msg, false);
2331 args[1] = wave_id;
2332 ac_build_intrinsic(ctx, "llvm.amdgcn.s.sendmsg", ctx->voidt, args, 2, 0);
2333 }
2334
2335 LLVMValueRef
2336 ac_build_imsb(struct ac_llvm_context *ctx,
2337 LLVMValueRef arg,
2338 LLVMTypeRef dst_type)
2339 {
2340 LLVMValueRef msb = ac_build_intrinsic(ctx, "llvm.amdgcn.sffbh.i32",
2341 dst_type, &arg, 1,
2342 AC_FUNC_ATTR_READNONE);
2343
2344 /* The HW returns the last bit index from MSB, but NIR/TGSI wants
2345 * the index from LSB. Invert it by doing "31 - msb". */
2346 msb = LLVMBuildSub(ctx->builder, LLVMConstInt(ctx->i32, 31, false),
2347 msb, "");
2348
2349 LLVMValueRef all_ones = LLVMConstInt(ctx->i32, -1, true);
2350 LLVMValueRef cond = LLVMBuildOr(ctx->builder,
2351 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
2352 arg, ctx->i32_0, ""),
2353 LLVMBuildICmp(ctx->builder, LLVMIntEQ,
2354 arg, all_ones, ""), "");
2355
2356 return LLVMBuildSelect(ctx->builder, cond, all_ones, msb, "");
2357 }
2358
2359 LLVMValueRef
2360 ac_build_umsb(struct ac_llvm_context *ctx,
2361 LLVMValueRef arg,
2362 LLVMTypeRef dst_type)
2363 {
2364 const char *intrin_name;
2365 LLVMTypeRef type;
2366 LLVMValueRef highest_bit;
2367 LLVMValueRef zero;
2368 unsigned bitsize;
2369
2370 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(arg));
2371 switch (bitsize) {
2372 case 64:
2373 intrin_name = "llvm.ctlz.i64";
2374 type = ctx->i64;
2375 highest_bit = LLVMConstInt(ctx->i64, 63, false);
2376 zero = ctx->i64_0;
2377 break;
2378 case 32:
2379 intrin_name = "llvm.ctlz.i32";
2380 type = ctx->i32;
2381 highest_bit = LLVMConstInt(ctx->i32, 31, false);
2382 zero = ctx->i32_0;
2383 break;
2384 case 16:
2385 intrin_name = "llvm.ctlz.i16";
2386 type = ctx->i16;
2387 highest_bit = LLVMConstInt(ctx->i16, 15, false);
2388 zero = ctx->i16_0;
2389 break;
2390 case 8:
2391 intrin_name = "llvm.ctlz.i8";
2392 type = ctx->i8;
2393 highest_bit = LLVMConstInt(ctx->i8, 7, false);
2394 zero = ctx->i8_0;
2395 break;
2396 default:
2397 unreachable(!"invalid bitsize");
2398 break;
2399 }
2400
2401 LLVMValueRef params[2] = {
2402 arg,
2403 ctx->i1true,
2404 };
2405
2406 LLVMValueRef msb = ac_build_intrinsic(ctx, intrin_name, type,
2407 params, 2,
2408 AC_FUNC_ATTR_READNONE);
2409
2410 /* The HW returns the last bit index from MSB, but TGSI/NIR wants
2411 * the index from LSB. Invert it by doing "31 - msb". */
2412 msb = LLVMBuildSub(ctx->builder, highest_bit, msb, "");
2413
2414 if (bitsize == 64) {
2415 msb = LLVMBuildTrunc(ctx->builder, msb, ctx->i32, "");
2416 } else if (bitsize < 32) {
2417 msb = LLVMBuildSExt(ctx->builder, msb, ctx->i32, "");
2418 }
2419
2420 /* check for zero */
2421 return LLVMBuildSelect(ctx->builder,
2422 LLVMBuildICmp(ctx->builder, LLVMIntEQ, arg, zero, ""),
2423 LLVMConstInt(ctx->i32, -1, true), msb, "");
2424 }
2425
2426 LLVMValueRef ac_build_fmin(struct ac_llvm_context *ctx, LLVMValueRef a,
2427 LLVMValueRef b)
2428 {
2429 char name[64];
2430 snprintf(name, sizeof(name), "llvm.minnum.f%d", ac_get_elem_bits(ctx, LLVMTypeOf(a)));
2431 LLVMValueRef args[2] = {a, b};
2432 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
2433 AC_FUNC_ATTR_READNONE);
2434 }
2435
2436 LLVMValueRef ac_build_fmax(struct ac_llvm_context *ctx, LLVMValueRef a,
2437 LLVMValueRef b)
2438 {
2439 char name[64];
2440 snprintf(name, sizeof(name), "llvm.maxnum.f%d", ac_get_elem_bits(ctx, LLVMTypeOf(a)));
2441 LLVMValueRef args[2] = {a, b};
2442 return ac_build_intrinsic(ctx, name, LLVMTypeOf(a), args, 2,
2443 AC_FUNC_ATTR_READNONE);
2444 }
2445
2446 LLVMValueRef ac_build_imin(struct ac_llvm_context *ctx, LLVMValueRef a,
2447 LLVMValueRef b)
2448 {
2449 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSLE, a, b, "");
2450 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2451 }
2452
2453 LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
2454 LLVMValueRef b)
2455 {
2456 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, a, b, "");
2457 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2458 }
2459
2460 LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a,
2461 LLVMValueRef b)
2462 {
2463 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntULE, a, b, "");
2464 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2465 }
2466
2467 LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a,
2468 LLVMValueRef b)
2469 {
2470 LLVMValueRef cmp = LLVMBuildICmp(ctx->builder, LLVMIntUGE, a, b, "");
2471 return LLVMBuildSelect(ctx->builder, cmp, a, b, "");
2472 }
2473
2474 LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value)
2475 {
2476 LLVMTypeRef t = LLVMTypeOf(value);
2477 return ac_build_fmin(ctx, ac_build_fmax(ctx, value, LLVMConstReal(t, 0.0)),
2478 LLVMConstReal(t, 1.0));
2479 }
2480
2481 void ac_build_export(struct ac_llvm_context *ctx, struct ac_export_args *a)
2482 {
2483 LLVMValueRef args[9];
2484
2485 args[0] = LLVMConstInt(ctx->i32, a->target, 0);
2486 args[1] = LLVMConstInt(ctx->i32, a->enabled_channels, 0);
2487
2488 if (a->compr) {
2489 LLVMTypeRef i16 = LLVMInt16TypeInContext(ctx->context);
2490 LLVMTypeRef v2i16 = LLVMVectorType(i16, 2);
2491
2492 args[2] = LLVMBuildBitCast(ctx->builder, a->out[0],
2493 v2i16, "");
2494 args[3] = LLVMBuildBitCast(ctx->builder, a->out[1],
2495 v2i16, "");
2496 args[4] = LLVMConstInt(ctx->i1, a->done, 0);
2497 args[5] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
2498
2499 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.compr.v2i16",
2500 ctx->voidt, args, 6, 0);
2501 } else {
2502 args[2] = a->out[0];
2503 args[3] = a->out[1];
2504 args[4] = a->out[2];
2505 args[5] = a->out[3];
2506 args[6] = LLVMConstInt(ctx->i1, a->done, 0);
2507 args[7] = LLVMConstInt(ctx->i1, a->valid_mask, 0);
2508
2509 ac_build_intrinsic(ctx, "llvm.amdgcn.exp.f32",
2510 ctx->voidt, args, 8, 0);
2511 }
2512 }
2513
2514 void ac_build_export_null(struct ac_llvm_context *ctx)
2515 {
2516 struct ac_export_args args;
2517
2518 args.enabled_channels = 0x0; /* enabled channels */
2519 args.valid_mask = 1; /* whether the EXEC mask is valid */
2520 args.done = 1; /* DONE bit */
2521 args.target = V_008DFC_SQ_EXP_NULL;
2522 args.compr = 0; /* COMPR flag (0 = 32-bit export) */
2523 args.out[0] = LLVMGetUndef(ctx->f32); /* R */
2524 args.out[1] = LLVMGetUndef(ctx->f32); /* G */
2525 args.out[2] = LLVMGetUndef(ctx->f32); /* B */
2526 args.out[3] = LLVMGetUndef(ctx->f32); /* A */
2527
2528 ac_build_export(ctx, &args);
2529 }
2530
2531 static unsigned ac_num_coords(enum ac_image_dim dim)
2532 {
2533 switch (dim) {
2534 case ac_image_1d:
2535 return 1;
2536 case ac_image_2d:
2537 case ac_image_1darray:
2538 return 2;
2539 case ac_image_3d:
2540 case ac_image_cube:
2541 case ac_image_2darray:
2542 case ac_image_2dmsaa:
2543 return 3;
2544 case ac_image_2darraymsaa:
2545 return 4;
2546 default:
2547 unreachable("ac_num_coords: bad dim");
2548 }
2549 }
2550
2551 static unsigned ac_num_derivs(enum ac_image_dim dim)
2552 {
2553 switch (dim) {
2554 case ac_image_1d:
2555 case ac_image_1darray:
2556 return 2;
2557 case ac_image_2d:
2558 case ac_image_2darray:
2559 case ac_image_cube:
2560 return 4;
2561 case ac_image_3d:
2562 return 6;
2563 case ac_image_2dmsaa:
2564 case ac_image_2darraymsaa:
2565 default:
2566 unreachable("derivatives not supported");
2567 }
2568 }
2569
2570 static const char *get_atomic_name(enum ac_atomic_op op)
2571 {
2572 switch (op) {
2573 case ac_atomic_swap: return "swap";
2574 case ac_atomic_add: return "add";
2575 case ac_atomic_sub: return "sub";
2576 case ac_atomic_smin: return "smin";
2577 case ac_atomic_umin: return "umin";
2578 case ac_atomic_smax: return "smax";
2579 case ac_atomic_umax: return "umax";
2580 case ac_atomic_and: return "and";
2581 case ac_atomic_or: return "or";
2582 case ac_atomic_xor: return "xor";
2583 }
2584 unreachable("bad atomic op");
2585 }
2586
2587 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
2588 struct ac_image_args *a)
2589 {
2590 const char *overload[3] = { "", "", "" };
2591 unsigned num_overloads = 0;
2592 LLVMValueRef args[18];
2593 unsigned num_args = 0;
2594 enum ac_image_dim dim = a->dim;
2595
2596 assert(!a->lod || a->lod == ctx->i32_0 || a->lod == ctx->f32_0 ||
2597 !a->level_zero);
2598 assert((a->opcode != ac_image_get_resinfo && a->opcode != ac_image_load_mip &&
2599 a->opcode != ac_image_store_mip) ||
2600 a->lod);
2601 assert(a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
2602 (!a->compare && !a->offset));
2603 assert((a->opcode == ac_image_sample || a->opcode == ac_image_gather4 ||
2604 a->opcode == ac_image_get_lod) ||
2605 !a->bias);
2606 assert((a->bias ? 1 : 0) +
2607 (a->lod ? 1 : 0) +
2608 (a->level_zero ? 1 : 0) +
2609 (a->derivs[0] ? 1 : 0) <= 1);
2610
2611 if (a->opcode == ac_image_get_lod) {
2612 switch (dim) {
2613 case ac_image_1darray:
2614 dim = ac_image_1d;
2615 break;
2616 case ac_image_2darray:
2617 case ac_image_cube:
2618 dim = ac_image_2d;
2619 break;
2620 default:
2621 break;
2622 }
2623 }
2624
2625 bool sample = a->opcode == ac_image_sample ||
2626 a->opcode == ac_image_gather4 ||
2627 a->opcode == ac_image_get_lod;
2628 bool atomic = a->opcode == ac_image_atomic ||
2629 a->opcode == ac_image_atomic_cmpswap;
2630 bool load = a->opcode == ac_image_sample ||
2631 a->opcode == ac_image_gather4 ||
2632 a->opcode == ac_image_load ||
2633 a->opcode == ac_image_load_mip;
2634 LLVMTypeRef coord_type = sample ? ctx->f32 : ctx->i32;
2635
2636 if (atomic || a->opcode == ac_image_store || a->opcode == ac_image_store_mip) {
2637 args[num_args++] = a->data[0];
2638 if (a->opcode == ac_image_atomic_cmpswap)
2639 args[num_args++] = a->data[1];
2640 }
2641
2642 if (!atomic)
2643 args[num_args++] = LLVMConstInt(ctx->i32, a->dmask, false);
2644
2645 if (a->offset)
2646 args[num_args++] = ac_to_integer(ctx, a->offset);
2647 if (a->bias) {
2648 args[num_args++] = ac_to_float(ctx, a->bias);
2649 overload[num_overloads++] = ".f32";
2650 }
2651 if (a->compare)
2652 args[num_args++] = ac_to_float(ctx, a->compare);
2653 if (a->derivs[0]) {
2654 unsigned count = ac_num_derivs(dim);
2655 for (unsigned i = 0; i < count; ++i)
2656 args[num_args++] = ac_to_float(ctx, a->derivs[i]);
2657 overload[num_overloads++] = ".f32";
2658 }
2659 unsigned num_coords =
2660 a->opcode != ac_image_get_resinfo ? ac_num_coords(dim) : 0;
2661 for (unsigned i = 0; i < num_coords; ++i)
2662 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->coords[i], coord_type, "");
2663 if (a->lod)
2664 args[num_args++] = LLVMBuildBitCast(ctx->builder, a->lod, coord_type, "");
2665 overload[num_overloads++] = sample ? ".f32" : ".i32";
2666
2667 args[num_args++] = a->resource;
2668 if (sample) {
2669 args[num_args++] = a->sampler;
2670 args[num_args++] = LLVMConstInt(ctx->i1, a->unorm, false);
2671 }
2672
2673 args[num_args++] = ctx->i32_0; /* texfailctrl */
2674 args[num_args++] = LLVMConstInt(ctx->i32,
2675 load ? get_load_cache_policy(ctx, a->cache_policy) :
2676 a->cache_policy, false);
2677
2678 const char *name;
2679 const char *atomic_subop = "";
2680 switch (a->opcode) {
2681 case ac_image_sample: name = "sample"; break;
2682 case ac_image_gather4: name = "gather4"; break;
2683 case ac_image_load: name = "load"; break;
2684 case ac_image_load_mip: name = "load.mip"; break;
2685 case ac_image_store: name = "store"; break;
2686 case ac_image_store_mip: name = "store.mip"; break;
2687 case ac_image_atomic:
2688 name = "atomic.";
2689 atomic_subop = get_atomic_name(a->atomic);
2690 break;
2691 case ac_image_atomic_cmpswap:
2692 name = "atomic.";
2693 atomic_subop = "cmpswap";
2694 break;
2695 case ac_image_get_lod: name = "getlod"; break;
2696 case ac_image_get_resinfo: name = "getresinfo"; break;
2697 default: unreachable("invalid image opcode");
2698 }
2699
2700 const char *dimname;
2701 switch (dim) {
2702 case ac_image_1d: dimname = "1d"; break;
2703 case ac_image_2d: dimname = "2d"; break;
2704 case ac_image_3d: dimname = "3d"; break;
2705 case ac_image_cube: dimname = "cube"; break;
2706 case ac_image_1darray: dimname = "1darray"; break;
2707 case ac_image_2darray: dimname = "2darray"; break;
2708 case ac_image_2dmsaa: dimname = "2dmsaa"; break;
2709 case ac_image_2darraymsaa: dimname = "2darraymsaa"; break;
2710 default: unreachable("invalid dim");
2711 }
2712
2713 bool lod_suffix =
2714 a->lod && (a->opcode == ac_image_sample || a->opcode == ac_image_gather4);
2715 char intr_name[96];
2716 snprintf(intr_name, sizeof(intr_name),
2717 "llvm.amdgcn.image.%s%s" /* base name */
2718 "%s%s%s" /* sample/gather modifiers */
2719 ".%s.%s%s%s%s", /* dimension and type overloads */
2720 name, atomic_subop,
2721 a->compare ? ".c" : "",
2722 a->bias ? ".b" :
2723 lod_suffix ? ".l" :
2724 a->derivs[0] ? ".d" :
2725 a->level_zero ? ".lz" : "",
2726 a->offset ? ".o" : "",
2727 dimname,
2728 atomic ? "i32" : "v4f32",
2729 overload[0], overload[1], overload[2]);
2730
2731 LLVMTypeRef retty;
2732 if (atomic)
2733 retty = ctx->i32;
2734 else if (a->opcode == ac_image_store || a->opcode == ac_image_store_mip)
2735 retty = ctx->voidt;
2736 else
2737 retty = ctx->v4f32;
2738
2739 LLVMValueRef result =
2740 ac_build_intrinsic(ctx, intr_name, retty, args, num_args,
2741 a->attributes);
2742 if (!sample && retty == ctx->v4f32) {
2743 result = LLVMBuildBitCast(ctx->builder, result,
2744 ctx->v4i32, "");
2745 }
2746 return result;
2747 }
2748
2749 LLVMValueRef ac_build_cvt_pkrtz_f16(struct ac_llvm_context *ctx,
2750 LLVMValueRef args[2])
2751 {
2752 LLVMTypeRef v2f16 =
2753 LLVMVectorType(LLVMHalfTypeInContext(ctx->context), 2);
2754
2755 return ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pkrtz", v2f16,
2756 args, 2, AC_FUNC_ATTR_READNONE);
2757 }
2758
2759 LLVMValueRef ac_build_cvt_pknorm_i16(struct ac_llvm_context *ctx,
2760 LLVMValueRef args[2])
2761 {
2762 LLVMValueRef res =
2763 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.i16",
2764 ctx->v2i16, args, 2,
2765 AC_FUNC_ATTR_READNONE);
2766 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2767 }
2768
2769 LLVMValueRef ac_build_cvt_pknorm_u16(struct ac_llvm_context *ctx,
2770 LLVMValueRef args[2])
2771 {
2772 LLVMValueRef res =
2773 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pknorm.u16",
2774 ctx->v2i16, args, 2,
2775 AC_FUNC_ATTR_READNONE);
2776 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2777 }
2778
2779 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2780 LLVMValueRef ac_build_cvt_pk_i16(struct ac_llvm_context *ctx,
2781 LLVMValueRef args[2], unsigned bits, bool hi)
2782 {
2783 assert(bits == 8 || bits == 10 || bits == 16);
2784
2785 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2786 bits == 8 ? 127 : bits == 10 ? 511 : 32767, 0);
2787 LLVMValueRef min_rgb = LLVMConstInt(ctx->i32,
2788 bits == 8 ? -128 : bits == 10 ? -512 : -32768, 0);
2789 LLVMValueRef max_alpha =
2790 bits != 10 ? max_rgb : ctx->i32_1;
2791 LLVMValueRef min_alpha =
2792 bits != 10 ? min_rgb : LLVMConstInt(ctx->i32, -2, 0);
2793
2794 /* Clamp. */
2795 if (bits != 16) {
2796 for (int i = 0; i < 2; i++) {
2797 bool alpha = hi && i == 1;
2798 args[i] = ac_build_imin(ctx, args[i],
2799 alpha ? max_alpha : max_rgb);
2800 args[i] = ac_build_imax(ctx, args[i],
2801 alpha ? min_alpha : min_rgb);
2802 }
2803 }
2804
2805 LLVMValueRef res =
2806 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.i16",
2807 ctx->v2i16, args, 2,
2808 AC_FUNC_ATTR_READNONE);
2809 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2810 }
2811
2812 /* The 8-bit and 10-bit clamping is for HW workarounds. */
2813 LLVMValueRef ac_build_cvt_pk_u16(struct ac_llvm_context *ctx,
2814 LLVMValueRef args[2], unsigned bits, bool hi)
2815 {
2816 assert(bits == 8 || bits == 10 || bits == 16);
2817
2818 LLVMValueRef max_rgb = LLVMConstInt(ctx->i32,
2819 bits == 8 ? 255 : bits == 10 ? 1023 : 65535, 0);
2820 LLVMValueRef max_alpha =
2821 bits != 10 ? max_rgb : LLVMConstInt(ctx->i32, 3, 0);
2822
2823 /* Clamp. */
2824 if (bits != 16) {
2825 for (int i = 0; i < 2; i++) {
2826 bool alpha = hi && i == 1;
2827 args[i] = ac_build_umin(ctx, args[i],
2828 alpha ? max_alpha : max_rgb);
2829 }
2830 }
2831
2832 LLVMValueRef res =
2833 ac_build_intrinsic(ctx, "llvm.amdgcn.cvt.pk.u16",
2834 ctx->v2i16, args, 2,
2835 AC_FUNC_ATTR_READNONE);
2836 return LLVMBuildBitCast(ctx->builder, res, ctx->i32, "");
2837 }
2838
2839 LLVMValueRef ac_build_wqm_vote(struct ac_llvm_context *ctx, LLVMValueRef i1)
2840 {
2841 return ac_build_intrinsic(ctx, "llvm.amdgcn.wqm.vote", ctx->i1,
2842 &i1, 1, AC_FUNC_ATTR_READNONE);
2843 }
2844
2845 void ac_build_kill_if_false(struct ac_llvm_context *ctx, LLVMValueRef i1)
2846 {
2847 ac_build_intrinsic(ctx, "llvm.amdgcn.kill", ctx->voidt,
2848 &i1, 1, 0);
2849 }
2850
2851 LLVMValueRef ac_build_bfe(struct ac_llvm_context *ctx, LLVMValueRef input,
2852 LLVMValueRef offset, LLVMValueRef width,
2853 bool is_signed)
2854 {
2855 LLVMValueRef args[] = {
2856 input,
2857 offset,
2858 width,
2859 };
2860
2861 LLVMValueRef result = ac_build_intrinsic(ctx,
2862 is_signed ? "llvm.amdgcn.sbfe.i32" :
2863 "llvm.amdgcn.ubfe.i32",
2864 ctx->i32, args, 3,
2865 AC_FUNC_ATTR_READNONE);
2866
2867 if (HAVE_LLVM < 0x0800) {
2868 /* FIXME: LLVM 7+ returns incorrect result when count is 0.
2869 * https://bugs.freedesktop.org/show_bug.cgi?id=107276
2870 */
2871 LLVMValueRef zero = ctx->i32_0;
2872 LLVMValueRef icond = LLVMBuildICmp(ctx->builder, LLVMIntEQ, width, zero, "");
2873 result = LLVMBuildSelect(ctx->builder, icond, zero, result, "");
2874 }
2875
2876 return result;
2877 }
2878
2879 LLVMValueRef ac_build_imad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2880 LLVMValueRef s1, LLVMValueRef s2)
2881 {
2882 return LLVMBuildAdd(ctx->builder,
2883 LLVMBuildMul(ctx->builder, s0, s1, ""), s2, "");
2884 }
2885
2886 LLVMValueRef ac_build_fmad(struct ac_llvm_context *ctx, LLVMValueRef s0,
2887 LLVMValueRef s1, LLVMValueRef s2)
2888 {
2889 return LLVMBuildFAdd(ctx->builder,
2890 LLVMBuildFMul(ctx->builder, s0, s1, ""), s2, "");
2891 }
2892
2893 void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags)
2894 {
2895 if (!wait_flags)
2896 return;
2897
2898 unsigned lgkmcnt = 63;
2899 unsigned vmcnt = ctx->chip_class >= GFX9 ? 63 : 15;
2900 unsigned vscnt = 63;
2901
2902 if (wait_flags & AC_WAIT_LGKM)
2903 lgkmcnt = 0;
2904 if (wait_flags & AC_WAIT_VLOAD)
2905 vmcnt = 0;
2906
2907 if (wait_flags & AC_WAIT_VSTORE) {
2908 if (ctx->chip_class >= GFX10)
2909 vscnt = 0;
2910 else
2911 vmcnt = 0;
2912 }
2913
2914 /* There is no intrinsic for vscnt(0), so use a fence. */
2915 if ((wait_flags & AC_WAIT_LGKM &&
2916 wait_flags & AC_WAIT_VLOAD &&
2917 wait_flags & AC_WAIT_VSTORE) ||
2918 vscnt == 0) {
2919 LLVMBuildFence(ctx->builder, LLVMAtomicOrderingRelease, false, "");
2920 return;
2921 }
2922
2923 unsigned simm16 = (lgkmcnt << 8) |
2924 (7 << 4) | /* expcnt */
2925 (vmcnt & 0xf) |
2926 ((vmcnt >> 4) << 14);
2927
2928 LLVMValueRef args[1] = {
2929 LLVMConstInt(ctx->i32, simm16, false),
2930 };
2931 ac_build_intrinsic(ctx, "llvm.amdgcn.s.waitcnt",
2932 ctx->voidt, args, 1, 0);
2933 }
2934
2935 LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
2936 LLVMValueRef src1, LLVMValueRef src2,
2937 unsigned bitsize)
2938 {
2939 LLVMTypeRef type;
2940 char *intr;
2941
2942 if (bitsize == 16) {
2943 intr = "llvm.amdgcn.fmed3.f16";
2944 type = ctx->f16;
2945 } else if (bitsize == 32) {
2946 intr = "llvm.amdgcn.fmed3.f32";
2947 type = ctx->f32;
2948 } else {
2949 intr = "llvm.amdgcn.fmed3.f64";
2950 type = ctx->f64;
2951 }
2952
2953 LLVMValueRef params[] = {
2954 src0,
2955 src1,
2956 src2,
2957 };
2958 return ac_build_intrinsic(ctx, intr, type, params, 3,
2959 AC_FUNC_ATTR_READNONE);
2960 }
2961
2962 LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
2963 unsigned bitsize)
2964 {
2965 LLVMTypeRef type;
2966 char *intr;
2967
2968 if (bitsize == 16) {
2969 intr = "llvm.amdgcn.fract.f16";
2970 type = ctx->f16;
2971 } else if (bitsize == 32) {
2972 intr = "llvm.amdgcn.fract.f32";
2973 type = ctx->f32;
2974 } else {
2975 intr = "llvm.amdgcn.fract.f64";
2976 type = ctx->f64;
2977 }
2978
2979 LLVMValueRef params[] = {
2980 src0,
2981 };
2982 return ac_build_intrinsic(ctx, intr, type, params, 1,
2983 AC_FUNC_ATTR_READNONE);
2984 }
2985
2986 LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
2987 unsigned bitsize)
2988 {
2989 LLVMTypeRef type = LLVMIntTypeInContext(ctx->context, bitsize);
2990 LLVMValueRef zero = LLVMConstInt(type, 0, false);
2991 LLVMValueRef one = LLVMConstInt(type, 1, false);
2992
2993 LLVMValueRef cmp, val;
2994 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGT, src0, zero, "");
2995 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
2996 cmp = LLVMBuildICmp(ctx->builder, LLVMIntSGE, val, zero, "");
2997 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstInt(type, -1, true), "");
2998 return val;
2999 }
3000
3001 LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
3002 unsigned bitsize)
3003 {
3004 LLVMValueRef cmp, val, zero, one;
3005 LLVMTypeRef type;
3006
3007 if (bitsize == 16) {
3008 type = ctx->f16;
3009 zero = ctx->f16_0;
3010 one = ctx->f16_1;
3011 } else if (bitsize == 32) {
3012 type = ctx->f32;
3013 zero = ctx->f32_0;
3014 one = ctx->f32_1;
3015 } else {
3016 type = ctx->f64;
3017 zero = ctx->f64_0;
3018 one = ctx->f64_1;
3019 }
3020
3021 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGT, src0, zero, "");
3022 val = LLVMBuildSelect(ctx->builder, cmp, one, src0, "");
3023 cmp = LLVMBuildFCmp(ctx->builder, LLVMRealOGE, val, zero, "");
3024 val = LLVMBuildSelect(ctx->builder, cmp, val, LLVMConstReal(type, -1.0), "");
3025 return val;
3026 }
3027
3028 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0)
3029 {
3030 LLVMValueRef result;
3031 unsigned bitsize;
3032
3033 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
3034
3035 switch (bitsize) {
3036 case 64:
3037 result = ac_build_intrinsic(ctx, "llvm.ctpop.i64", ctx->i64,
3038 (LLVMValueRef []) { src0 }, 1,
3039 AC_FUNC_ATTR_READNONE);
3040
3041 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
3042 break;
3043 case 32:
3044 result = ac_build_intrinsic(ctx, "llvm.ctpop.i32", ctx->i32,
3045 (LLVMValueRef []) { src0 }, 1,
3046 AC_FUNC_ATTR_READNONE);
3047 break;
3048 case 16:
3049 result = ac_build_intrinsic(ctx, "llvm.ctpop.i16", ctx->i16,
3050 (LLVMValueRef []) { src0 }, 1,
3051 AC_FUNC_ATTR_READNONE);
3052
3053 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
3054 break;
3055 case 8:
3056 result = ac_build_intrinsic(ctx, "llvm.ctpop.i8", ctx->i8,
3057 (LLVMValueRef []) { src0 }, 1,
3058 AC_FUNC_ATTR_READNONE);
3059
3060 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
3061 break;
3062 default:
3063 unreachable(!"invalid bitsize");
3064 break;
3065 }
3066
3067 return result;
3068 }
3069
3070 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
3071 LLVMValueRef src0)
3072 {
3073 LLVMValueRef result;
3074 unsigned bitsize;
3075
3076 bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
3077
3078 switch (bitsize) {
3079 case 64:
3080 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i64", ctx->i64,
3081 (LLVMValueRef []) { src0 }, 1,
3082 AC_FUNC_ATTR_READNONE);
3083
3084 result = LLVMBuildTrunc(ctx->builder, result, ctx->i32, "");
3085 break;
3086 case 32:
3087 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i32", ctx->i32,
3088 (LLVMValueRef []) { src0 }, 1,
3089 AC_FUNC_ATTR_READNONE);
3090 break;
3091 case 16:
3092 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i16", ctx->i16,
3093 (LLVMValueRef []) { src0 }, 1,
3094 AC_FUNC_ATTR_READNONE);
3095
3096 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
3097 break;
3098 case 8:
3099 result = ac_build_intrinsic(ctx, "llvm.bitreverse.i8", ctx->i8,
3100 (LLVMValueRef []) { src0 }, 1,
3101 AC_FUNC_ATTR_READNONE);
3102
3103 result = LLVMBuildZExt(ctx->builder, result, ctx->i32, "");
3104 break;
3105 default:
3106 unreachable(!"invalid bitsize");
3107 break;
3108 }
3109
3110 return result;
3111 }
3112
3113 #define AC_EXP_TARGET 0
3114 #define AC_EXP_ENABLED_CHANNELS 1
3115 #define AC_EXP_OUT0 2
3116
3117 enum ac_ir_type {
3118 AC_IR_UNDEF,
3119 AC_IR_CONST,
3120 AC_IR_VALUE,
3121 };
3122
3123 struct ac_vs_exp_chan
3124 {
3125 LLVMValueRef value;
3126 float const_float;
3127 enum ac_ir_type type;
3128 };
3129
3130 struct ac_vs_exp_inst {
3131 unsigned offset;
3132 LLVMValueRef inst;
3133 struct ac_vs_exp_chan chan[4];
3134 };
3135
3136 struct ac_vs_exports {
3137 unsigned num;
3138 struct ac_vs_exp_inst exp[VARYING_SLOT_MAX];
3139 };
3140
3141 /* Return true if the PARAM export has been eliminated. */
3142 static bool ac_eliminate_const_output(uint8_t *vs_output_param_offset,
3143 uint32_t num_outputs,
3144 struct ac_vs_exp_inst *exp)
3145 {
3146 unsigned i, default_val; /* SPI_PS_INPUT_CNTL_i.DEFAULT_VAL */
3147 bool is_zero[4] = {}, is_one[4] = {};
3148
3149 for (i = 0; i < 4; i++) {
3150 /* It's a constant expression. Undef outputs are eliminated too. */
3151 if (exp->chan[i].type == AC_IR_UNDEF) {
3152 is_zero[i] = true;
3153 is_one[i] = true;
3154 } else if (exp->chan[i].type == AC_IR_CONST) {
3155 if (exp->chan[i].const_float == 0)
3156 is_zero[i] = true;
3157 else if (exp->chan[i].const_float == 1)
3158 is_one[i] = true;
3159 else
3160 return false; /* other constant */
3161 } else
3162 return false;
3163 }
3164
3165 /* Only certain combinations of 0 and 1 can be eliminated. */
3166 if (is_zero[0] && is_zero[1] && is_zero[2])
3167 default_val = is_zero[3] ? 0 : 1;
3168 else if (is_one[0] && is_one[1] && is_one[2])
3169 default_val = is_zero[3] ? 2 : 3;
3170 else
3171 return false;
3172
3173 /* The PARAM export can be represented as DEFAULT_VAL. Kill it. */
3174 LLVMInstructionEraseFromParent(exp->inst);
3175
3176 /* Change OFFSET to DEFAULT_VAL. */
3177 for (i = 0; i < num_outputs; i++) {
3178 if (vs_output_param_offset[i] == exp->offset) {
3179 vs_output_param_offset[i] =
3180 AC_EXP_PARAM_DEFAULT_VAL_0000 + default_val;
3181 break;
3182 }
3183 }
3184 return true;
3185 }
3186
3187 static bool ac_eliminate_duplicated_output(struct ac_llvm_context *ctx,
3188 uint8_t *vs_output_param_offset,
3189 uint32_t num_outputs,
3190 struct ac_vs_exports *processed,
3191 struct ac_vs_exp_inst *exp)
3192 {
3193 unsigned p, copy_back_channels = 0;
3194
3195 /* See if the output is already in the list of processed outputs.
3196 * The LLVMValueRef comparison relies on SSA.
3197 */
3198 for (p = 0; p < processed->num; p++) {
3199 bool different = false;
3200
3201 for (unsigned j = 0; j < 4; j++) {
3202 struct ac_vs_exp_chan *c1 = &processed->exp[p].chan[j];
3203 struct ac_vs_exp_chan *c2 = &exp->chan[j];
3204
3205 /* Treat undef as a match. */
3206 if (c2->type == AC_IR_UNDEF)
3207 continue;
3208
3209 /* If c1 is undef but c2 isn't, we can copy c2 to c1
3210 * and consider the instruction duplicated.
3211 */
3212 if (c1->type == AC_IR_UNDEF) {
3213 copy_back_channels |= 1 << j;
3214 continue;
3215 }
3216
3217 /* Test whether the channels are not equal. */
3218 if (c1->type != c2->type ||
3219 (c1->type == AC_IR_CONST &&
3220 c1->const_float != c2->const_float) ||
3221 (c1->type == AC_IR_VALUE &&
3222 c1->value != c2->value)) {
3223 different = true;
3224 break;
3225 }
3226 }
3227 if (!different)
3228 break;
3229
3230 copy_back_channels = 0;
3231 }
3232 if (p == processed->num)
3233 return false;
3234
3235 /* If a match was found, but the matching export has undef where the new
3236 * one has a normal value, copy the normal value to the undef channel.
3237 */
3238 struct ac_vs_exp_inst *match = &processed->exp[p];
3239
3240 /* Get current enabled channels mask. */
3241 LLVMValueRef arg = LLVMGetOperand(match->inst, AC_EXP_ENABLED_CHANNELS);
3242 unsigned enabled_channels = LLVMConstIntGetZExtValue(arg);
3243
3244 while (copy_back_channels) {
3245 unsigned chan = u_bit_scan(&copy_back_channels);
3246
3247 assert(match->chan[chan].type == AC_IR_UNDEF);
3248 LLVMSetOperand(match->inst, AC_EXP_OUT0 + chan,
3249 exp->chan[chan].value);
3250 match->chan[chan] = exp->chan[chan];
3251
3252 /* Update number of enabled channels because the original mask
3253 * is not always 0xf.
3254 */
3255 enabled_channels |= (1 << chan);
3256 LLVMSetOperand(match->inst, AC_EXP_ENABLED_CHANNELS,
3257 LLVMConstInt(ctx->i32, enabled_channels, 0));
3258 }
3259
3260 /* The PARAM export is duplicated. Kill it. */
3261 LLVMInstructionEraseFromParent(exp->inst);
3262
3263 /* Change OFFSET to the matching export. */
3264 for (unsigned i = 0; i < num_outputs; i++) {
3265 if (vs_output_param_offset[i] == exp->offset) {
3266 vs_output_param_offset[i] = match->offset;
3267 break;
3268 }
3269 }
3270 return true;
3271 }
3272
3273 void ac_optimize_vs_outputs(struct ac_llvm_context *ctx,
3274 LLVMValueRef main_fn,
3275 uint8_t *vs_output_param_offset,
3276 uint32_t num_outputs,
3277 uint8_t *num_param_exports)
3278 {
3279 LLVMBasicBlockRef bb;
3280 bool removed_any = false;
3281 struct ac_vs_exports exports;
3282
3283 exports.num = 0;
3284
3285 /* Process all LLVM instructions. */
3286 bb = LLVMGetFirstBasicBlock(main_fn);
3287 while (bb) {
3288 LLVMValueRef inst = LLVMGetFirstInstruction(bb);
3289
3290 while (inst) {
3291 LLVMValueRef cur = inst;
3292 inst = LLVMGetNextInstruction(inst);
3293 struct ac_vs_exp_inst exp;
3294
3295 if (LLVMGetInstructionOpcode(cur) != LLVMCall)
3296 continue;
3297
3298 LLVMValueRef callee = ac_llvm_get_called_value(cur);
3299
3300 if (!ac_llvm_is_function(callee))
3301 continue;
3302
3303 const char *name = LLVMGetValueName(callee);
3304 unsigned num_args = LLVMCountParams(callee);
3305
3306 /* Check if this is an export instruction. */
3307 if ((num_args != 9 && num_args != 8) ||
3308 (strcmp(name, "llvm.SI.export") &&
3309 strcmp(name, "llvm.amdgcn.exp.f32")))
3310 continue;
3311
3312 LLVMValueRef arg = LLVMGetOperand(cur, AC_EXP_TARGET);
3313 unsigned target = LLVMConstIntGetZExtValue(arg);
3314
3315 if (target < V_008DFC_SQ_EXP_PARAM)
3316 continue;
3317
3318 target -= V_008DFC_SQ_EXP_PARAM;
3319
3320 /* Parse the instruction. */
3321 memset(&exp, 0, sizeof(exp));
3322 exp.offset = target;
3323 exp.inst = cur;
3324
3325 for (unsigned i = 0; i < 4; i++) {
3326 LLVMValueRef v = LLVMGetOperand(cur, AC_EXP_OUT0 + i);
3327
3328 exp.chan[i].value = v;
3329
3330 if (LLVMIsUndef(v)) {
3331 exp.chan[i].type = AC_IR_UNDEF;
3332 } else if (LLVMIsAConstantFP(v)) {
3333 LLVMBool loses_info;
3334 exp.chan[i].type = AC_IR_CONST;
3335 exp.chan[i].const_float =
3336 LLVMConstRealGetDouble(v, &loses_info);
3337 } else {
3338 exp.chan[i].type = AC_IR_VALUE;
3339 }
3340 }
3341
3342 /* Eliminate constant and duplicated PARAM exports. */
3343 if (ac_eliminate_const_output(vs_output_param_offset,
3344 num_outputs, &exp) ||
3345 ac_eliminate_duplicated_output(ctx,
3346 vs_output_param_offset,
3347 num_outputs, &exports,
3348 &exp)) {
3349 removed_any = true;
3350 } else {
3351 exports.exp[exports.num++] = exp;
3352 }
3353 }
3354 bb = LLVMGetNextBasicBlock(bb);
3355 }
3356
3357 /* Remove holes in export memory due to removed PARAM exports.
3358 * This is done by renumbering all PARAM exports.
3359 */
3360 if (removed_any) {
3361 uint8_t old_offset[VARYING_SLOT_MAX];
3362 unsigned out, i;
3363
3364 /* Make a copy of the offsets. We need the old version while
3365 * we are modifying some of them. */
3366 memcpy(old_offset, vs_output_param_offset,
3367 sizeof(old_offset));
3368
3369 for (i = 0; i < exports.num; i++) {
3370 unsigned offset = exports.exp[i].offset;
3371
3372 /* Update vs_output_param_offset. Multiple outputs can
3373 * have the same offset.
3374 */
3375 for (out = 0; out < num_outputs; out++) {
3376 if (old_offset[out] == offset)
3377 vs_output_param_offset[out] = i;
3378 }
3379
3380 /* Change the PARAM offset in the instruction. */
3381 LLVMSetOperand(exports.exp[i].inst, AC_EXP_TARGET,
3382 LLVMConstInt(ctx->i32,
3383 V_008DFC_SQ_EXP_PARAM + i, 0));
3384 }
3385 *num_param_exports = exports.num;
3386 }
3387 }
3388
3389 void ac_init_exec_full_mask(struct ac_llvm_context *ctx)
3390 {
3391 LLVMValueRef full_mask = LLVMConstInt(ctx->i64, ~0ull, 0);
3392 ac_build_intrinsic(ctx,
3393 "llvm.amdgcn.init.exec", ctx->voidt,
3394 &full_mask, 1, AC_FUNC_ATTR_CONVERGENT);
3395 }
3396
3397 void ac_declare_lds_as_pointer(struct ac_llvm_context *ctx)
3398 {
3399 unsigned lds_size = ctx->chip_class >= GFX7 ? 65536 : 32768;
3400 ctx->lds = LLVMBuildIntToPtr(ctx->builder, ctx->i32_0,
3401 LLVMPointerType(LLVMArrayType(ctx->i32, lds_size / 4), AC_ADDR_SPACE_LDS),
3402 "lds");
3403 }
3404
3405 LLVMValueRef ac_lds_load(struct ac_llvm_context *ctx,
3406 LLVMValueRef dw_addr)
3407 {
3408 return LLVMBuildLoad(ctx->builder, ac_build_gep0(ctx, ctx->lds, dw_addr), "");
3409 }
3410
3411 void ac_lds_store(struct ac_llvm_context *ctx,
3412 LLVMValueRef dw_addr,
3413 LLVMValueRef value)
3414 {
3415 value = ac_to_integer(ctx, value);
3416 ac_build_indexed_store(ctx, ctx->lds,
3417 dw_addr, value);
3418 }
3419
3420 LLVMValueRef ac_find_lsb(struct ac_llvm_context *ctx,
3421 LLVMTypeRef dst_type,
3422 LLVMValueRef src0)
3423 {
3424 unsigned src0_bitsize = ac_get_elem_bits(ctx, LLVMTypeOf(src0));
3425 const char *intrin_name;
3426 LLVMTypeRef type;
3427 LLVMValueRef zero;
3428
3429 switch (src0_bitsize) {
3430 case 64:
3431 intrin_name = "llvm.cttz.i64";
3432 type = ctx->i64;
3433 zero = ctx->i64_0;
3434 break;
3435 case 32:
3436 intrin_name = "llvm.cttz.i32";
3437 type = ctx->i32;
3438 zero = ctx->i32_0;
3439 break;
3440 case 16:
3441 intrin_name = "llvm.cttz.i16";
3442 type = ctx->i16;
3443 zero = ctx->i16_0;
3444 break;
3445 case 8:
3446 intrin_name = "llvm.cttz.i8";
3447 type = ctx->i8;
3448 zero = ctx->i8_0;
3449 break;
3450 default:
3451 unreachable(!"invalid bitsize");
3452 }
3453
3454 LLVMValueRef params[2] = {
3455 src0,
3456
3457 /* The value of 1 means that ffs(x=0) = undef, so LLVM won't
3458 * add special code to check for x=0. The reason is that
3459 * the LLVM behavior for x=0 is different from what we
3460 * need here. However, LLVM also assumes that ffs(x) is
3461 * in [0, 31], but GLSL expects that ffs(0) = -1, so
3462 * a conditional assignment to handle 0 is still required.
3463 *
3464 * The hardware already implements the correct behavior.
3465 */
3466 ctx->i1true,
3467 };
3468
3469 LLVMValueRef lsb = ac_build_intrinsic(ctx, intrin_name, type,
3470 params, 2,
3471 AC_FUNC_ATTR_READNONE);
3472
3473 if (src0_bitsize == 64) {
3474 lsb = LLVMBuildTrunc(ctx->builder, lsb, ctx->i32, "");
3475 } else if (src0_bitsize < 32) {
3476 lsb = LLVMBuildSExt(ctx->builder, lsb, ctx->i32, "");
3477 }
3478
3479 /* TODO: We need an intrinsic to skip this conditional. */
3480 /* Check for zero: */
3481 return LLVMBuildSelect(ctx->builder, LLVMBuildICmp(ctx->builder,
3482 LLVMIntEQ, src0,
3483 zero, ""),
3484 LLVMConstInt(ctx->i32, -1, 0), lsb, "");
3485 }
3486
3487 LLVMTypeRef ac_array_in_const_addr_space(LLVMTypeRef elem_type)
3488 {
3489 return LLVMPointerType(elem_type, AC_ADDR_SPACE_CONST);
3490 }
3491
3492 LLVMTypeRef ac_array_in_const32_addr_space(LLVMTypeRef elem_type)
3493 {
3494 return LLVMPointerType(elem_type, AC_ADDR_SPACE_CONST_32BIT);
3495 }
3496
3497 static struct ac_llvm_flow *
3498 get_current_flow(struct ac_llvm_context *ctx)
3499 {
3500 if (ctx->flow->depth > 0)
3501 return &ctx->flow->stack[ctx->flow->depth - 1];
3502 return NULL;
3503 }
3504
3505 static struct ac_llvm_flow *
3506 get_innermost_loop(struct ac_llvm_context *ctx)
3507 {
3508 for (unsigned i = ctx->flow->depth; i > 0; --i) {
3509 if (ctx->flow->stack[i - 1].loop_entry_block)
3510 return &ctx->flow->stack[i - 1];
3511 }
3512 return NULL;
3513 }
3514
3515 static struct ac_llvm_flow *
3516 push_flow(struct ac_llvm_context *ctx)
3517 {
3518 struct ac_llvm_flow *flow;
3519
3520 if (ctx->flow->depth >= ctx->flow->depth_max) {
3521 unsigned new_max = MAX2(ctx->flow->depth << 1,
3522 AC_LLVM_INITIAL_CF_DEPTH);
3523
3524 ctx->flow->stack = realloc(ctx->flow->stack, new_max * sizeof(*ctx->flow->stack));
3525 ctx->flow->depth_max = new_max;
3526 }
3527
3528 flow = &ctx->flow->stack[ctx->flow->depth];
3529 ctx->flow->depth++;
3530
3531 flow->next_block = NULL;
3532 flow->loop_entry_block = NULL;
3533 return flow;
3534 }
3535
3536 static void set_basicblock_name(LLVMBasicBlockRef bb, const char *base,
3537 int label_id)
3538 {
3539 char buf[32];
3540 snprintf(buf, sizeof(buf), "%s%d", base, label_id);
3541 LLVMSetValueName(LLVMBasicBlockAsValue(bb), buf);
3542 }
3543
3544 /* Append a basic block at the level of the parent flow.
3545 */
3546 static LLVMBasicBlockRef append_basic_block(struct ac_llvm_context *ctx,
3547 const char *name)
3548 {
3549 assert(ctx->flow->depth >= 1);
3550
3551 if (ctx->flow->depth >= 2) {
3552 struct ac_llvm_flow *flow = &ctx->flow->stack[ctx->flow->depth - 2];
3553
3554 return LLVMInsertBasicBlockInContext(ctx->context,
3555 flow->next_block, name);
3556 }
3557
3558 LLVMValueRef main_fn =
3559 LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx->builder));
3560 return LLVMAppendBasicBlockInContext(ctx->context, main_fn, name);
3561 }
3562
3563 /* Emit a branch to the given default target for the current block if
3564 * applicable -- that is, if the current block does not already contain a
3565 * branch from a break or continue.
3566 */
3567 static void emit_default_branch(LLVMBuilderRef builder,
3568 LLVMBasicBlockRef target)
3569 {
3570 if (!LLVMGetBasicBlockTerminator(LLVMGetInsertBlock(builder)))
3571 LLVMBuildBr(builder, target);
3572 }
3573
3574 void ac_build_bgnloop(struct ac_llvm_context *ctx, int label_id)
3575 {
3576 struct ac_llvm_flow *flow = push_flow(ctx);
3577 flow->loop_entry_block = append_basic_block(ctx, "LOOP");
3578 flow->next_block = append_basic_block(ctx, "ENDLOOP");
3579 set_basicblock_name(flow->loop_entry_block, "loop", label_id);
3580 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
3581 LLVMPositionBuilderAtEnd(ctx->builder, flow->loop_entry_block);
3582 }
3583
3584 void ac_build_break(struct ac_llvm_context *ctx)
3585 {
3586 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
3587 LLVMBuildBr(ctx->builder, flow->next_block);
3588 }
3589
3590 void ac_build_continue(struct ac_llvm_context *ctx)
3591 {
3592 struct ac_llvm_flow *flow = get_innermost_loop(ctx);
3593 LLVMBuildBr(ctx->builder, flow->loop_entry_block);
3594 }
3595
3596 void ac_build_else(struct ac_llvm_context *ctx, int label_id)
3597 {
3598 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
3599 LLVMBasicBlockRef endif_block;
3600
3601 assert(!current_branch->loop_entry_block);
3602
3603 endif_block = append_basic_block(ctx, "ENDIF");
3604 emit_default_branch(ctx->builder, endif_block);
3605
3606 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
3607 set_basicblock_name(current_branch->next_block, "else", label_id);
3608
3609 current_branch->next_block = endif_block;
3610 }
3611
3612 void ac_build_endif(struct ac_llvm_context *ctx, int label_id)
3613 {
3614 struct ac_llvm_flow *current_branch = get_current_flow(ctx);
3615
3616 assert(!current_branch->loop_entry_block);
3617
3618 emit_default_branch(ctx->builder, current_branch->next_block);
3619 LLVMPositionBuilderAtEnd(ctx->builder, current_branch->next_block);
3620 set_basicblock_name(current_branch->next_block, "endif", label_id);
3621
3622 ctx->flow->depth--;
3623 }
3624
3625 void ac_build_endloop(struct ac_llvm_context *ctx, int label_id)
3626 {
3627 struct ac_llvm_flow *current_loop = get_current_flow(ctx);
3628
3629 assert(current_loop->loop_entry_block);
3630
3631 emit_default_branch(ctx->builder, current_loop->loop_entry_block);
3632
3633 LLVMPositionBuilderAtEnd(ctx->builder, current_loop->next_block);
3634 set_basicblock_name(current_loop->next_block, "endloop", label_id);
3635 ctx->flow->depth--;
3636 }
3637
3638 void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id)
3639 {
3640 struct ac_llvm_flow *flow = push_flow(ctx);
3641 LLVMBasicBlockRef if_block;
3642
3643 if_block = append_basic_block(ctx, "IF");
3644 flow->next_block = append_basic_block(ctx, "ELSE");
3645 set_basicblock_name(if_block, "if", label_id);
3646 LLVMBuildCondBr(ctx->builder, cond, if_block, flow->next_block);
3647 LLVMPositionBuilderAtEnd(ctx->builder, if_block);
3648 }
3649
3650 void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
3651 int label_id)
3652 {
3653 LLVMValueRef cond = LLVMBuildFCmp(ctx->builder, LLVMRealUNE,
3654 value, ctx->f32_0, "");
3655 ac_build_ifcc(ctx, cond, label_id);
3656 }
3657
3658 void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
3659 int label_id)
3660 {
3661 LLVMValueRef cond = LLVMBuildICmp(ctx->builder, LLVMIntNE,
3662 ac_to_integer(ctx, value),
3663 ctx->i32_0, "");
3664 ac_build_ifcc(ctx, cond, label_id);
3665 }
3666
3667 LLVMValueRef ac_build_alloca_undef(struct ac_llvm_context *ac, LLVMTypeRef type,
3668 const char *name)
3669 {
3670 LLVMBuilderRef builder = ac->builder;
3671 LLVMBasicBlockRef current_block = LLVMGetInsertBlock(builder);
3672 LLVMValueRef function = LLVMGetBasicBlockParent(current_block);
3673 LLVMBasicBlockRef first_block = LLVMGetEntryBasicBlock(function);
3674 LLVMValueRef first_instr = LLVMGetFirstInstruction(first_block);
3675 LLVMBuilderRef first_builder = LLVMCreateBuilderInContext(ac->context);
3676 LLVMValueRef res;
3677
3678 if (first_instr) {
3679 LLVMPositionBuilderBefore(first_builder, first_instr);
3680 } else {
3681 LLVMPositionBuilderAtEnd(first_builder, first_block);
3682 }
3683
3684 res = LLVMBuildAlloca(first_builder, type, name);
3685 LLVMDisposeBuilder(first_builder);
3686 return res;
3687 }
3688
3689 LLVMValueRef ac_build_alloca(struct ac_llvm_context *ac,
3690 LLVMTypeRef type, const char *name)
3691 {
3692 LLVMValueRef ptr = ac_build_alloca_undef(ac, type, name);
3693 LLVMBuildStore(ac->builder, LLVMConstNull(type), ptr);
3694 return ptr;
3695 }
3696
3697 LLVMValueRef ac_cast_ptr(struct ac_llvm_context *ctx, LLVMValueRef ptr,
3698 LLVMTypeRef type)
3699 {
3700 int addr_space = LLVMGetPointerAddressSpace(LLVMTypeOf(ptr));
3701 return LLVMBuildBitCast(ctx->builder, ptr,
3702 LLVMPointerType(type, addr_space), "");
3703 }
3704
3705 LLVMValueRef ac_trim_vector(struct ac_llvm_context *ctx, LLVMValueRef value,
3706 unsigned count)
3707 {
3708 unsigned num_components = ac_get_llvm_num_components(value);
3709 if (count == num_components)
3710 return value;
3711
3712 LLVMValueRef masks[MAX2(count, 2)];
3713 masks[0] = ctx->i32_0;
3714 masks[1] = ctx->i32_1;
3715 for (unsigned i = 2; i < count; i++)
3716 masks[i] = LLVMConstInt(ctx->i32, i, false);
3717
3718 if (count == 1)
3719 return LLVMBuildExtractElement(ctx->builder, value, masks[0],
3720 "");
3721
3722 LLVMValueRef swizzle = LLVMConstVector(masks, count);
3723 return LLVMBuildShuffleVector(ctx->builder, value, value, swizzle, "");
3724 }
3725
3726 LLVMValueRef ac_unpack_param(struct ac_llvm_context *ctx, LLVMValueRef param,
3727 unsigned rshift, unsigned bitwidth)
3728 {
3729 LLVMValueRef value = param;
3730 if (rshift)
3731 value = LLVMBuildLShr(ctx->builder, value,
3732 LLVMConstInt(ctx->i32, rshift, false), "");
3733
3734 if (rshift + bitwidth < 32) {
3735 unsigned mask = (1 << bitwidth) - 1;
3736 value = LLVMBuildAnd(ctx->builder, value,
3737 LLVMConstInt(ctx->i32, mask, false), "");
3738 }
3739 return value;
3740 }
3741
3742 /* Adjust the sample index according to FMASK.
3743 *
3744 * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
3745 * which is the identity mapping. Each nibble says which physical sample
3746 * should be fetched to get that sample.
3747 *
3748 * For example, 0x11111100 means there are only 2 samples stored and
3749 * the second sample covers 3/4 of the pixel. When reading samples 0
3750 * and 1, return physical sample 0 (determined by the first two 0s
3751 * in FMASK), otherwise return physical sample 1.
3752 *
3753 * The sample index should be adjusted as follows:
3754 * addr[sample_index] = (fmask >> (addr[sample_index] * 4)) & 0xF;
3755 */
3756 void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
3757 LLVMValueRef *addr, bool is_array_tex)
3758 {
3759 struct ac_image_args fmask_load = {};
3760 fmask_load.opcode = ac_image_load;
3761 fmask_load.resource = fmask;
3762 fmask_load.dmask = 0xf;
3763 fmask_load.dim = is_array_tex ? ac_image_2darray : ac_image_2d;
3764 fmask_load.attributes = AC_FUNC_ATTR_READNONE;
3765
3766 fmask_load.coords[0] = addr[0];
3767 fmask_load.coords[1] = addr[1];
3768 if (is_array_tex)
3769 fmask_load.coords[2] = addr[2];
3770
3771 LLVMValueRef fmask_value = ac_build_image_opcode(ac, &fmask_load);
3772 fmask_value = LLVMBuildExtractElement(ac->builder, fmask_value,
3773 ac->i32_0, "");
3774
3775 /* Apply the formula. */
3776 unsigned sample_chan = is_array_tex ? 3 : 2;
3777 LLVMValueRef final_sample;
3778 final_sample = LLVMBuildMul(ac->builder, addr[sample_chan],
3779 LLVMConstInt(ac->i32, 4, 0), "");
3780 final_sample = LLVMBuildLShr(ac->builder, fmask_value, final_sample, "");
3781 /* Mask the sample index by 0x7, because 0x8 means an unknown value
3782 * with EQAA, so those will map to 0. */
3783 final_sample = LLVMBuildAnd(ac->builder, final_sample,
3784 LLVMConstInt(ac->i32, 0x7, 0), "");
3785
3786 /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
3787 * resource descriptor is 0 (invalid).
3788 */
3789 LLVMValueRef tmp;
3790 tmp = LLVMBuildBitCast(ac->builder, fmask, ac->v8i32, "");
3791 tmp = LLVMBuildExtractElement(ac->builder, tmp, ac->i32_1, "");
3792 tmp = LLVMBuildICmp(ac->builder, LLVMIntNE, tmp, ac->i32_0, "");
3793
3794 /* Replace the MSAA sample index. */
3795 addr[sample_chan] = LLVMBuildSelect(ac->builder, tmp, final_sample,
3796 addr[sample_chan], "");
3797 }
3798
3799 static LLVMValueRef
3800 _ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
3801 {
3802 ac_build_optimization_barrier(ctx, &src);
3803 return ac_build_intrinsic(ctx,
3804 lane == NULL ? "llvm.amdgcn.readfirstlane" : "llvm.amdgcn.readlane",
3805 LLVMTypeOf(src), (LLVMValueRef []) {
3806 src, lane },
3807 lane == NULL ? 1 : 2,
3808 AC_FUNC_ATTR_READNONE |
3809 AC_FUNC_ATTR_CONVERGENT);
3810 }
3811
3812 /**
3813 * Builds the "llvm.amdgcn.readlane" or "llvm.amdgcn.readfirstlane" intrinsic.
3814 * @param ctx
3815 * @param src
3816 * @param lane - id of the lane or NULL for the first active lane
3817 * @return value of the lane
3818 */
3819 LLVMValueRef
3820 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane)
3821 {
3822 LLVMTypeRef src_type = LLVMTypeOf(src);
3823 src = ac_to_integer(ctx, src);
3824 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3825 LLVMValueRef ret;
3826
3827 if (bits == 32) {
3828 ret = _ac_build_readlane(ctx, src, lane);
3829 } else {
3830 assert(bits % 32 == 0);
3831 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3832 LLVMValueRef src_vector =
3833 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3834 ret = LLVMGetUndef(vec_type);
3835 for (unsigned i = 0; i < bits / 32; i++) {
3836 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3837 LLVMConstInt(ctx->i32, i, 0), "");
3838 LLVMValueRef ret_comp = _ac_build_readlane(ctx, src, lane);
3839 ret = LLVMBuildInsertElement(ctx->builder, ret, ret_comp,
3840 LLVMConstInt(ctx->i32, i, 0), "");
3841 }
3842 }
3843 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3844 }
3845
3846 LLVMValueRef
3847 ac_build_writelane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef value, LLVMValueRef lane)
3848 {
3849 if (HAVE_LLVM >= 0x0800) {
3850 return ac_build_intrinsic(ctx, "llvm.amdgcn.writelane", ctx->i32,
3851 (LLVMValueRef []) {value, lane, src}, 3,
3852 AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3853 }
3854
3855 LLVMValueRef pred = LLVMBuildICmp(ctx->builder, LLVMIntEQ, lane,
3856 ac_get_thread_id(ctx), "");
3857 return LLVMBuildSelect(ctx->builder, pred, value, src, "");
3858 }
3859
3860 LLVMValueRef
3861 ac_build_mbcnt(struct ac_llvm_context *ctx, LLVMValueRef mask)
3862 {
3863 if (ctx->wave_size == 32) {
3864 return ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
3865 (LLVMValueRef []) { mask, ctx->i32_0 },
3866 2, AC_FUNC_ATTR_READNONE);
3867 }
3868 LLVMValueRef mask_vec = LLVMBuildBitCast(ctx->builder, mask,
3869 LLVMVectorType(ctx->i32, 2),
3870 "");
3871 LLVMValueRef mask_lo = LLVMBuildExtractElement(ctx->builder, mask_vec,
3872 ctx->i32_0, "");
3873 LLVMValueRef mask_hi = LLVMBuildExtractElement(ctx->builder, mask_vec,
3874 ctx->i32_1, "");
3875 LLVMValueRef val =
3876 ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.lo", ctx->i32,
3877 (LLVMValueRef []) { mask_lo, ctx->i32_0 },
3878 2, AC_FUNC_ATTR_READNONE);
3879 val = ac_build_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi", ctx->i32,
3880 (LLVMValueRef []) { mask_hi, val },
3881 2, AC_FUNC_ATTR_READNONE);
3882 return val;
3883 }
3884
3885 enum dpp_ctrl {
3886 _dpp_quad_perm = 0x000,
3887 _dpp_row_sl = 0x100,
3888 _dpp_row_sr = 0x110,
3889 _dpp_row_rr = 0x120,
3890 dpp_wf_sl1 = 0x130,
3891 dpp_wf_rl1 = 0x134,
3892 dpp_wf_sr1 = 0x138,
3893 dpp_wf_rr1 = 0x13C,
3894 dpp_row_mirror = 0x140,
3895 dpp_row_half_mirror = 0x141,
3896 dpp_row_bcast15 = 0x142,
3897 dpp_row_bcast31 = 0x143
3898 };
3899
3900 static inline enum dpp_ctrl
3901 dpp_quad_perm(unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
3902 {
3903 assert(lane0 < 4 && lane1 < 4 && lane2 < 4 && lane3 < 4);
3904 return _dpp_quad_perm | lane0 | (lane1 << 2) | (lane2 << 4) | (lane3 << 6);
3905 }
3906
3907 static inline enum dpp_ctrl
3908 dpp_row_sl(unsigned amount)
3909 {
3910 assert(amount > 0 && amount < 16);
3911 return _dpp_row_sl | amount;
3912 }
3913
3914 static inline enum dpp_ctrl
3915 dpp_row_sr(unsigned amount)
3916 {
3917 assert(amount > 0 && amount < 16);
3918 return _dpp_row_sr | amount;
3919 }
3920
3921 static LLVMValueRef
3922 _ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3923 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3924 bool bound_ctrl)
3925 {
3926 return ac_build_intrinsic(ctx, "llvm.amdgcn.update.dpp.i32",
3927 LLVMTypeOf(old),
3928 (LLVMValueRef[]) {
3929 old, src,
3930 LLVMConstInt(ctx->i32, dpp_ctrl, 0),
3931 LLVMConstInt(ctx->i32, row_mask, 0),
3932 LLVMConstInt(ctx->i32, bank_mask, 0),
3933 LLVMConstInt(ctx->i1, bound_ctrl, 0) },
3934 6, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3935 }
3936
3937 static LLVMValueRef
3938 ac_build_dpp(struct ac_llvm_context *ctx, LLVMValueRef old, LLVMValueRef src,
3939 enum dpp_ctrl dpp_ctrl, unsigned row_mask, unsigned bank_mask,
3940 bool bound_ctrl)
3941 {
3942 LLVMTypeRef src_type = LLVMTypeOf(src);
3943 src = ac_to_integer(ctx, src);
3944 old = ac_to_integer(ctx, old);
3945 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
3946 LLVMValueRef ret;
3947 if (bits == 32) {
3948 ret = _ac_build_dpp(ctx, old, src, dpp_ctrl, row_mask,
3949 bank_mask, bound_ctrl);
3950 } else {
3951 assert(bits % 32 == 0);
3952 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
3953 LLVMValueRef src_vector =
3954 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
3955 LLVMValueRef old_vector =
3956 LLVMBuildBitCast(ctx->builder, old, vec_type, "");
3957 ret = LLVMGetUndef(vec_type);
3958 for (unsigned i = 0; i < bits / 32; i++) {
3959 src = LLVMBuildExtractElement(ctx->builder, src_vector,
3960 LLVMConstInt(ctx->i32, i,
3961 0), "");
3962 old = LLVMBuildExtractElement(ctx->builder, old_vector,
3963 LLVMConstInt(ctx->i32, i,
3964 0), "");
3965 LLVMValueRef ret_comp = _ac_build_dpp(ctx, old, src,
3966 dpp_ctrl,
3967 row_mask,
3968 bank_mask,
3969 bound_ctrl);
3970 ret = LLVMBuildInsertElement(ctx->builder, ret,
3971 ret_comp,
3972 LLVMConstInt(ctx->i32, i,
3973 0), "");
3974 }
3975 }
3976 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
3977 }
3978
3979 static LLVMValueRef
3980 _ac_build_permlane16(struct ac_llvm_context *ctx, LLVMValueRef src, uint64_t sel,
3981 bool exchange_rows, bool bound_ctrl)
3982 {
3983 LLVMValueRef args[6] = {
3984 src,
3985 src,
3986 LLVMConstInt(ctx->i32, sel, false),
3987 LLVMConstInt(ctx->i32, sel >> 32, false),
3988 ctx->i1true, /* fi */
3989 bound_ctrl ? ctx->i1true : ctx->i1false,
3990 };
3991 return ac_build_intrinsic(ctx, exchange_rows ? "llvm.amdgcn.permlanex16"
3992 : "llvm.amdgcn.permlane16",
3993 ctx->i32, args, 6,
3994 AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
3995 }
3996
3997 static LLVMValueRef
3998 ac_build_permlane16(struct ac_llvm_context *ctx, LLVMValueRef src, uint64_t sel,
3999 bool exchange_rows, bool bound_ctrl)
4000 {
4001 LLVMTypeRef src_type = LLVMTypeOf(src);
4002 src = ac_to_integer(ctx, src);
4003 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
4004 LLVMValueRef ret;
4005 if (bits == 32) {
4006 ret = _ac_build_permlane16(ctx, src, sel, exchange_rows,
4007 bound_ctrl);
4008 } else {
4009 assert(bits % 32 == 0);
4010 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
4011 LLVMValueRef src_vector =
4012 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
4013 ret = LLVMGetUndef(vec_type);
4014 for (unsigned i = 0; i < bits / 32; i++) {
4015 src = LLVMBuildExtractElement(ctx->builder, src_vector,
4016 LLVMConstInt(ctx->i32, i,
4017 0), "");
4018 LLVMValueRef ret_comp =
4019 _ac_build_permlane16(ctx, src, sel,
4020 exchange_rows,
4021 bound_ctrl);
4022 ret = LLVMBuildInsertElement(ctx->builder, ret,
4023 ret_comp,
4024 LLVMConstInt(ctx->i32, i,
4025 0), "");
4026 }
4027 }
4028 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
4029 }
4030
4031 static inline unsigned
4032 ds_pattern_bitmode(unsigned and_mask, unsigned or_mask, unsigned xor_mask)
4033 {
4034 assert(and_mask < 32 && or_mask < 32 && xor_mask < 32);
4035 return and_mask | (or_mask << 5) | (xor_mask << 10);
4036 }
4037
4038 static LLVMValueRef
4039 _ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
4040 {
4041 return ac_build_intrinsic(ctx, "llvm.amdgcn.ds.swizzle",
4042 LLVMTypeOf(src), (LLVMValueRef []) {
4043 src, LLVMConstInt(ctx->i32, mask, 0) },
4044 2, AC_FUNC_ATTR_READNONE | AC_FUNC_ATTR_CONVERGENT);
4045 }
4046
4047 LLVMValueRef
4048 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask)
4049 {
4050 LLVMTypeRef src_type = LLVMTypeOf(src);
4051 src = ac_to_integer(ctx, src);
4052 unsigned bits = LLVMGetIntTypeWidth(LLVMTypeOf(src));
4053 LLVMValueRef ret;
4054 if (bits == 32) {
4055 ret = _ac_build_ds_swizzle(ctx, src, mask);
4056 } else {
4057 assert(bits % 32 == 0);
4058 LLVMTypeRef vec_type = LLVMVectorType(ctx->i32, bits / 32);
4059 LLVMValueRef src_vector =
4060 LLVMBuildBitCast(ctx->builder, src, vec_type, "");
4061 ret = LLVMGetUndef(vec_type);
4062 for (unsigned i = 0; i < bits / 32; i++) {
4063 src = LLVMBuildExtractElement(ctx->builder, src_vector,
4064 LLVMConstInt(ctx->i32, i,
4065 0), "");
4066 LLVMValueRef ret_comp = _ac_build_ds_swizzle(ctx, src,
4067 mask);
4068 ret = LLVMBuildInsertElement(ctx->builder, ret,
4069 ret_comp,
4070 LLVMConstInt(ctx->i32, i,
4071 0), "");
4072 }
4073 }
4074 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
4075 }
4076
4077 static LLVMValueRef
4078 ac_build_wwm(struct ac_llvm_context *ctx, LLVMValueRef src)
4079 {
4080 char name[32], type[8];
4081 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
4082 snprintf(name, sizeof(name), "llvm.amdgcn.wwm.%s", type);
4083 return ac_build_intrinsic(ctx, name, LLVMTypeOf(src),
4084 (LLVMValueRef []) { src }, 1,
4085 AC_FUNC_ATTR_READNONE);
4086 }
4087
4088 static LLVMValueRef
4089 ac_build_set_inactive(struct ac_llvm_context *ctx, LLVMValueRef src,
4090 LLVMValueRef inactive)
4091 {
4092 char name[33], type[8];
4093 LLVMTypeRef src_type = LLVMTypeOf(src);
4094 src = ac_to_integer(ctx, src);
4095 inactive = ac_to_integer(ctx, inactive);
4096 ac_build_type_name_for_intr(LLVMTypeOf(src), type, sizeof(type));
4097 snprintf(name, sizeof(name), "llvm.amdgcn.set.inactive.%s", type);
4098 LLVMValueRef ret =
4099 ac_build_intrinsic(ctx, name,
4100 LLVMTypeOf(src), (LLVMValueRef []) {
4101 src, inactive }, 2,
4102 AC_FUNC_ATTR_READNONE |
4103 AC_FUNC_ATTR_CONVERGENT);
4104 return LLVMBuildBitCast(ctx->builder, ret, src_type, "");
4105 }
4106
4107 static LLVMValueRef
4108 get_reduction_identity(struct ac_llvm_context *ctx, nir_op op, unsigned type_size)
4109 {
4110 if (type_size == 4) {
4111 switch (op) {
4112 case nir_op_iadd: return ctx->i32_0;
4113 case nir_op_fadd: return ctx->f32_0;
4114 case nir_op_imul: return ctx->i32_1;
4115 case nir_op_fmul: return ctx->f32_1;
4116 case nir_op_imin: return LLVMConstInt(ctx->i32, INT32_MAX, 0);
4117 case nir_op_umin: return LLVMConstInt(ctx->i32, UINT32_MAX, 0);
4118 case nir_op_fmin: return LLVMConstReal(ctx->f32, INFINITY);
4119 case nir_op_imax: return LLVMConstInt(ctx->i32, INT32_MIN, 0);
4120 case nir_op_umax: return ctx->i32_0;
4121 case nir_op_fmax: return LLVMConstReal(ctx->f32, -INFINITY);
4122 case nir_op_iand: return LLVMConstInt(ctx->i32, -1, 0);
4123 case nir_op_ior: return ctx->i32_0;
4124 case nir_op_ixor: return ctx->i32_0;
4125 default:
4126 unreachable("bad reduction intrinsic");
4127 }
4128 } else { /* type_size == 64bit */
4129 switch (op) {
4130 case nir_op_iadd: return ctx->i64_0;
4131 case nir_op_fadd: return ctx->f64_0;
4132 case nir_op_imul: return ctx->i64_1;
4133 case nir_op_fmul: return ctx->f64_1;
4134 case nir_op_imin: return LLVMConstInt(ctx->i64, INT64_MAX, 0);
4135 case nir_op_umin: return LLVMConstInt(ctx->i64, UINT64_MAX, 0);
4136 case nir_op_fmin: return LLVMConstReal(ctx->f64, INFINITY);
4137 case nir_op_imax: return LLVMConstInt(ctx->i64, INT64_MIN, 0);
4138 case nir_op_umax: return ctx->i64_0;
4139 case nir_op_fmax: return LLVMConstReal(ctx->f64, -INFINITY);
4140 case nir_op_iand: return LLVMConstInt(ctx->i64, -1, 0);
4141 case nir_op_ior: return ctx->i64_0;
4142 case nir_op_ixor: return ctx->i64_0;
4143 default:
4144 unreachable("bad reduction intrinsic");
4145 }
4146 }
4147 }
4148
4149 static LLVMValueRef
4150 ac_build_alu_op(struct ac_llvm_context *ctx, LLVMValueRef lhs, LLVMValueRef rhs, nir_op op)
4151 {
4152 bool _64bit = ac_get_type_size(LLVMTypeOf(lhs)) == 8;
4153 switch (op) {
4154 case nir_op_iadd: return LLVMBuildAdd(ctx->builder, lhs, rhs, "");
4155 case nir_op_fadd: return LLVMBuildFAdd(ctx->builder, lhs, rhs, "");
4156 case nir_op_imul: return LLVMBuildMul(ctx->builder, lhs, rhs, "");
4157 case nir_op_fmul: return LLVMBuildFMul(ctx->builder, lhs, rhs, "");
4158 case nir_op_imin: return LLVMBuildSelect(ctx->builder,
4159 LLVMBuildICmp(ctx->builder, LLVMIntSLT, lhs, rhs, ""),
4160 lhs, rhs, "");
4161 case nir_op_umin: return LLVMBuildSelect(ctx->builder,
4162 LLVMBuildICmp(ctx->builder, LLVMIntULT, lhs, rhs, ""),
4163 lhs, rhs, "");
4164 case nir_op_fmin: return ac_build_intrinsic(ctx,
4165 _64bit ? "llvm.minnum.f64" : "llvm.minnum.f32",
4166 _64bit ? ctx->f64 : ctx->f32,
4167 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
4168 case nir_op_imax: return LLVMBuildSelect(ctx->builder,
4169 LLVMBuildICmp(ctx->builder, LLVMIntSGT, lhs, rhs, ""),
4170 lhs, rhs, "");
4171 case nir_op_umax: return LLVMBuildSelect(ctx->builder,
4172 LLVMBuildICmp(ctx->builder, LLVMIntUGT, lhs, rhs, ""),
4173 lhs, rhs, "");
4174 case nir_op_fmax: return ac_build_intrinsic(ctx,
4175 _64bit ? "llvm.maxnum.f64" : "llvm.maxnum.f32",
4176 _64bit ? ctx->f64 : ctx->f32,
4177 (LLVMValueRef[]){lhs, rhs}, 2, AC_FUNC_ATTR_READNONE);
4178 case nir_op_iand: return LLVMBuildAnd(ctx->builder, lhs, rhs, "");
4179 case nir_op_ior: return LLVMBuildOr(ctx->builder, lhs, rhs, "");
4180 case nir_op_ixor: return LLVMBuildXor(ctx->builder, lhs, rhs, "");
4181 default:
4182 unreachable("bad reduction intrinsic");
4183 }
4184 }
4185
4186 /**
4187 * \param maxprefix specifies that the result only needs to be correct for a
4188 * prefix of this many threads
4189 *
4190 * TODO: add inclusive and excluse scan functions for GFX6.
4191 */
4192 static LLVMValueRef
4193 ac_build_scan(struct ac_llvm_context *ctx, nir_op op, LLVMValueRef src, LLVMValueRef identity,
4194 unsigned maxprefix, bool inclusive)
4195 {
4196 LLVMValueRef result, tmp;
4197
4198 if (ctx->chip_class >= GFX10) {
4199 result = inclusive ? src : identity;
4200 } else {
4201 if (inclusive)
4202 result = src;
4203 else
4204 result = ac_build_dpp(ctx, identity, src, dpp_wf_sr1, 0xf, 0xf, false);
4205 }
4206 if (maxprefix <= 1)
4207 return result;
4208 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(1), 0xf, 0xf, false);
4209 result = ac_build_alu_op(ctx, result, tmp, op);
4210 if (maxprefix <= 2)
4211 return result;
4212 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(2), 0xf, 0xf, false);
4213 result = ac_build_alu_op(ctx, result, tmp, op);
4214 if (maxprefix <= 3)
4215 return result;
4216 tmp = ac_build_dpp(ctx, identity, src, dpp_row_sr(3), 0xf, 0xf, false);
4217 result = ac_build_alu_op(ctx, result, tmp, op);
4218 if (maxprefix <= 4)
4219 return result;
4220 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(4), 0xf, 0xe, false);
4221 result = ac_build_alu_op(ctx, result, tmp, op);
4222 if (maxprefix <= 8)
4223 return result;
4224 tmp = ac_build_dpp(ctx, identity, result, dpp_row_sr(8), 0xf, 0xc, false);
4225 result = ac_build_alu_op(ctx, result, tmp, op);
4226 if (maxprefix <= 16)
4227 return result;
4228
4229 if (ctx->chip_class >= GFX10) {
4230 /* dpp_row_bcast{15,31} are not supported on gfx10. */
4231 LLVMBuilderRef builder = ctx->builder;
4232 LLVMValueRef tid = ac_get_thread_id(ctx);
4233 LLVMValueRef cc;
4234 /* TODO-GFX10: Can we get better code-gen by putting this into
4235 * a branch so that LLVM generates EXEC mask manipulations? */
4236 if (inclusive)
4237 tmp = result;
4238 else
4239 tmp = ac_build_alu_op(ctx, result, src, op);
4240 tmp = ac_build_permlane16(ctx, tmp, ~(uint64_t)0, true, false);
4241 tmp = ac_build_alu_op(ctx, result, tmp, op);
4242 cc = LLVMBuildAnd(builder, tid, LLVMConstInt(ctx->i32, 16, false), "");
4243 cc = LLVMBuildICmp(builder, LLVMIntNE, cc, ctx->i32_0, "");
4244 result = LLVMBuildSelect(builder, cc, tmp, result, "");
4245 if (maxprefix <= 32)
4246 return result;
4247
4248 if (inclusive)
4249 tmp = result;
4250 else
4251 tmp = ac_build_alu_op(ctx, result, src, op);
4252 tmp = ac_build_readlane(ctx, tmp, LLVMConstInt(ctx->i32, 31, false));
4253 tmp = ac_build_alu_op(ctx, result, tmp, op);
4254 cc = LLVMBuildICmp(builder, LLVMIntUGE, tid,
4255 LLVMConstInt(ctx->i32, 32, false), "");
4256 result = LLVMBuildSelect(builder, cc, tmp, result, "");
4257 return result;
4258 }
4259
4260 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
4261 result = ac_build_alu_op(ctx, result, tmp, op);
4262 if (maxprefix <= 32)
4263 return result;
4264 tmp = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
4265 result = ac_build_alu_op(ctx, result, tmp, op);
4266 return result;
4267 }
4268
4269 LLVMValueRef
4270 ac_build_inclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
4271 {
4272 LLVMValueRef result;
4273
4274 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
4275 LLVMBuilderRef builder = ctx->builder;
4276 src = LLVMBuildZExt(builder, src, ctx->i32, "");
4277 result = ac_build_ballot(ctx, src);
4278 result = ac_build_mbcnt(ctx, result);
4279 result = LLVMBuildAdd(builder, result, src, "");
4280 return result;
4281 }
4282
4283 ac_build_optimization_barrier(ctx, &src);
4284
4285 LLVMValueRef identity =
4286 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
4287 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
4288 LLVMTypeOf(identity), "");
4289 result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, true);
4290
4291 return ac_build_wwm(ctx, result);
4292 }
4293
4294 LLVMValueRef
4295 ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op)
4296 {
4297 LLVMValueRef result;
4298
4299 if (LLVMTypeOf(src) == ctx->i1 && op == nir_op_iadd) {
4300 LLVMBuilderRef builder = ctx->builder;
4301 src = LLVMBuildZExt(builder, src, ctx->i32, "");
4302 result = ac_build_ballot(ctx, src);
4303 result = ac_build_mbcnt(ctx, result);
4304 return result;
4305 }
4306
4307 ac_build_optimization_barrier(ctx, &src);
4308
4309 LLVMValueRef identity =
4310 get_reduction_identity(ctx, op, ac_get_type_size(LLVMTypeOf(src)));
4311 result = LLVMBuildBitCast(ctx->builder, ac_build_set_inactive(ctx, src, identity),
4312 LLVMTypeOf(identity), "");
4313 result = ac_build_scan(ctx, op, result, identity, ctx->wave_size, false);
4314
4315 return ac_build_wwm(ctx, result);
4316 }
4317
4318 LLVMValueRef
4319 ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size)
4320 {
4321 if (cluster_size == 1) return src;
4322 ac_build_optimization_barrier(ctx, &src);
4323 LLVMValueRef result, swap;
4324 LLVMValueRef identity = get_reduction_identity(ctx, op,
4325 ac_get_type_size(LLVMTypeOf(src)));
4326 result = LLVMBuildBitCast(ctx->builder,
4327 ac_build_set_inactive(ctx, src, identity),
4328 LLVMTypeOf(identity), "");
4329 swap = ac_build_quad_swizzle(ctx, result, 1, 0, 3, 2);
4330 result = ac_build_alu_op(ctx, result, swap, op);
4331 if (cluster_size == 2) return ac_build_wwm(ctx, result);
4332
4333 swap = ac_build_quad_swizzle(ctx, result, 2, 3, 0, 1);
4334 result = ac_build_alu_op(ctx, result, swap, op);
4335 if (cluster_size == 4) return ac_build_wwm(ctx, result);
4336
4337 if (ctx->chip_class >= GFX8)
4338 swap = ac_build_dpp(ctx, identity, result, dpp_row_half_mirror, 0xf, 0xf, false);
4339 else
4340 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x04));
4341 result = ac_build_alu_op(ctx, result, swap, op);
4342 if (cluster_size == 8) return ac_build_wwm(ctx, result);
4343
4344 if (ctx->chip_class >= GFX8)
4345 swap = ac_build_dpp(ctx, identity, result, dpp_row_mirror, 0xf, 0xf, false);
4346 else
4347 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x08));
4348 result = ac_build_alu_op(ctx, result, swap, op);
4349 if (cluster_size == 16) return ac_build_wwm(ctx, result);
4350
4351 if (ctx->chip_class >= GFX10)
4352 swap = ac_build_permlane16(ctx, result, 0, true, false);
4353 else if (ctx->chip_class >= GFX8 && cluster_size != 32)
4354 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast15, 0xa, 0xf, false);
4355 else
4356 swap = ac_build_ds_swizzle(ctx, result, ds_pattern_bitmode(0x1f, 0, 0x10));
4357 result = ac_build_alu_op(ctx, result, swap, op);
4358 if (cluster_size == 32) return ac_build_wwm(ctx, result);
4359
4360 if (ctx->chip_class >= GFX8) {
4361 if (ctx->chip_class >= GFX10)
4362 swap = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 31, false));
4363 else
4364 swap = ac_build_dpp(ctx, identity, result, dpp_row_bcast31, 0xc, 0xf, false);
4365 result = ac_build_alu_op(ctx, result, swap, op);
4366 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 63, 0));
4367 return ac_build_wwm(ctx, result);
4368 } else {
4369 swap = ac_build_readlane(ctx, result, ctx->i32_0);
4370 result = ac_build_readlane(ctx, result, LLVMConstInt(ctx->i32, 32, 0));
4371 result = ac_build_alu_op(ctx, result, swap, op);
4372 return ac_build_wwm(ctx, result);
4373 }
4374 }
4375
4376 /**
4377 * "Top half" of a scan that reduces per-wave values across an entire
4378 * workgroup.
4379 *
4380 * The source value must be present in the highest lane of the wave, and the
4381 * highest lane must be live.
4382 */
4383 void
4384 ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4385 {
4386 if (ws->maxwaves <= 1)
4387 return;
4388
4389 const LLVMValueRef last_lane = LLVMConstInt(ctx->i32, ctx->wave_size - 1, false);
4390 LLVMBuilderRef builder = ctx->builder;
4391 LLVMValueRef tid = ac_get_thread_id(ctx);
4392 LLVMValueRef tmp;
4393
4394 tmp = LLVMBuildICmp(builder, LLVMIntEQ, tid, last_lane, "");
4395 ac_build_ifcc(ctx, tmp, 1000);
4396 LLVMBuildStore(builder, ws->src, LLVMBuildGEP(builder, ws->scratch, &ws->waveidx, 1, ""));
4397 ac_build_endif(ctx, 1000);
4398 }
4399
4400 /**
4401 * "Bottom half" of a scan that reduces per-wave values across an entire
4402 * workgroup.
4403 *
4404 * The caller must place a barrier between the top and bottom halves.
4405 */
4406 void
4407 ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4408 {
4409 const LLVMTypeRef type = LLVMTypeOf(ws->src);
4410 const LLVMValueRef identity =
4411 get_reduction_identity(ctx, ws->op, ac_get_type_size(type));
4412
4413 if (ws->maxwaves <= 1) {
4414 ws->result_reduce = ws->src;
4415 ws->result_inclusive = ws->src;
4416 ws->result_exclusive = identity;
4417 return;
4418 }
4419 assert(ws->maxwaves <= 32);
4420
4421 LLVMBuilderRef builder = ctx->builder;
4422 LLVMValueRef tid = ac_get_thread_id(ctx);
4423 LLVMBasicBlockRef bbs[2];
4424 LLVMValueRef phivalues_scan[2];
4425 LLVMValueRef tmp, tmp2;
4426
4427 bbs[0] = LLVMGetInsertBlock(builder);
4428 phivalues_scan[0] = LLVMGetUndef(type);
4429
4430 if (ws->enable_reduce)
4431 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->numwaves, "");
4432 else if (ws->enable_inclusive)
4433 tmp = LLVMBuildICmp(builder, LLVMIntULE, tid, ws->waveidx, "");
4434 else
4435 tmp = LLVMBuildICmp(builder, LLVMIntULT, tid, ws->waveidx, "");
4436 ac_build_ifcc(ctx, tmp, 1001);
4437 {
4438 tmp = LLVMBuildLoad(builder, LLVMBuildGEP(builder, ws->scratch, &tid, 1, ""), "");
4439
4440 ac_build_optimization_barrier(ctx, &tmp);
4441
4442 bbs[1] = LLVMGetInsertBlock(builder);
4443 phivalues_scan[1] = ac_build_scan(ctx, ws->op, tmp, identity, ws->maxwaves, true);
4444 }
4445 ac_build_endif(ctx, 1001);
4446
4447 const LLVMValueRef scan = ac_build_phi(ctx, type, 2, phivalues_scan, bbs);
4448
4449 if (ws->enable_reduce) {
4450 tmp = LLVMBuildSub(builder, ws->numwaves, ctx->i32_1, "");
4451 ws->result_reduce = ac_build_readlane(ctx, scan, tmp);
4452 }
4453 if (ws->enable_inclusive)
4454 ws->result_inclusive = ac_build_readlane(ctx, scan, ws->waveidx);
4455 if (ws->enable_exclusive) {
4456 tmp = LLVMBuildSub(builder, ws->waveidx, ctx->i32_1, "");
4457 tmp = ac_build_readlane(ctx, scan, tmp);
4458 tmp2 = LLVMBuildICmp(builder, LLVMIntEQ, ws->waveidx, ctx->i32_0, "");
4459 ws->result_exclusive = LLVMBuildSelect(builder, tmp2, identity, tmp, "");
4460 }
4461 }
4462
4463 /**
4464 * Inclusive scan of a per-wave value across an entire workgroup.
4465 *
4466 * This implies an s_barrier instruction.
4467 *
4468 * Unlike ac_build_inclusive_scan, the caller \em must ensure that all threads
4469 * of the workgroup are live. (This requirement cannot easily be relaxed in a
4470 * useful manner because of the barrier in the algorithm.)
4471 */
4472 void
4473 ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4474 {
4475 ac_build_wg_wavescan_top(ctx, ws);
4476 ac_build_s_barrier(ctx);
4477 ac_build_wg_wavescan_bottom(ctx, ws);
4478 }
4479
4480 /**
4481 * "Top half" of a scan that reduces per-thread values across an entire
4482 * workgroup.
4483 *
4484 * All lanes must be active when this code runs.
4485 */
4486 void
4487 ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4488 {
4489 if (ws->enable_exclusive) {
4490 ws->extra = ac_build_exclusive_scan(ctx, ws->src, ws->op);
4491 if (LLVMTypeOf(ws->src) == ctx->i1 && ws->op == nir_op_iadd)
4492 ws->src = LLVMBuildZExt(ctx->builder, ws->src, ctx->i32, "");
4493 ws->src = ac_build_alu_op(ctx, ws->extra, ws->src, ws->op);
4494 } else {
4495 ws->src = ac_build_inclusive_scan(ctx, ws->src, ws->op);
4496 }
4497
4498 bool enable_inclusive = ws->enable_inclusive;
4499 bool enable_exclusive = ws->enable_exclusive;
4500 ws->enable_inclusive = false;
4501 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
4502 ac_build_wg_wavescan_top(ctx, ws);
4503 ws->enable_inclusive = enable_inclusive;
4504 ws->enable_exclusive = enable_exclusive;
4505 }
4506
4507 /**
4508 * "Bottom half" of a scan that reduces per-thread values across an entire
4509 * workgroup.
4510 *
4511 * The caller must place a barrier between the top and bottom halves.
4512 */
4513 void
4514 ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4515 {
4516 bool enable_inclusive = ws->enable_inclusive;
4517 bool enable_exclusive = ws->enable_exclusive;
4518 ws->enable_inclusive = false;
4519 ws->enable_exclusive = ws->enable_exclusive || enable_inclusive;
4520 ac_build_wg_wavescan_bottom(ctx, ws);
4521 ws->enable_inclusive = enable_inclusive;
4522 ws->enable_exclusive = enable_exclusive;
4523
4524 /* ws->result_reduce is already the correct value */
4525 if (ws->enable_inclusive)
4526 ws->result_inclusive = ac_build_alu_op(ctx, ws->result_inclusive, ws->src, ws->op);
4527 if (ws->enable_exclusive)
4528 ws->result_exclusive = ac_build_alu_op(ctx, ws->result_exclusive, ws->extra, ws->op);
4529 }
4530
4531 /**
4532 * A scan that reduces per-thread values across an entire workgroup.
4533 *
4534 * The caller must ensure that all lanes are active when this code runs
4535 * (WWM is insufficient!), because there is an implied barrier.
4536 */
4537 void
4538 ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws)
4539 {
4540 ac_build_wg_scan_top(ctx, ws);
4541 ac_build_s_barrier(ctx);
4542 ac_build_wg_scan_bottom(ctx, ws);
4543 }
4544
4545 LLVMValueRef
4546 ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
4547 unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3)
4548 {
4549 unsigned mask = dpp_quad_perm(lane0, lane1, lane2, lane3);
4550 if (ctx->chip_class >= GFX8) {
4551 return ac_build_dpp(ctx, src, src, mask, 0xf, 0xf, false);
4552 } else {
4553 return ac_build_ds_swizzle(ctx, src, (1 << 15) | mask);
4554 }
4555 }
4556
4557 LLVMValueRef
4558 ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index)
4559 {
4560 index = LLVMBuildMul(ctx->builder, index, LLVMConstInt(ctx->i32, 4, 0), "");
4561 return ac_build_intrinsic(ctx,
4562 "llvm.amdgcn.ds.bpermute", ctx->i32,
4563 (LLVMValueRef []) {index, src}, 2,
4564 AC_FUNC_ATTR_READNONE |
4565 AC_FUNC_ATTR_CONVERGENT);
4566 }
4567
4568 LLVMValueRef
4569 ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
4570 unsigned bitsize)
4571 {
4572 LLVMTypeRef type;
4573 char *intr;
4574
4575 if (bitsize == 16) {
4576 intr = "llvm.amdgcn.frexp.exp.i16.f16";
4577 type = ctx->i16;
4578 } else if (bitsize == 32) {
4579 intr = "llvm.amdgcn.frexp.exp.i32.f32";
4580 type = ctx->i32;
4581 } else {
4582 intr = "llvm.amdgcn.frexp.exp.i32.f64";
4583 type = ctx->i32;
4584 }
4585
4586 LLVMValueRef params[] = {
4587 src0,
4588 };
4589 return ac_build_intrinsic(ctx, intr, type, params, 1,
4590 AC_FUNC_ATTR_READNONE);
4591 }
4592 LLVMValueRef
4593 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
4594 unsigned bitsize)
4595 {
4596 LLVMTypeRef type;
4597 char *intr;
4598
4599 if (bitsize == 16) {
4600 intr = "llvm.amdgcn.frexp.mant.f16";
4601 type = ctx->f16;
4602 } else if (bitsize == 32) {
4603 intr = "llvm.amdgcn.frexp.mant.f32";
4604 type = ctx->f32;
4605 } else {
4606 intr = "llvm.amdgcn.frexp.mant.f64";
4607 type = ctx->f64;
4608 }
4609
4610 LLVMValueRef params[] = {
4611 src0,
4612 };
4613 return ac_build_intrinsic(ctx, intr, type, params, 1,
4614 AC_FUNC_ATTR_READNONE);
4615 }
4616
4617 /*
4618 * this takes an I,J coordinate pair,
4619 * and works out the X and Y derivatives.
4620 * it returns DDX(I), DDX(J), DDY(I), DDY(J).
4621 */
4622 LLVMValueRef
4623 ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij)
4624 {
4625 LLVMValueRef result[4], a;
4626 unsigned i;
4627
4628 for (i = 0; i < 2; i++) {
4629 a = LLVMBuildExtractElement(ctx->builder, interp_ij,
4630 LLVMConstInt(ctx->i32, i, false), "");
4631 result[i] = ac_build_ddxy(ctx, AC_TID_MASK_TOP_LEFT, 1, a);
4632 result[2+i] = ac_build_ddxy(ctx, AC_TID_MASK_TOP_LEFT, 2, a);
4633 }
4634 return ac_build_gather_values(ctx, result, 4);
4635 }
4636
4637 LLVMValueRef
4638 ac_build_load_helper_invocation(struct ac_llvm_context *ctx)
4639 {
4640 LLVMValueRef result = ac_build_intrinsic(ctx, "llvm.amdgcn.ps.live",
4641 ctx->i1, NULL, 0,
4642 AC_FUNC_ATTR_READNONE);
4643 result = LLVMBuildNot(ctx->builder, result, "");
4644 return LLVMBuildSExt(ctx->builder, result, ctx->i32, "");
4645 }
4646
4647 LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
4648 LLVMValueRef *args, unsigned num_args)
4649 {
4650 LLVMValueRef ret = LLVMBuildCall(ctx->builder, func, args, num_args, "");
4651 LLVMSetInstructionCallConv(ret, LLVMGetFunctionCallConv(func));
4652 return ret;
4653 }